Merge branch 'develop' into pr/5008

.github/contributors/Jan-711.md

@@ -0,0 +1,106 @@
+# spaCy contributor agreement
+
+This spaCy Contributor Agreement (**"SCA"**) is based on the
+[Oracle Contributor Agreement](http://www.oracle.com/technetwork/oca-405177.pdf).
+The SCA applies to any contribution that you make to any product or project
+managed by us (the **"project"**), and sets out the intellectual property rights
+you grant to us in the contributed materials. The term **"us"** shall mean
+[ExplosionAI GmbH](https://explosion.ai/legal). The term
+**"you"** shall mean the person or entity identified below.
+
+If you agree to be bound by these terms, fill in the information requested
+below and include the filled-in version with your first pull request, under the
+folder [`.github/contributors/`](/.github/contributors/). The name of the file
+should be your GitHub username, with the extension `.md`. For example, the user
+example_user would create the file `.github/contributors/example_user.md`.
+
+Read this agreement carefully before signing. These terms and conditions
+constitute a binding legal agreement.
+
+## Contributor Agreement
+
+1. The term "contribution" or "contributed materials" means any source code,
+object code, patch, tool, sample, graphic, specification, manual,
+documentation, or any other material posted or submitted by you to the project.
+
+2. With respect to any worldwide copyrights, or copyright applications and
+registrations, in your contribution:
+
+    * you hereby assign to us joint ownership, and to the extent that such
+    assignment is or becomes invalid, ineffective or unenforceable, you hereby
+    grant to us a perpetual, irrevocable, non-exclusive, worldwide, no-charge,
+    royalty-free, unrestricted license to exercise all rights under those
+    copyrights. This includes, at our option, the right to sublicense these same
+    rights to third parties through multiple levels of sublicensees or other
+    licensing arrangements;
+
+    * you agree that each of us can do all things in relation to your
+    contribution as if each of us were the sole owners, and if one of us makes
+    a derivative work of your contribution, the one who makes the derivative
+    work (or has it made will be the sole owner of that derivative work;
+
+    * you agree that you will not assert any moral rights in your contribution
+    against us, our licensees or transferees;
+
+    * you agree that we may register a copyright in your contribution and
+    exercise all ownership rights associated with it; and
+
+    * you agree that neither of us has any duty to consult with, obtain the
+    consent of, pay or render an accounting to the other for any use or
+    distribution of your contribution.
+
+3. With respect to any patents you own, or that you can license without payment
+to any third party, you hereby grant to us a perpetual, irrevocable,
+non-exclusive, worldwide, no-charge, royalty-free license to:
+
+    * make, have made, use, sell, offer to sell, import, and otherwise transfer
+    your contribution in whole or in part, alone or in combination with or
+    included in any product, work or materials arising out of the project to
+    which your contribution was submitted, and
+
+    * at our option, to sublicense these same rights to third parties through
+    multiple levels of sublicensees or other licensing arrangements.
+
+4. Except as set out above, you keep all right, title, and interest in your
+contribution. The rights that you grant to us under these terms are effective
+on the date you first submitted a contribution to us, even if your submission
+took place before the date you sign these terms.
+
+5. You covenant, represent, warrant and agree that:
+
+    * Each contribution that you submit is and shall be an original work of
+    authorship and you can legally grant the rights set out in this SCA;
+
+    * to the best of your knowledge, each contribution will not violate any
+    third party's copyrights, trademarks, patents, or other intellectual
+    property rights; and
+
+    * each contribution shall be in compliance with U.S. export control laws and
+    other applicable export and import laws. You agree to notify us if you
+    become aware of any circumstance which would make any of the foregoing
+    representations inaccurate in any respect. We may publicly disclose your
+    participation in the project, including the fact that you have signed the SCA.
+
+6. This SCA is governed by the laws of the State of California and applicable
+U.S. Federal law. Any choice of law rules will not apply.
+
+7. Please place an “x” on one of the applicable statement below. Please do NOT
+mark both statements:
+
+    * [x] I am signing on behalf of myself as an individual and no other person
+    or entity, including my employer, has or will have rights with respect to my
+    contributions.
+
+    * [ ] I am signing on behalf of my employer or a legal entity and I have the
+    actual authority to contractually bind that entity.
+
+## Contributor Details
+
+| Field                          | Entry                |
+|------------------------------- | -------------------- |
+| Name                           | Jan Jessewitsch      |
+| Company name (if applicable)   |                      |
+| Title or role (if applicable)  |                      |
+| Date                           | 16.02.2020           |
+| GitHub username                | Jan-711              |
+| Website (optional)             |                      |

.gitignore

@@ -39,6 +39,7 @@ __pycache__/
 .env*
 .~env/
 .venv
+env3.6/
 venv/
 .dev
 .denv
@@ -111,3 +112,6 @@ Desktop.ini
 
 # Pycharm project files
 *.idea
+
+# IPython
+.ipynb_checkpoints/

.travis.yml

@@ -1,23 +0,0 @@
-language: python
-sudo: false
-cache: pip
-dist: trusty
-group: edge
-python:
-   - "2.7"
-os:
-  - linux
-install:
-  - "pip install -r requirements.txt"
-  - "python setup.py build_ext --inplace"
-  - "pip install -e ."
-script:
-  - "cat /proc/cpuinfo | grep flags | head -n 1"
-  - "python -m pytest --tb=native spacy"
-branches:
-  except:
-    - spacy.io
-notifications:
-  slack:
-    secure: F8GvqnweSdzImuLL64TpfG0i5rYl89liyr9tmFVsHl4c0DNiDuGhZivUz0M1broS8svE3OPOllLfQbACG/4KxD890qfF9MoHzvRDlp7U+RtwMV/YAkYn8MGWjPIbRbX0HpGdY7O2Rc9Qy4Kk0T8ZgiqXYIqAz2Eva9/9BlSmsJQ=
-  email: false

CONTRIBUTING.md

@@ -280,23 +280,7 @@ except:  # noqa: E722
 
 ### Python conventions
 
-All Python code must be written in an **intersection of Python 2 and Python 3**.
+All Python code must be written **compatible with Python 3.6+**.
-This is easy in Cython, but somewhat ugly in Python. Logic that deals with
-Python or platform compatibility should only live in
-[`spacy.compat`](spacy/compat.py). To distinguish them from the builtin
-functions, replacement functions are suffixed with an underscore, for example
-`unicode_`. If you need to access the user's version or platform information,
-for example to show more specific error messages, you can use the `is_config()`
-helper function.
-
-```python
-from .compat import unicode_, is_config
-
-compatible_unicode = unicode_('hello world')
-if is_config(windows=True, python2=True):
-    print("You are using Python 2 on Windows.")
-```
-
 Code that interacts with the file-system should accept objects that follow the
 `pathlib.Path` API, without assuming that the object inherits from `pathlib.Path`.
 If the function is user-facing and takes a path as an argument, it should check

README.md

@@ -15,7 +15,6 @@ It's commercial open-source software, released under the MIT license.
 [Check out the release notes here.](https://github.com/explosion/spaCy/releases)
 
 [![Azure Pipelines](<https://img.shields.io/azure-devops/build/explosion-ai/public/8/master.svg?logo=azure-pipelines&style=flat-square&label=build+(3.x)>)](https://dev.azure.com/explosion-ai/public/_build?definitionId=8)
-[![Travis Build Status](<https://img.shields.io/travis/explosion/spaCy/master.svg?style=flat-square&logo=travis-ci&logoColor=white&label=build+(2.7)>)](https://travis-ci.org/explosion/spaCy)
 [![Current Release Version](https://img.shields.io/github/release/explosion/spacy.svg?style=flat-square&logo=github)](https://github.com/explosion/spaCy/releases)
 [![pypi Version](https://img.shields.io/pypi/v/spacy.svg?style=flat-square&logo=pypi&logoColor=white)](https://pypi.org/project/spacy/)
 [![conda Version](https://img.shields.io/conda/vn/conda-forge/spacy.svg?style=flat-square&logo=conda-forge&logoColor=white)](https://anaconda.org/conda-forge/spacy)
@@ -98,12 +97,19 @@ For detailed installation instructions, see the
 
 - **Operating system**: macOS / OS X · Linux · Windows (Cygwin, MinGW, Visual
   Studio)
-- **Python version**: Python 2.7, 3.5+ (only 64 bit)
+- **Python version**: Python 3.6+ (only 64 bit)
 - **Package managers**: [pip] · [conda] (via `conda-forge`)
 
 [pip]: https://pypi.org/project/spacy/
 [conda]: https://anaconda.org/conda-forge/spacy
 
+> ⚠️ **Important note for Python 3.8:** We can't yet ship pre-compiled binary
+> wheels for spaCy that work on Python 3.8, as we're still waiting for our CI
+> providers and other tooling to support it. This means that in order to run
+> spaCy on Python 3.8, you'll need [a compiler installed](#source) and compile
+> the library and its Cython dependencies locally. If this is causing problems
+> for you, the easiest solution is to **use Python 3.7** in the meantime.
+
 ### pip
 
 Using pip, spaCy releases are available as source packages and binary wheels (as
@@ -262,9 +268,7 @@ and git preinstalled.
 Install a version of the
 [Visual C++ Build Tools](https://visualstudio.microsoft.com/visual-cpp-build-tools/)
 or [Visual Studio Express](https://visualstudio.microsoft.com/vs/express/) that
-matches the version that was used to compile your Python interpreter. For
+matches the version that was used to compile your Python interpreter.
-official distributions these are VS 2008 (Python 2.7), VS 2010 (Python 3.4) and
-VS 2015 (Python 3.5).
 
 ## Run tests
 

azure-pipelines.yml

@@ -35,12 +35,6 @@ jobs:
   dependsOn: 'Validate'
   strategy:
     matrix:
-      Python35Linux:
-        imageName: 'ubuntu-16.04'
-        python.version: '3.5'
-      Python35Windows:
-        imageName: 'vs2017-win2016'
-        python.version: '3.5'
       Python36Linux:
         imageName: 'ubuntu-16.04'
         python.version: '3.6'

bin/cythonize.py

@@ -1,169 +0,0 @@
-#!/usr/bin/env python
-""" cythonize.py
-
-Cythonize pyx files into C++ files as needed.
-
-Usage: cythonize.py [root]
-
-Checks pyx files to see if they have been changed relative to their
-corresponding C++ files. If they have, then runs cython on these files to
-recreate the C++ files.
-
-Additionally, checks pxd files and setup.py if they have been changed. If
-they have, rebuilds everything.
-
-Change detection based on file hashes stored in JSON format.
-
-For now, this script should be run by developers when changing Cython files
-and the resulting C++ files checked in, so that end-users (and Python-only
-developers) do not get the Cython dependencies.
-
-Based upon:
-
-https://raw.github.com/dagss/private-scipy-refactor/cythonize/cythonize.py
-https://raw.githubusercontent.com/numpy/numpy/master/tools/cythonize.py
-
-Note: this script does not check any of the dependent C++ libraries.
-"""
-from __future__ import print_function
-
-import os
-import sys
-import json
-import hashlib
-import subprocess
-import argparse
-
-
-HASH_FILE = "cythonize.json"
-
-
-def process_pyx(fromfile, tofile, language_level="-2"):
-    print("Processing %s" % fromfile)
-    try:
-        from Cython.Compiler.Version import version as cython_version
-        from distutils.version import LooseVersion
-
-        if LooseVersion(cython_version) < LooseVersion("0.19"):
-            raise Exception("Require Cython >= 0.19")
-
-    except ImportError:
-        pass
-
-    flags = ["--fast-fail", language_level]
-    if tofile.endswith(".cpp"):
-        flags += ["--cplus"]
-
-    try:
-        try:
-            r = subprocess.call(
-                ["cython"] + flags + ["-o", tofile, fromfile], env=os.environ
-            )  # See Issue #791
-            if r != 0:
-                raise Exception("Cython failed")
-        except OSError:
-            # There are ways of installing Cython that don't result in a cython
-            # executable on the path, see gh-2397.
-            r = subprocess.call(
-                [
-                    sys.executable,
-                    "-c",
-                    "import sys; from Cython.Compiler.Main import "
-                    "setuptools_main as main; sys.exit(main())",
-                ]
-                + flags
-                + ["-o", tofile, fromfile]
-            )
-            if r != 0:
-                raise Exception("Cython failed")
-    except OSError:
-        raise OSError("Cython needs to be installed")
-
-
-def preserve_cwd(path, func, *args):
-    orig_cwd = os.getcwd()
-    try:
-        os.chdir(path)
-        func(*args)
-    finally:
-        os.chdir(orig_cwd)
-
-
-def load_hashes(filename):
-    try:
-        return json.load(open(filename))
-    except (ValueError, IOError):
-        return {}
-
-
-def save_hashes(hash_db, filename):
-    with open(filename, "w") as f:
-        f.write(json.dumps(hash_db))
-
-
-def get_hash(path):
-    return hashlib.md5(open(path, "rb").read()).hexdigest()
-
-
-def hash_changed(base, path, db):
-    full_path = os.path.normpath(os.path.join(base, path))
-    return not get_hash(full_path) == db.get(full_path)
-
-
-def hash_add(base, path, db):
-    full_path = os.path.normpath(os.path.join(base, path))
-    db[full_path] = get_hash(full_path)
-
-
-def process(base, filename, db):
-    root, ext = os.path.splitext(filename)
-    if ext in [".pyx", ".cpp"]:
-        if hash_changed(base, filename, db) or not os.path.isfile(
-            os.path.join(base, root + ".cpp")
-        ):
-            preserve_cwd(base, process_pyx, root + ".pyx", root + ".cpp")
-            hash_add(base, root + ".cpp", db)
-            hash_add(base, root + ".pyx", db)
-
-
-def check_changes(root, db):
-    res = False
-    new_db = {}
-
-    setup_filename = "setup.py"
-    hash_add(".", setup_filename, new_db)
-    if hash_changed(".", setup_filename, db):
-        res = True
-
-    for base, _, files in os.walk(root):
-        for filename in files:
-            if filename.endswith(".pxd"):
-                hash_add(base, filename, new_db)
-                if hash_changed(base, filename, db):
-                    res = True
-
-    if res:
-        db.clear()
-        db.update(new_db)
-    return res
-
-
-def run(root):
-    db = load_hashes(HASH_FILE)
-
-    try:
-        check_changes(root, db)
-        for base, _, files in os.walk(root):
-            for filename in files:
-                process(base, filename, db)
-    finally:
-        save_hashes(db, HASH_FILE)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Cythonize pyx files into C++ files as needed"
-    )
-    parser.add_argument("root", help="root directory")
-    args = parser.parse_args()
-    run(args.root)

bin/ud/ud_run_test.py

@@ -13,23 +13,12 @@ import srsly
 import spacy
 import spacy.util
 from spacy.tokens import Token, Doc
-from spacy.gold import GoldParse
-from spacy.util import compounding, minibatch_by_words
-from spacy.syntax.nonproj import projectivize
 from spacy.matcher import Matcher
 
-# from spacy.morphology import Fused_begin, Fused_inside
-from spacy import displacy
-from collections import defaultdict, Counter
-from timeit import default_timer as timer
 
 Fused_begin = None
 Fused_inside = None
 
-import itertools
-import random
-import numpy.random
-
 from . import conll17_ud_eval
 
 from spacy import lang
@@ -268,7 +257,7 @@ def load_nlp(experiments_dir, corpus):
     return nlp
 
 
-def initialize_pipeline(nlp, docs, golds, config, device):
+def initialize_pipeline(nlp, examples, config, device):
     nlp.add_pipe(nlp.create_pipe("parser"))
     return nlp
 

bin/ud/ud_train.py

@@ -14,7 +14,7 @@ import spacy
 import spacy.util
 from bin.ud import conll17_ud_eval
 from spacy.tokens import Token, Doc
-from spacy.gold import GoldParse
+from spacy.gold import GoldParse, Example
 from spacy.util import compounding, minibatch, minibatch_by_words
 from spacy.syntax.nonproj import projectivize
 from spacy.matcher import Matcher
@@ -53,7 +53,7 @@ def read_data(
     max_doc_length=None,
     limit=None,
 ):
-    """Read the CONLLU format into (Doc, GoldParse) tuples. If raw_text=True,
+    """Read the CONLLU format into Example objects. If raw_text=True,
     include Doc objects created using nlp.make_doc and then aligned against
     the gold-standard sequences. If oracle_segments=True, include Doc objects
     created from the gold-standard segments. At least one must be True."""
@@ -98,15 +98,16 @@ def read_data(
                 docs.append(doc)
                 golds.append(gold)
                 if limit and len(docs) >= limit:
-                    return docs, golds
+                    return golds_to_gold_data(docs, golds)
 
         if raw_text and sent_annots:
             doc, gold = _make_gold(nlp, None, sent_annots)
             docs.append(doc)
             golds.append(gold)
         if limit and len(docs) >= limit:
-            return docs, golds
+            return golds_to_gold_data(docs, golds)
-    return docs, golds
+    return golds_to_gold_data(docs, golds)
+
 
 def _parse_morph_string(morph_string):
     if morph_string == '_':
@@ -120,6 +121,7 @@ def _parse_morph_string(morph_string):
         output.append('%s_%s' % (key, value.lower()))
     return set(output)
 
+
 def read_conllu(file_):
     docs = []
     sent = []
@@ -180,16 +182,18 @@ def _make_gold(nlp, text, sent_annots, drop_deps=0.0):
 #############################
 
 
-def golds_to_gold_tuples(docs, golds):
+def golds_to_gold_data(docs, golds):
-    """Get out the annoying 'tuples' format used by begin_training, given the
+    """Get out the training data format used by begin_training, given the
     GoldParse objects."""
-    tuples = []
+    data = []
     for doc, gold in zip(docs, golds):
-        text = doc.text
+        example = Example(doc=doc)
-        ids, words, tags, heads, labels, iob = zip(*gold.orig_annot)
+        example.add_doc_annotation(cats=gold.cats)
-        sents = [((ids, words, tags, heads, labels, iob), [])]
+        token_annotation_dict = gold.orig.to_dict()
-        tuples.append((text, sents))
+        example.add_token_annotation(**token_annotation_dict)
-    return tuples
+        example.goldparse = gold
+        data.append(example)
+    return data
 
 
 ##############
@@ -327,7 +331,6 @@ def get_token_conllu(token, i):
     return "\n".join(lines)
 
 
-
 ##################
 # Initialization #
 ##################
@@ -348,7 +351,7 @@ def load_nlp(corpus, config, vectors=None):
     return nlp
 
 
-def initialize_pipeline(nlp, docs, golds, config, device):
+def initialize_pipeline(nlp, examples, config, device):
     nlp.add_pipe(nlp.create_pipe("tagger", config={"set_morphology": False}))
     nlp.add_pipe(nlp.create_pipe("morphologizer"))
     nlp.add_pipe(nlp.create_pipe("parser"))
@@ -356,14 +359,15 @@ def initialize_pipeline(nlp, docs, golds, config, device):
         nlp.parser.add_multitask_objective("tag")
     if config.multitask_sent:
         nlp.parser.add_multitask_objective("sent_start")
-    for gold in golds:
+    for ex in examples:
+        gold = ex.gold
         for tag in gold.tags:
             if tag is not None:
                 nlp.tagger.add_label(tag)
     if torch is not None and device != -1:
         torch.set_default_tensor_type("torch.cuda.FloatTensor")
     optimizer = nlp.begin_training(
-        lambda: golds_to_gold_tuples(docs, golds),
+        lambda: examples,
         device=device,
         subword_features=config.subword_features,
         conv_depth=config.conv_depth,
@@ -491,6 +495,10 @@ def main(
     Token.set_extension("begins_fused", default=False)
     Token.set_extension("inside_fused", default=False)
 
+    Token.set_extension("get_conllu_lines", method=get_token_conllu)
+    Token.set_extension("begins_fused", default=False)
+    Token.set_extension("inside_fused", default=False)
+
     spacy.util.fix_random_seed()
     lang.zh.Chinese.Defaults.use_jieba = False
     lang.ja.Japanese.Defaults.use_janome = False
@@ -505,7 +513,7 @@ def main(
     print("Train and evaluate", corpus, "using lang", paths.lang)
     nlp = load_nlp(paths.lang, config, vectors=vectors_dir)
 
-    docs, golds = read_data(
+    examples = read_data(
         nlp,
         paths.train.conllu.open(encoding="utf8"),
         paths.train.text.open(encoding="utf8"),
@@ -513,12 +521,12 @@ def main(
         limit=limit,
     )
 
-    optimizer = initialize_pipeline(nlp, docs, golds, config, gpu_device)
+    optimizer = initialize_pipeline(nlp, examples, config, gpu_device)
 
     batch_sizes = compounding(config.min_batch_size, config.max_batch_size, 1.001)
     beam_prob = compounding(0.2, 0.8, 1.001)
     for i in range(config.nr_epoch):
-        docs, golds = read_data(
+        examples = read_data(
             nlp,
             paths.train.conllu.open(encoding="utf8"),
             paths.train.text.open(encoding="utf8"),
@@ -527,22 +535,19 @@ def main(
             oracle_segments=use_oracle_segments,
             raw_text=not use_oracle_segments,
         )
-        Xs = list(zip(docs, golds))
+        random.shuffle(examples)
-        random.shuffle(Xs)
         if config.batch_by_words:
-            batches = minibatch_by_words(Xs, size=batch_sizes)
+            batches = minibatch_by_words(examples, size=batch_sizes)
         else:
-            batches = minibatch(Xs, size=batch_sizes)
+            batches = minibatch(examples, size=batch_sizes)
         losses = {}
-        n_train_words = sum(len(doc) for doc in docs)
+        n_train_words = sum(len(ex.doc) for ex in examples)
         with tqdm.tqdm(total=n_train_words, leave=False) as pbar:
             for batch in batches:
-                batch_docs, batch_gold = zip(*batch)
+                pbar.update(sum(len(ex.doc) for ex in batch))
-                pbar.update(sum(len(doc) for doc in batch_docs))
                 nlp.parser.cfg["beam_update_prob"] = next(beam_prob)
                 nlp.update(
-                    batch_docs,
+                    batch,
-                    batch_gold,
                     sgd=optimizer,
                     drop=config.dropout,
                     losses=losses,

bin/wiki_entity_linking/kb_creator.py

@@ -46,7 +46,7 @@ def _define_entities(nlp, kb, entity_def_path, entity_descr_path, min_entity_fre
             " cf. https://spacy.io/usage/models#languages."
         )
 
-    logger.info("Filtering entities with fewer than {} mentions".format(min_entity_freq))
+    logger.info("Filtering entities with fewer than {} mentions or no description".format(min_entity_freq))
     entity_frequencies = io.read_entity_to_count(entity_freq_path)
     # filter the entities for in the KB by frequency, because there's just too much data (8M entities) otherwise
     filtered_title_to_id, entity_list, description_list, frequency_list = get_filtered_entities(

bin/wiki_entity_linking/train_descriptions.py

@@ -4,12 +4,12 @@ from random import shuffle
 import logging
 import numpy as np
 
-from spacy._ml import zero_init, create_default_optimizer
+from thinc.model import Model
-from spacy.cli.pretrain import get_cossim_loss
-
-from thinc.v2v import Model
 from thinc.api import chain
-from thinc.neural._classes.affine import Affine
+from thinc.loss import CosineDistance
+from thinc.layers import Linear
+
+from spacy.util import create_default_optimizer
 
 logger = logging.getLogger(__name__)
 
@@ -34,6 +34,7 @@ class EntityEncoder:
         self.input_dim = input_dim
         self.desc_width = desc_width
         self.epochs = epochs
+        self.distance = CosineDistance(ignore_zeros=True, normalize=False)
 
     def apply_encoder(self, description_list):
         if self.encoder is None:
@@ -132,21 +133,17 @@ class EntityEncoder:
     def _build_network(self, orig_width, hidden_with):
         with Model.define_operators({">>": chain}):
             # very simple encoder-decoder model
-            self.encoder = Affine(hidden_with, orig_width)
+            self.encoder = Linear(hidden_with, orig_width)
-            self.model = self.encoder >> zero_init(
+            # TODO: removed the zero_init here - is oK?
-                Affine(orig_width, hidden_with, drop_factor=0.0)
+            self.model = self.encoder >> Linear(orig_width, hidden_with)
-            )
+        self.sgd = create_default_optimizer()
-        self.sgd = create_default_optimizer(self.model.ops)
 
     def _update(self, vectors):
+        truths = self.model.ops.asarray(vectors)
         predictions, bp_model = self.model.begin_update(
-            np.asarray(vectors), drop=self.DROP
+            truths, drop=self.DROP
         )
-        loss, d_scores = self._get_loss(scores=predictions, golds=np.asarray(vectors))
+        d_scores, loss = self.distance(predictions, truths)
         bp_model(d_scores, sgd=self.sgd)
         return loss / len(vectors)
 
-    @staticmethod
-    def _get_loss(golds, scores):
-        loss, gradients = get_cossim_loss(scores, golds)
-        return loss, gradients

bin/wiki_entity_linking/wikidata_train_entity_linker.py

@@ -17,7 +17,13 @@ import plac
 from tqdm import tqdm
 
 from bin.wiki_entity_linking import wikipedia_processor
-from bin.wiki_entity_linking import TRAINING_DATA_FILE, KB_MODEL_DIR, KB_FILE, LOG_FORMAT, OUTPUT_MODEL_DIR
+from bin.wiki_entity_linking import (
+    TRAINING_DATA_FILE,
+    KB_MODEL_DIR,
+    KB_FILE,
+    LOG_FORMAT,
+    OUTPUT_MODEL_DIR,
+)
 from bin.wiki_entity_linking.entity_linker_evaluation import measure_performance
 from bin.wiki_entity_linking.kb_creator import read_kb
 
@@ -48,10 +54,12 @@ def main(
     l2=1e-6,
     train_articles=None,
     dev_articles=None,
-    labels_discard=None
+    labels_discard=None,
 ):
     if not output_dir:
-        logger.warning("No output dir specified so no results will be written, are you sure about this ?")
+        logger.warning(
+            "No output dir specified so no results will be written, are you sure about this ?"
+        )
 
     logger.info("Creating Entity Linker with Wikipedia and WikiData")
 
@@ -68,7 +76,11 @@ def main(
     # STEP 1 : load the NLP object
     logger.info("STEP 1a: Loading model from {}".format(nlp_dir))
     nlp = spacy.load(nlp_dir)
-    logger.info("Original NLP pipeline has following pipeline components: {}".format(nlp.pipe_names))
+    logger.info(
+        "Original NLP pipeline has following pipeline components: {}".format(
+            nlp.pipe_names
+        )
+    )
 
     # check that there is a NER component in the pipeline
     if "ner" not in nlp.pipe_names:
@@ -79,25 +91,42 @@ def main(
 
     # STEP 2: read the training dataset previously created from WP
     logger.info("STEP 2: Reading training & dev dataset from {}".format(training_path))
-    train_indices, dev_indices = wikipedia_processor.read_training_indices(training_path)
+    train_indices, dev_indices = wikipedia_processor.read_training_indices(
-    logger.info("Training set has {} articles, limit set to roughly {} articles per epoch"
+        training_path
-                .format(len(train_indices), train_articles if train_articles else "all"))
+    )
-    logger.info("Dev set has {} articles, limit set to rougly {} articles for evaluation"
+    logger.info(
-                .format(len(dev_indices), dev_articles if dev_articles else "all"))
+        "Training set has {} articles, limit set to roughly {} articles per epoch".format(
+            len(train_indices), train_articles if train_articles else "all"
+        )
+    )
+    logger.info(
+        "Dev set has {} articles, limit set to rougly {} articles for evaluation".format(
+            len(dev_indices), dev_articles if dev_articles else "all"
+        )
+    )
     if dev_articles:
         dev_indices = dev_indices[0:dev_articles]
 
     # STEP 3: create and train an entity linking pipe
-    logger.info("STEP 3: Creating and training an Entity Linking pipe for {} epochs".format(epochs))
+    logger.info(
+        "STEP 3: Creating and training an Entity Linking pipe for {} epochs".format(
+            epochs
+        )
+    )
     if labels_discard:
         labels_discard = [x.strip() for x in labels_discard.split(",")]
-        logger.info("Discarding {} NER types: {}".format(len(labels_discard), labels_discard))
+        logger.info(
+            "Discarding {} NER types: {}".format(len(labels_discard), labels_discard)
+        )
     else:
         labels_discard = []
 
     el_pipe = nlp.create_pipe(
-        name="entity_linker", config={"pretrained_vectors": nlp.vocab.vectors.name,
+        name="entity_linker",
-                                      "labels_discard": labels_discard}
+        config={
+            "pretrained_vectors": nlp.vocab.vectors,
+            "labels_discard": labels_discard,
+        },
     )
     el_pipe.set_kb(kb)
     nlp.add_pipe(el_pipe, last=True)
@@ -109,11 +138,18 @@ def main(
         optimizer.L2 = l2
 
     logger.info("Dev Baseline Accuracies:")
-    dev_data = wikipedia_processor.read_el_docs_golds(nlp=nlp, entity_file_path=training_path,
+    dev_data = wikipedia_processor.read_el_docs_golds(
-                                                      dev=True, line_ids=dev_indices,
+        nlp=nlp,
-                                                      kb=kb, labels_discard=labels_discard)
+        entity_file_path=training_path,
+        dev=True,
+        line_ids=dev_indices,
+        kb=kb,
+        labels_discard=labels_discard,
+    )
 
-    measure_performance(dev_data, kb, el_pipe, baseline=True, context=False, dev_limit=len(dev_indices))
+    measure_performance(
+        dev_data, kb, el_pipe, baseline=True, context=False, dev_limit=len(dev_indices)
+    )
 
     for itn in range(epochs):
         random.shuffle(train_indices)
@@ -127,13 +163,18 @@ def main(
         if train_articles:
             bar_total = train_articles
 
-        with tqdm(total=bar_total, leave=False, desc='Epoch ' + str(itn)) as pbar:
+        with tqdm(total=bar_total, leave=False, desc=f"Epoch {itn}") as pbar:
             for batch in batches:
                 if not train_articles or articles_processed < train_articles:
                     with nlp.disable_pipes("entity_linker"):
-                        train_batch = wikipedia_processor.read_el_docs_golds(nlp=nlp, entity_file_path=training_path,
+                        train_batch = wikipedia_processor.read_el_docs_golds(
-                                                                             dev=False, line_ids=batch,
+                            nlp=nlp,
-                                                                             kb=kb, labels_discard=labels_discard)
+                            entity_file_path=training_path,
+                            dev=False,
+                            line_ids=batch,
+                            kb=kb,
+                            labels_discard=labels_discard,
+                        )
                         docs, golds = zip(*train_batch)
                     try:
                         with nlp.disable_pipes(*other_pipes):
@@ -150,17 +191,36 @@ def main(
                     except Exception as e:
                         logger.error("Error updating batch:" + str(e))
         if batchnr > 0:
-            logging.info("Epoch {} trained on {} articles, train loss {}"
+            logging.info(
-                         .format(itn, articles_processed, round(losses["entity_linker"] / batchnr, 2)))
+                "Epoch {} trained on {} articles, train loss {}".format(
+                    itn, articles_processed, round(losses["entity_linker"] / batchnr, 2)
+                )
+            )
             # re-read the dev_data (data is returned as a generator)
-            dev_data = wikipedia_processor.read_el_docs_golds(nlp=nlp, entity_file_path=training_path,
+            dev_data = wikipedia_processor.read_el_docs_golds(
-                                                              dev=True, line_ids=dev_indices,
+                nlp=nlp,
-                                                              kb=kb, labels_discard=labels_discard)
+                entity_file_path=training_path,
-            measure_performance(dev_data, kb, el_pipe, baseline=False, context=True, dev_limit=len(dev_indices))
+                dev=True,
+                line_ids=dev_indices,
+                kb=kb,
+                labels_discard=labels_discard,
+            )
+            measure_performance(
+                dev_data,
+                kb,
+                el_pipe,
+                baseline=False,
+                context=True,
+                dev_limit=len(dev_indices),
+            )
 
     if output_dir:
         # STEP 4: write the NLP pipeline (now including an EL model) to file
-        logger.info("Final NLP pipeline has following pipeline components: {}".format(nlp.pipe_names))
+        logger.info(
+            "Final NLP pipeline has following pipeline components: {}".format(
+                nlp.pipe_names
+            )
+        )
         logger.info("STEP 4: Writing trained NLP to {}".format(nlp_output_dir))
         nlp.to_disk(nlp_output_dir)
 

examples/deep_learning_keras.py

@@ -14,7 +14,7 @@ pip install keras==2.0.9
 
 Compatible with: spaCy v2.0.0+
 """
-
+import ml_datasets
 import plac
 import random
 import pathlib
@@ -24,7 +24,6 @@ from keras.models import Sequential, model_from_json
 from keras.layers import LSTM, Dense, Embedding, Bidirectional
 from keras.layers import TimeDistributed
 from keras.optimizers import Adam
-import thinc.extra.datasets
 from spacy.compat import pickle
 import spacy
 
@@ -224,7 +223,7 @@ def main(
     if model_dir is not None:
         model_dir = pathlib.Path(model_dir)
     if train_dir is None or dev_dir is None:
-        imdb_data = thinc.extra.datasets.imdb()
+        imdb_data = ml_datasets.imdb()
     if is_runtime:
         if dev_dir is None:
             dev_texts, dev_labels = zip(*imdb_data[1])

examples/experiments/ptb-joint-pos-dep/bilstm_tok2vec.cfg

@@ -0,0 +1,63 @@
+[training]
+patience = 10000
+eval_frequency = 200
+dropout = 0.2
+init_tok2vec = null
+vectors = null
+max_epochs = 100
+orth_variant_level = 0.0
+gold_preproc = true
+max_length = 0
+use_gpu = 0
+scores = ["tags_acc", "uas", "las"]
+score_weights = {"las": 0.8, "tags_acc": 0.2}
+limit = 0
+
+[training.batch_size]
+@schedules = "compounding.v1"
+start = 100
+stop = 1000
+compound = 1.001
+
+[optimizer]
+@optimizers = "Adam.v1"
+learn_rate = 0.001
+beta1 = 0.9
+beta2 = 0.999
+
+[nlp]
+lang = "en"
+vectors = ${training:vectors}
+
+[nlp.pipeline.tok2vec]
+factory = "tok2vec"
+
+[nlp.pipeline.tagger]
+factory = "tagger"
+
+[nlp.pipeline.parser]
+factory = "parser"
+
+[nlp.pipeline.tagger.model]
+@architectures = "tagger_model.v1"
+
+[nlp.pipeline.tagger.model.tok2vec]
+@architectures = "tok2vec_tensors.v1"
+width = ${nlp.pipeline.tok2vec.model:width}
+
+[nlp.pipeline.parser.model]
+@architectures = "transition_based_parser.v1"
+nr_feature_tokens = 8
+hidden_width = 64
+maxout_pieces = 3
+
+[nlp.pipeline.parser.model.tok2vec]
+@architectures = "tok2vec_tensors.v1"
+width = ${nlp.pipeline.tok2vec.model:width}
+
+[nlp.pipeline.tok2vec.model]
+@architectures = "hash_embed_bilstm.v1"
+pretrained_vectors = ${nlp:vectors}
+width = 96
+depth = 4
+embed_size = 2000

examples/experiments/ptb-joint-pos-dep/defaults.cfg

@@ -0,0 +1,65 @@
+[training]
+patience = 10000
+eval_frequency = 200
+dropout = 0.2
+init_tok2vec = null
+vectors = null
+max_epochs = 100
+orth_variant_level = 0.0
+gold_preproc = true
+max_length = 0
+use_gpu = -1
+scores = ["tags_acc", "uas", "las"]
+score_weights = {"las": 0.8, "tags_acc": 0.2}
+limit = 0
+
+[training.batch_size]
+@schedules = "compounding.v1"
+start = 100
+stop = 1000
+compound = 1.001
+
+[optimizer]
+@optimizers = "Adam.v1"
+learn_rate = 0.001
+beta1 = 0.9
+beta2 = 0.999
+
+[nlp]
+lang = "en"
+vectors = ${training:vectors}
+
+[nlp.pipeline.tok2vec]
+factory = "tok2vec"
+
+[nlp.pipeline.tagger]
+factory = "tagger"
+
+[nlp.pipeline.parser]
+factory = "parser"
+
+[nlp.pipeline.tagger.model]
+@architectures = "tagger_model.v1"
+
+[nlp.pipeline.tagger.model.tok2vec]
+@architectures = "tok2vec_tensors.v1"
+width = ${nlp.pipeline.tok2vec.model:width}
+
+[nlp.pipeline.parser.model]
+@architectures = "transition_based_parser.v1"
+nr_feature_tokens = 8
+hidden_width = 64
+maxout_pieces = 3
+
+[nlp.pipeline.parser.model.tok2vec]
+@architectures = "tok2vec_tensors.v1"
+width = ${nlp.pipeline.tok2vec.model:width}
+
+[nlp.pipeline.tok2vec.model]
+@architectures = "hash_embed_cnn.v1"
+pretrained_vectors = ${nlp:vectors}
+width = 96
+depth = 4
+window_size = 1
+embed_size = 2000
+maxout_pieces = 3

examples/pipeline/multi_processing.py

@@ -13,9 +13,10 @@ Prerequisites: pip install joblib
 from __future__ import print_function, unicode_literals
 
 from pathlib import Path
+
+import ml_datasets
 from joblib import Parallel, delayed
 from functools import partial
-import thinc.extra.datasets
 import plac
 import spacy
 from spacy.util import minibatch
@@ -35,7 +36,7 @@ def main(output_dir, model="en_core_web_sm", n_jobs=4, batch_size=1000, limit=10
         output_dir.mkdir()
     # load and pre-process the IMBD dataset
     print("Loading IMDB data...")
-    data, _ = thinc.extra.datasets.imdb()
+    data, _ = ml_datasets.imdb()
     texts, _ = zip(*data[-limit:])
     print("Processing texts...")
     partitions = minibatch(texts, size=batch_size)

examples/streamlit_spacy.py

@@ -26,12 +26,12 @@ DEFAULT_TEXT = "Mark Zuckerberg is the CEO of Facebook."
 HTML_WRAPPER = """<div style="overflow-x: auto; border: 1px solid #e6e9ef; border-radius: 0.25rem; padding: 1rem; margin-bottom: 2.5rem">{}</div>"""
 
 
-@st.cache(ignore_hash=True)
+@st.cache(allow_output_mutation=True)
 def load_model(name):
     return spacy.load(name)
 
 
-@st.cache(ignore_hash=True)
+@st.cache(allow_output_mutation=True)
 def process_text(model_name, text):
     nlp = load_model(model_name)
     return nlp(text)
@@ -79,7 +79,9 @@ if "ner" in nlp.pipe_names:
     st.header("Named Entities")
     st.sidebar.header("Named Entities")
     label_set = nlp.get_pipe("ner").labels
-    labels = st.sidebar.multiselect("Entity labels", label_set, label_set)
+    labels = st.sidebar.multiselect(
+        "Entity labels", options=label_set, default=list(label_set)
+    )
     html = displacy.render(doc, style="ent", options={"ents": labels})
     # Newlines seem to mess with the rendering
     html = html.replace("\n", " ")

examples/training/conllu.py

@@ -12,7 +12,7 @@ import tqdm
 import spacy
 import spacy.util
 from spacy.tokens import Token, Doc
-from spacy.gold import GoldParse
+from spacy.gold import GoldParse, Example
 from spacy.syntax.nonproj import projectivize
 from collections import defaultdict
 from spacy.matcher import Matcher
@@ -33,25 +33,25 @@ random.seed(0)
 numpy.random.seed(0)
 
 
-def minibatch_by_words(items, size=5000):
+def minibatch_by_words(examples, size=5000):
-    random.shuffle(items)
+    random.shuffle(examples)
     if isinstance(size, int):
         size_ = itertools.repeat(size)
     else:
         size_ = size
-    items = iter(items)
+    examples = iter(examples)
     while True:
         batch_size = next(size_)
         batch = []
         while batch_size >= 0:
             try:
-                doc, gold = next(items)
+                example = next(examples)
             except StopIteration:
                 if batch:
                     yield batch
                 return
-            batch_size -= len(doc)
+            batch_size -= len(example.doc)
-            batch.append((doc, gold))
+            batch.append(example)
         if batch:
             yield batch
         else:
@@ -78,7 +78,7 @@ def read_data(
     max_doc_length=None,
     limit=None,
 ):
-    """Read the CONLLU format into (Doc, GoldParse) tuples. If raw_text=True,
+    """Read the CONLLU format into Example objects. If raw_text=True,
     include Doc objects created using nlp.make_doc and then aligned against
     the gold-standard sequences. If oracle_segments=True, include Doc objects
     created from the gold-standard segments. At least one must be True."""
@@ -119,15 +119,15 @@ def read_data(
                 docs.append(doc)
                 golds.append(gold)
                 if limit and len(docs) >= limit:
-                    return docs, golds
+                    return golds_to_gold_data(docs, golds)
 
         if raw_text and sent_annots:
             doc, gold = _make_gold(nlp, None, sent_annots)
             docs.append(doc)
             golds.append(gold)
         if limit and len(docs) >= limit:
-            return docs, golds
+            return golds_to_gold_data(docs, golds)
-    return docs, golds
+    return golds_to_gold_data(docs, golds)
 
 
 def read_conllu(file_):
@@ -181,16 +181,18 @@ def _make_gold(nlp, text, sent_annots):
 #############################
 
 
-def golds_to_gold_tuples(docs, golds):
+def golds_to_gold_data(docs, golds):
-    """Get out the annoying 'tuples' format used by begin_training, given the
+    """Get out the training data format used by begin_training, given the
     GoldParse objects."""
-    tuples = []
+    data = []
     for doc, gold in zip(docs, golds):
-        text = doc.text
+        example = Example(doc=doc)
-        ids, words, tags, heads, labels, iob = zip(*gold.orig_annot)
+        example.add_doc_annotation(cats=gold.cats)
-        sents = [((ids, words, tags, heads, labels, iob), [])]
+        token_annotation_dict = gold.orig.to_dict()
-        tuples.append((text, sents))
+        example.add_token_annotation(**token_annotation_dict)
-    return tuples
+        example.goldparse = gold
+        data.append(example)
+    return data
 
 
 ##############
@@ -303,7 +305,7 @@ def load_nlp(corpus, config):
     return nlp
 
 
-def initialize_pipeline(nlp, docs, golds, config):
+def initialize_pipeline(nlp, examples, config):
     nlp.add_pipe(nlp.create_pipe("parser"))
     if config.multitask_tag:
         nlp.parser.add_multitask_objective("tag")
@@ -311,18 +313,19 @@ def initialize_pipeline(nlp, docs, golds, config):
         nlp.parser.add_multitask_objective("sent_start")
     nlp.parser.moves.add_action(2, "subtok")
     nlp.add_pipe(nlp.create_pipe("tagger"))
-    for gold in golds:
+    for ex in examples:
-        for tag in gold.tags:
+        for tag in ex.gold.tags:
             if tag is not None:
                 nlp.tagger.add_label(tag)
     # Replace labels that didn't make the frequency cutoff
     actions = set(nlp.parser.labels)
     label_set = set([act.split("-")[1] for act in actions if "-" in act])
-    for gold in golds:
+    for ex in examples:
+        gold = ex.gold
         for i, label in enumerate(gold.labels):
             if label is not None and label not in label_set:
                 gold.labels[i] = label.split("||")[0]
-    return nlp.begin_training(lambda: golds_to_gold_tuples(docs, golds))
+    return nlp.begin_training(lambda: examples)
 
 
 ########################
@@ -391,13 +394,17 @@ def main(ud_dir, parses_dir, config, corpus, limit=0):
     Token.set_extension("begins_fused", default=False)
     Token.set_extension("inside_fused", default=False)
 
+    Token.set_extension("get_conllu_lines", method=get_token_conllu)
+    Token.set_extension("begins_fused", default=False)
+    Token.set_extension("inside_fused", default=False)
+
     paths = TreebankPaths(ud_dir, corpus)
     if not (parses_dir / corpus).exists():
         (parses_dir / corpus).mkdir()
     print("Train and evaluate", corpus, "using lang", paths.lang)
     nlp = load_nlp(paths.lang, config)
 
-    docs, golds = read_data(
+    examples = read_data(
         nlp,
         paths.train.conllu.open(encoding="utf8"),
         paths.train.text.open(encoding="utf8"),
@@ -405,23 +412,18 @@ def main(ud_dir, parses_dir, config, corpus, limit=0):
         limit=limit,
     )
 
-    optimizer = initialize_pipeline(nlp, docs, golds, config)
+    optimizer = initialize_pipeline(nlp, examples, config)
 
     for i in range(config.nr_epoch):
-        docs = [nlp.make_doc(doc.text) for doc in docs]
+        docs = [nlp.make_doc(example.doc.text) for example in examples]
-        batches = minibatch_by_words(list(zip(docs, golds)), size=config.batch_size)
+        batches = minibatch_by_words(examples, size=config.batch_size)
         losses = {}
         n_train_words = sum(len(doc) for doc in docs)
         with tqdm.tqdm(total=n_train_words, leave=False) as pbar:
             for batch in batches:
-                batch_docs, batch_gold = zip(*batch)
+                pbar.update(sum(len(ex.doc) for ex in batch))
-                pbar.update(sum(len(doc) for doc in batch_docs))
                 nlp.update(
-                    batch_docs,
+                    examples=batch, sgd=optimizer, drop=config.dropout, losses=losses,
-                    batch_gold,
-                    sgd=optimizer,
-                    drop=config.dropout,
-                    losses=losses,
                 )
 
         out_path = parses_dir / corpus / "epoch-{i}.conllu".format(i=i)

examples/training/ner_multitask_objective.py

@@ -31,14 +31,13 @@ random.seed(0)
 
 PWD = os.path.dirname(__file__)
 
-TRAIN_DATA = list(read_json_file(
+TRAIN_DATA = list(read_json_file(os.path.join(PWD, "training-data.json")))
-    os.path.join(PWD, "ner_example_data", "ner-sent-per-line.json")))
 
 
-def get_position_label(i, words, tags, heads, labels, ents):
+def get_position_label(i, token_annotation):
     """Return labels indicating the position of the word in the document.
     """
-    if len(words) < 20:
+    if len(token_annotation.words) < 20:
         return "short-doc"
     elif i == 0:
         return "first-word"
@@ -46,7 +45,7 @@ def get_position_label(i, words, tags, heads, labels, ents):
         return "early-word"
     elif i < 20:
         return "mid-word"
-    elif i == len(words) - 1:
+    elif i == len(token_annotation.words) - 1:
         return "last-word"
     else:
         return "late-word"
@@ -60,17 +59,17 @@ def main(n_iter=10):
     print(nlp.pipeline)
 
     print("Create data", len(TRAIN_DATA))
-    optimizer = nlp.begin_training(get_gold_tuples=lambda: TRAIN_DATA)
+    optimizer = nlp.begin_training(get_examples=lambda: TRAIN_DATA)
     for itn in range(n_iter):
         random.shuffle(TRAIN_DATA)
         losses = {}
-        for text, annot_brackets in TRAIN_DATA:
+        for example in TRAIN_DATA:
-            for annotations, _ in annot_brackets:
+            for token_annotation in example.token_annotations:
-                doc = Doc(nlp.vocab, words=annotations[1])
+                doc = Doc(nlp.vocab, words=token_annotation.words)
-                gold = GoldParse.from_annot_tuples(doc, annotations)
+                gold = GoldParse.from_annotation(doc, example.doc_annotation, token_annotation)
+
                 nlp.update(
-                    [doc],  # batch of texts
+                    examples=[(doc, gold)],  # 1 example
-                    [gold],  # batch of annotations
                     drop=0.2,  # dropout - make it harder to memorise data
                     sgd=optimizer,  # callable to update weights
                     losses=losses,
@@ -78,9 +77,9 @@ def main(n_iter=10):
         print(losses.get("nn_labeller", 0.0), losses["ner"])
 
     # test the trained model
-    for text, _ in TRAIN_DATA:
+    for example in TRAIN_DATA:
-        if text is not None:
+        if example.text is not None:
-            doc = nlp(text)
+            doc = nlp(example.text)
             print("Entities", [(ent.text, ent.label_) for ent in doc.ents])
             print("Tokens", [(t.text, t.ent_type_, t.ent_iob) for t in doc])
 

examples/training/pretrain_textcat.py

@@ -16,16 +16,18 @@ the development labels, after all --- only the unlabelled text.
 import plac
 import tqdm
 import random
+
+import ml_datasets
+
 import spacy
-import thinc.extra.datasets
 from spacy.util import minibatch, use_gpu, compounding
-from spacy._ml import Tok2Vec
 from spacy.pipeline import TextCategorizer
+from spacy.ml.tok2vec import Tok2Vec
 import numpy
 
 
 def load_texts(limit=0):
-    train, dev = thinc.extra.datasets.imdb()
+    train, dev = ml_datasets.imdb()
     train_texts, train_labels = zip(*train)
     dev_texts, dev_labels = zip(*train)
     train_texts = list(train_texts)
@@ -41,7 +43,7 @@ def load_texts(limit=0):
 def load_textcat_data(limit=0):
     """Load data from the IMDB dataset."""
     # Partition off part of the train data for evaluation
-    train_data, eval_data = thinc.extra.datasets.imdb()
+    train_data, eval_data = ml_datasets.imdb()
     random.shuffle(train_data)
     train_data = train_data[-limit:]
     texts, labels = zip(*train_data)
@@ -63,17 +65,15 @@ def prefer_gpu():
 
 
 def build_textcat_model(tok2vec, nr_class, width):
-    from thinc.v2v import Model, Softmax, Maxout
+    from thinc.model import Model
-    from thinc.api import flatten_add_lengths, chain
+    from thinc.layers import Softmax, chain, reduce_mean
-    from thinc.t2v import Pooling, sum_pool, mean_pool, max_pool
+    from thinc.layers import list2ragged
-    from thinc.misc import Residual, LayerNorm
-    from spacy._ml import logistic, zero_init
 
     with Model.define_operators({">>": chain}):
         model = (
             tok2vec
-            >> flatten_add_lengths
+            >> list2ragged()
-            >> Pooling(mean_pool)
+            >> reduce_mean()
             >> Softmax(nr_class, width)
         )
     model.tok2vec = tok2vec
@@ -81,7 +81,7 @@ def build_textcat_model(tok2vec, nr_class, width):
 
 
 def block_gradients(model):
-    from thinc.api import wrap
+    from thinc.api import wrap  # TODO FIX
 
     def forward(X, drop=0.0):
         Y, _ = model.begin_update(X, drop=drop)
@@ -114,7 +114,7 @@ def train_tensorizer(nlp, texts, dropout, n_iter):
         losses = {}
         for i, batch in enumerate(minibatch(tqdm.tqdm(texts))):
             docs = [nlp.make_doc(text) for text in batch]
-            tensorizer.update(docs, None, losses=losses, sgd=optimizer, drop=dropout)
+            tensorizer.update((docs, None), losses=losses, sgd=optimizer, drop=dropout)
         print(losses)
     return optimizer
 
@@ -143,8 +143,7 @@ def train_textcat(nlp, n_texts, n_iter=10):
             # batch up the examples using spaCy's minibatch
             batches = minibatch(tqdm.tqdm(train_data), size=2)
             for batch in batches:
-                texts, annotations = zip(*batch)
+                nlp.update(batch, sgd=optimizer, drop=0.2, losses=losses)
-                nlp.update(texts, annotations, sgd=optimizer, drop=0.2, losses=losses)
             with textcat.model.use_params(optimizer.averages):
                 # evaluate on the dev data split off in load_data()
                 scores = evaluate_textcat(nlp.tokenizer, textcat, dev_texts, dev_cats)

examples/training/rehearsal.py

@@ -58,7 +58,7 @@ def main(model_name, unlabelled_loc):
     # yet, but I'm getting weird results from Adam. Try commenting out the
     # nlp.update(), and using Adam -- you'll find the models drift apart.
     # I guess Adam is losing precision, introducing gradient noise?
-    optimizer.alpha = 0.1
+    optimizer.learn_rate = 0.1
     optimizer.b1 = 0.0
     optimizer.b2 = 0.0
 
@@ -75,8 +75,7 @@ def main(model_name, unlabelled_loc):
             # batch up the examples using spaCy's minibatch
             raw_batches = minibatch(raw_docs, size=4)
             for batch in minibatch(TRAIN_DATA, size=sizes):
-                docs, golds = zip(*batch)
+                nlp.update(batch, sgd=optimizer, drop=dropout, losses=losses)
-                nlp.update(docs, golds, sgd=optimizer, drop=dropout, losses=losses)
                 raw_batch = list(next(raw_batches))
                 nlp.rehearse(raw_batch, sgd=optimizer, losses=r_losses)
             print("Losses", losses)

examples/training/train_entity_linker.py

@@ -17,7 +17,7 @@ import plac
 import random
 from pathlib import Path
 
-from spacy.symbols import PERSON
+import srsly
 from spacy.vocab import Vocab
 
 import spacy
@@ -68,7 +68,7 @@ def main(kb_path, vocab_path=None, output_dir=None, n_iter=50):
     vocab = Vocab().from_disk(vocab_path)
     # create blank Language class with correct vocab
     nlp = spacy.blank("en", vocab=vocab)
-    nlp.vocab.vectors.name = "spacy_pretrained_vectors"
+    nlp.vocab.vectors.name = "nel_vectors"
     print("Created blank 'en' model with vocab from '%s'" % vocab_path)
 
     # Add a sentencizer component. Alternatively, add a dependency parser for higher accuracy.
@@ -93,7 +93,7 @@ def main(kb_path, vocab_path=None, output_dir=None, n_iter=50):
         nlp.add_pipe(entity_linker, last=True)
 
     # Convert the texts to docs to make sure we have doc.ents set for the training examples.
-    # Also ensure that the annotated examples correspond to known identifiers in the knowlege base.
+    # Also ensure that the annotated examples correspond to known identifiers in the knowledge base.
     kb_ids = nlp.get_pipe("entity_linker").kb.get_entity_strings()
     TRAIN_DOCS = []
     for text, annotation in TRAIN_DATA:
@@ -118,16 +118,15 @@ def main(kb_path, vocab_path=None, output_dir=None, n_iter=50):
     with nlp.disable_pipes(*other_pipes):  # only train entity linker
         # reset and initialize the weights randomly
         optimizer = nlp.begin_training()
+
         for itn in range(n_iter):
             random.shuffle(TRAIN_DOCS)
             losses = {}
             # batch up the examples using spaCy's minibatch
             batches = minibatch(TRAIN_DOCS, size=compounding(4.0, 32.0, 1.001))
             for batch in batches:
-                texts, annotations = zip(*batch)
                 nlp.update(
-                    texts,  # batch of texts
+                    batch,
-                    annotations,  # batch of annotations
                     drop=0.2,  # dropout - make it harder to memorise data
                     losses=losses,
                     sgd=optimizer,

examples/training/train_intent_parser.py

@@ -134,8 +134,7 @@ def main(model=None, output_dir=None, n_iter=15):
             # batch up the examples using spaCy's minibatch
             batches = minibatch(TRAIN_DATA, size=compounding(4.0, 32.0, 1.001))
             for batch in batches:
-                texts, annotations = zip(*batch)
+                nlp.update(batch, sgd=optimizer, losses=losses)
-                nlp.update(texts, annotations, sgd=optimizer, losses=losses)
             print("Losses", losses)
 
     # test the trained model

examples/training/train_ner.py

@@ -68,10 +68,8 @@ def main(model=None, output_dir=None, n_iter=100):
             # batch up the examples using spaCy's minibatch
             batches = minibatch(TRAIN_DATA, size=compounding(4.0, 32.0, 1.001))
             for batch in batches:
-                texts, annotations = zip(*batch)
                 nlp.update(
-                    texts,  # batch of texts
+                    batch,
-                    annotations,  # batch of annotations
                     drop=0.5,  # dropout - make it harder to memorise data
                     losses=losses,
                 )

examples/training/train_new_entity_type.py

@@ -105,8 +105,7 @@ def main(model=None, new_model_name="animal", output_dir=None, n_iter=30):
             batches = minibatch(TRAIN_DATA, size=sizes)
             losses = {}
             for batch in batches:
-                texts, annotations = zip(*batch)
+                nlp.update(batch, sgd=optimizer, drop=0.35, losses=losses)
-                nlp.update(texts, annotations, sgd=optimizer, drop=0.35, losses=losses)
             print("Losses", losses)
 
     # test the trained model

examples/training/train_parser.py

@@ -75,8 +75,7 @@ def main(model=None, output_dir=None, n_iter=15):
             # batch up the examples using spaCy's minibatch
             batches = minibatch(TRAIN_DATA, size=compounding(4.0, 32.0, 1.001))
             for batch in batches:
-                texts, annotations = zip(*batch)
+                nlp.update(batch, sgd=optimizer, losses=losses)
-                nlp.update(texts, annotations, sgd=optimizer, losses=losses)
             print("Losses", losses)
 
     # test the trained model

examples/training/train_tagger.py

@@ -65,8 +65,7 @@ def main(lang="en", output_dir=None, n_iter=25):
         # batch up the examples using spaCy's minibatch
         batches = minibatch(TRAIN_DATA, size=compounding(4.0, 32.0, 1.001))
         for batch in batches:
-            texts, annotations = zip(*batch)
+            nlp.update(batch, sgd=optimizer, losses=losses)
-            nlp.update(texts, annotations, sgd=optimizer, losses=losses)
         print("Losses", losses)
 
     # test the trained model

examples/training/train_textcat.py

@@ -10,10 +10,11 @@ see the documentation:
 Compatible with: spaCy v2.0.0+
 """
 from __future__ import unicode_literals, print_function
+
+import ml_datasets
 import plac
 import random
 from pathlib import Path
-import thinc.extra.datasets
 
 import spacy
 from spacy.util import minibatch, compounding
@@ -83,8 +84,7 @@ def main(model=None, output_dir=None, n_iter=20, n_texts=2000, init_tok2vec=None
             random.shuffle(train_data)
             batches = minibatch(train_data, size=batch_sizes)
             for batch in batches:
-                texts, annotations = zip(*batch)
+                nlp.update(batch, sgd=optimizer, drop=0.2, losses=losses)
-                nlp.update(texts, annotations, sgd=optimizer, drop=0.2, losses=losses)
             with textcat.model.use_params(optimizer.averages):
                 # evaluate on the dev data split off in load_data()
                 scores = evaluate(nlp.tokenizer, textcat, dev_texts, dev_cats)
@@ -117,7 +117,7 @@ def main(model=None, output_dir=None, n_iter=20, n_texts=2000, init_tok2vec=None
 def load_data(limit=0, split=0.8):
     """Load data from the IMDB dataset."""
     # Partition off part of the train data for evaluation
-    train_data, _ = thinc.extra.datasets.imdb()
+    train_data, _ = ml_datasets.imdb()
     random.shuffle(train_data)
     train_data = train_data[-limit:]
     texts, labels = zip(*train_data)

fabfile.py

@@ -1,9 +1,6 @@
-# coding: utf-8
-from __future__ import unicode_literals, print_function
-
 import contextlib
 from pathlib import Path
-from fabric.api import local, lcd, env, settings, prefix
+from fabric.api import local, lcd
 from os import path, environ
 import shutil
 import sys
@@ -82,9 +79,7 @@ def pex():
     with virtualenv(VENV_DIR) as venv_local:
         with lcd(path.dirname(__file__)):
             sha = local("git rev-parse --short HEAD", capture=True)
-            venv_local(
+            venv_local(f"pex dist/*.whl -e spacy -o dist/spacy-{sha}.pex", direct=True)
-                "pex dist/*.whl -e spacy -o dist/spacy-%s.pex" % sha, direct=True
-            )
 
 
 def clean():

requirements.txt

@@ -1,20 +1,21 @@
 # Our libraries
 cymem>=2.0.2,<2.1.0
 preshed>=3.0.2,<3.1.0
-thinc==7.4.0.dev0
+thinc==8.0.0a0
 blis>=0.4.0,<0.5.0
+ml_datasets>=0.1.1
 murmurhash>=0.28.0,<1.1.0
 wasabi>=0.4.0,<1.1.0
-srsly>=1.0.1,<1.1.0
+srsly>=2.0.0,<3.0.0
 catalogue>=0.0.7,<1.1.0
 # Third party dependencies
 numpy>=1.15.0
 requests>=2.13.0,<3.0.0
 plac>=0.9.6,<1.2.0
-pathlib==1.0.1; python_version < "3.4"
 tqdm>=4.38.0,<5.0.0
 # Optional dependencies
 jsonschema>=2.6.0,<3.1.0
+pydantic>=1.0.0,<2.0.0
 # Development dependencies
 cython>=0.25
 pytest>=4.6.5

setup.cfg

@@ -16,10 +16,7 @@ classifiers =
     Operating System :: MacOS :: MacOS X
     Operating System :: Microsoft :: Windows
     Programming Language :: Cython
-    Programming Language :: Python :: 2
-    Programming Language :: Python :: 2.7
     Programming Language :: Python :: 3
-    Programming Language :: Python :: 3.5
     Programming Language :: Python :: 3.6
     Programming Language :: Python :: 3.7
     Programming Language :: Python :: 3.8
@@ -30,32 +27,35 @@ zip_safe = false
 include_package_data = true
 scripts =
     bin/spacy
-python_requires = >=2.7,!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*
+python_requires = >=3.6
 setup_requires =
     wheel
     cython>=0.25
+    numpy>=1.15.0
     # We also need our Cython packages here to compile against
     cymem>=2.0.2,<2.1.0
     preshed>=3.0.2,<3.1.0
     murmurhash>=0.28.0,<1.1.0
-    thinc==7.4.0.dev0
+    thinc==8.0.0a0
 install_requires =
     # Our libraries
     murmurhash>=0.28.0,<1.1.0
     cymem>=2.0.2,<2.1.0
     preshed>=3.0.2,<3.1.0
-    thinc==7.4.0.dev0
+    thinc==8.0.0a0
     blis>=0.4.0,<0.5.0
     wasabi>=0.4.0,<1.1.0
-    srsly>=1.0.1,<1.1.0
+    srsly>=2.0.0,<3.0.0
     catalogue>=0.0.7,<1.1.0
+    ml_datasets
     # Third-party dependencies
     tqdm>=4.38.0,<5.0.0
     setuptools
     numpy>=1.15.0
     plac>=0.9.6,<1.2.0
     requests>=2.13.0,<3.0.0
-    pathlib==1.0.1; python_version < "3.4"
+    pydantic>=1.3.0,<2.0.0
+    tqdm>=4.38.0,<5.0.0
 
 [options.extras_require]
 lookups =

setup.py

@@ -1,37 +1,23 @@
 #!/usr/bin/env python
-from __future__ import print_function
-import io
-import os
-import subprocess
 import sys
-import contextlib
 from distutils.command.build_ext import build_ext
 from distutils.sysconfig import get_python_inc
 import distutils.util
 from distutils import ccompiler, msvccompiler
 from setuptools import Extension, setup, find_packages
+import numpy
+from pathlib import Path
+from Cython.Build import cythonize
+from Cython.Compiler import Options
 
 
-def is_new_osx():
+# Preserve `__doc__` on functions and classes
-    """Check whether we're on OSX >= 10.10"""
+# http://docs.cython.org/en/latest/src/userguide/source_files_and_compilation.html#compiler-options
-    name = distutils.util.get_platform()
+Options.docstrings = True
-    if sys.platform != "darwin":
-        return False
-    elif name.startswith("macosx-10"):
-        minor_version = int(name.split("-")[1].split(".")[1])
-        if minor_version >= 7:
-            return True
-        else:
-            return False
-    else:
-        return False
 
 
 PACKAGES = find_packages()
-
-
 MOD_NAMES = [
-    "spacy._align",
     "spacy.parts_of_speech",
     "spacy.strings",
     "spacy.lexeme",
@@ -63,16 +49,32 @@ MOD_NAMES = [
     "spacy.symbols",
     "spacy.vectors",
 ]
-
-
 COMPILE_OPTIONS = {
     "msvc": ["/Ox", "/EHsc"],
     "mingw32": ["-O2", "-Wno-strict-prototypes", "-Wno-unused-function"],
     "other": ["-O2", "-Wno-strict-prototypes", "-Wno-unused-function"],
 }
-
-
 LINK_OPTIONS = {"msvc": [], "mingw32": [], "other": []}
+COMPILER_DIRECTIVES = {
+    "language_level": -3,
+    "embedsignature": True,
+    "annotation_typing": False,
+}
+
+
+def is_new_osx():
+    """Check whether we're on OSX >= 10.10"""
+    name = distutils.util.get_platform()
+    if sys.platform != "darwin":
+        return False
+    elif name.startswith("macosx-10"):
+        minor_version = int(name.split("-")[1].split(".")[1])
+        if minor_version >= 7:
+            return True
+        else:
+            return False
+    else:
+        return False
 
 
 if is_new_osx():
@@ -105,87 +107,40 @@ class build_ext_subclass(build_ext, build_ext_options):
         build_ext.build_extensions(self)
 
 
-def generate_cython(root, source):
-    print("Cythonizing sources")
-    p = subprocess.call(
-        [sys.executable, os.path.join(root, "bin", "cythonize.py"), source],
-        env=os.environ,
-    )
-    if p != 0:
-        raise RuntimeError("Running cythonize failed")
-
-
-def is_source_release(path):
-    return os.path.exists(os.path.join(path, "PKG-INFO"))
-
-
 def clean(path):
-    for name in MOD_NAMES:
+    for path in path.glob("**/*"):
-        name = name.replace(".", "/")
+        if path.is_file() and path.suffix in (".so", ".cpp"):
-        for ext in [".so", ".html", ".cpp", ".c"]:
+            print(f"Deleting {path.name}")
-            file_path = os.path.join(path, name + ext)
+            path.unlink()
-            if os.path.exists(file_path):
-                os.unlink(file_path)
-
-
-@contextlib.contextmanager
-def chdir(new_dir):
-    old_dir = os.getcwd()
-    try:
-        os.chdir(new_dir)
-        sys.path.insert(0, new_dir)
-        yield
-    finally:
-        del sys.path[0]
-        os.chdir(old_dir)
 
 
 def setup_package():
-    root = os.path.abspath(os.path.dirname(__file__))
+    root = Path(__file__).parent
 
     if len(sys.argv) > 1 and sys.argv[1] == "clean":
-        return clean(root)
+        return clean(root / "spacy")
 
-    with chdir(root):
+    with (root / "spacy" / "about.py").open("r") as f:
-        with io.open(os.path.join(root, "spacy", "about.py"), encoding="utf8") as f:
         about = {}
         exec(f.read(), about)
 
     include_dirs = [
         get_python_inc(plat_specific=True),
-            os.path.join(root, "include"),
+        numpy.get_include(),
+        str(root / "include"),
     ]
-
     if (
         ccompiler.new_compiler().compiler_type == "msvc"
         and msvccompiler.get_build_version() == 9
     ):
-            include_dirs.append(os.path.join(root, "include", "msvc9"))
+        include_dirs.append(str(root / "include" / "msvc9"))
-
     ext_modules = []
-        for mod_name in MOD_NAMES:
+    for name in MOD_NAMES:
-            mod_path = mod_name.replace(".", "/") + ".cpp"
+        mod_path = name.replace(".", "/") + ".pyx"
-            extra_link_args = []
+        ext = Extension(name, [mod_path], language="c++")
-            # ???
+        ext_modules.append(ext)
-            # Imported from patch from @mikepb
+    print("Cythonizing sources")
-            # See Issue #267. Running blind here...
+    ext_modules = cythonize(ext_modules, compiler_directives=COMPILER_DIRECTIVES)
-            if sys.platform == "darwin":
-                dylib_path = [".." for _ in range(mod_name.count("."))]
-                dylib_path = "/".join(dylib_path)
-                dylib_path = "@loader_path/%s/spacy/platform/darwin/lib" % dylib_path
-                extra_link_args.append("-Wl,-rpath,%s" % dylib_path)
-            ext_modules.append(
-                Extension(
-                    mod_name,
-                    [mod_path],
-                    language="c++",
-                    include_dirs=include_dirs,
-                    extra_link_args=extra_link_args,
-                )
-            )
-
-        if not is_source_release(root):
-            generate_cython(root, "spacy")
 
     setup(
         name="spacy",
@@ -193,6 +148,8 @@ def setup_package():
         version=about["__version__"],
         ext_modules=ext_modules,
         cmdclass={"build_ext": build_ext_subclass},
+        include_dirs=include_dirs,
+        package_data={"": ["*.pyx", "*.pxd", "*.pxi", "*.cpp"]},
     )
 
 

spacy/__init__.py

@@ -1,5 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
 import warnings
 import sys
 
@@ -7,7 +5,7 @@ warnings.filterwarnings("ignore", message="numpy.dtype size changed")
 warnings.filterwarnings("ignore", message="numpy.ufunc size changed")
 
 # These are imported as part of the API
-from thinc.neural.util import prefer_gpu, require_gpu
+from thinc.api import prefer_gpu, require_gpu
 
 from . import pipeline
 from .cli.info import info as cli_info
@@ -23,6 +21,9 @@ if sys.maxunicode == 65535:
     raise SystemError(Errors.E130)
 
 
+config = registry
+
+
 def load(name, **overrides):
     depr_path = overrides.get("path")
     if depr_path not in (True, False, None):

spacy/__main__.py

@@ -1,21 +1,17 @@
-# coding: utf8
-from __future__ import print_function
-
-# NB! This breaks in plac on Python 2!!
-# from __future__ import unicode_literals
-
 if __name__ == "__main__":
     import plac
     import sys
     from wasabi import msg
     from spacy.cli import download, link, info, package, train, pretrain, convert
     from spacy.cli import init_model, profile, evaluate, validate, debug_data
+    from spacy.cli import train_from_config_cli
 
     commands = {
         "download": download,
         "link": link,
         "info": info,
         "train": train,
+        "train-from-config": train_from_config_cli,
         "pretrain": pretrain,
         "debug-data": debug_data,
         "evaluate": evaluate,
@@ -28,9 +24,9 @@ if __name__ == "__main__":
     if len(sys.argv) == 1:
         msg.info("Available commands", ", ".join(commands), exits=1)
     command = sys.argv.pop(1)
-    sys.argv[0] = "spacy %s" % command
+    sys.argv[0] = f"spacy {command}"
     if command in commands:
         plac.call(commands[command], sys.argv[1:])
     else:
-        available = "Available: {}".format(", ".join(commands))
+        available = f"Available: {', '.join(commands)}"
-        msg.fail("Unknown command: {}".format(command), available, exits=1)
+        msg.fail(f"Unknown command: {command}", available, exits=1)

spacy/_align.pyx

@@ -1,255 +0,0 @@
-# cython: infer_types=True
-'''Do Levenshtein alignment, for evaluation of tokenized input.
-
-Random notes:
-
-  r i n g
-  0 1 2 3 4
-r 1 0 1 2 3
-a 2 1 1 2 3
-n 3 2 2 1 2
-g 4 3 3 2 1
-
-0,0: (1,1)=min(0+0,1+1,1+1)=0 S
-1,0: (2,1)=min(1+1,0+1,2+1)=1 D
-2,0: (3,1)=min(2+1,3+1,1+1)=2 D
-3,0: (4,1)=min(3+1,4+1,2+1)=3 D
-0,1: (1,2)=min(1+1,2+1,0+1)=1 D
-1,1: (2,2)=min(0+1,1+1,1+1)=1 S
-2,1: (3,2)=min(1+1,1+1,2+1)=2 S or I
-3,1: (4,2)=min(2+1,2+1,3+1)=3 S or I
-0,2: (1,3)=min(2+1,3+1,1+1)=2 I
-1,2: (2,3)=min(1+1,2+1,1+1)=2 S or I
-2,2: (3,3)
-3,2: (4,3)
-At state (i, j) we're asking "How do I transform S[:i+1] to T[:j+1]?"
-
-We know the costs to transition:
-
-S[:i]   -> T[:j]   (at D[i,j])
-S[:i+1] -> T[:j]   (at D[i+1,j])
-S[:i]   -> T[:j+1] (at D[i,j+1])
-    
-Further, now we can transform:
-S[:i+1] -> S[:i] (DEL) for 1,
-T[:j+1] -> T[:j] (INS) for 1.
-S[i+1]  -> T[j+1] (SUB) for 0 or 1
-
-Therefore we have the costs:
-SUB: Cost(S[:i]->T[:j])   + Cost(S[i]->S[j])
-i.e. D[i, j] + S[i+1] != T[j+1]
-INS: Cost(S[:i+1]->T[:j]) + Cost(T[:j+1]->T[:j])
-i.e. D[i+1,j] + 1
-DEL: Cost(S[:i]->T[:j+1]) + Cost(S[:i+1]->S[:i]) 
-i.e. D[i,j+1] + 1
-
-    Source string S has length m, with index i
-    Target string T has length n, with index j
-
-    Output two alignment vectors: i2j (length m) and j2i (length n)
-    # function LevenshteinDistance(char s[1..m], char t[1..n]):
-    # for all i and j, d[i,j] will hold the Levenshtein distance between
-    # the first i characters of s and the first j characters of t
-    # note that d has (m+1)*(n+1) values
-    # set each element in d to zero
-    ring rang
-      - r i n g
-    - 0 0 0 0 0
-    r 0 0 0 0 0
-    a 0 0 0 0 0
-    n 0 0 0 0 0
-    g 0 0 0 0 0
-
-    # source prefixes can be transformed into empty string by
-    # dropping all characters
-    # d[i, 0] := i
-    ring rang
-      - r i n g
-    - 0 0 0 0 0
-    r 1 0 0 0 0
-    a 2 0 0 0 0
-    n 3 0 0 0 0
-    g 4 0 0 0 0
-
-    # target prefixes can be reached from empty source prefix
-    # by inserting every character
-    # d[0, j] := j
-      - r i n g
-    - 0 1 2 3 4
-    r 1 0 0 0 0
-    a 2 0 0 0 0
-    n 3 0 0 0 0
-    g 4 0 0 0 0
-
-'''
-from __future__ import unicode_literals
-from libc.stdint cimport uint32_t
-import numpy
-cimport numpy as np
-from .compat import unicode_
-from murmurhash.mrmr cimport hash32
-
-
-def align(S, T):
-    cdef int m = len(S)
-    cdef int n = len(T)
-    cdef np.ndarray matrix = numpy.zeros((m+1, n+1), dtype='int32')
-    cdef np.ndarray i2j = numpy.zeros((m,), dtype='i')
-    cdef np.ndarray j2i = numpy.zeros((n,), dtype='i')
-
-    cdef np.ndarray S_arr = _convert_sequence(S)
-    cdef np.ndarray T_arr = _convert_sequence(T)
-
-    fill_matrix(<int*>matrix.data,
-        <const int*>S_arr.data, m, <const int*>T_arr.data, n)
-    fill_i2j(i2j, matrix)
-    fill_j2i(j2i, matrix)
-    for i in range(i2j.shape[0]):
-        if i2j[i] >= 0 and len(S[i]) != len(T[i2j[i]]):
-            i2j[i] = -1
-    for j in range(j2i.shape[0]):
-        if j2i[j] >= 0 and len(T[j]) != len(S[j2i[j]]):
-            j2i[j] = -1
-    return matrix[-1,-1], i2j, j2i, matrix
-
-
-def multi_align(np.ndarray i2j, np.ndarray j2i, i_lengths, j_lengths):
-    '''Let's say we had:
-
-    Guess: [aa bb cc dd]
-    Truth: [aa bbcc dd]
-    i2j: [0, None, -2, 2]
-    j2i: [0, -2, 3]
-
-    We want:
-
-    i2j_multi: {1: 1, 2: 1}
-    j2i_multi: {}
-    '''
-    i2j_miss = _get_regions(i2j, i_lengths)
-    j2i_miss = _get_regions(j2i, j_lengths)
-
-    i2j_multi, j2i_multi = _get_mapping(i2j_miss, j2i_miss, i_lengths, j_lengths)
-    return i2j_multi, j2i_multi
-
-
-def _get_regions(alignment, lengths):
-    regions = {}
-    start = None
-    offset = 0
-    for i in range(len(alignment)):
-        if alignment[i] < 0:
-            if start is None:
-                start = offset
-                regions.setdefault(start, [])
-            regions[start].append(i)
-        else:
-            start = None
-        offset += lengths[i]
-    return regions
-
-
-def _get_mapping(miss1, miss2, lengths1, lengths2):
-    i2j = {}
-    j2i = {}
-    for start, region1 in miss1.items():
-        if not region1 or start not in miss2:
-            continue
-        region2 = miss2[start]
-        if sum(lengths1[i] for i in region1) == sum(lengths2[i] for i in region2):
-            j = region2.pop(0)
-            buff = []
-            # Consume tokens from region 1, until we meet the length of the
-            # first token in region2. If we do, align the tokens. If
-            # we exceed the length, break.
-            while region1:
-                buff.append(region1.pop(0))
-                if sum(lengths1[i] for i in buff) == lengths2[j]:
-                    for i in buff:
-                        i2j[i] = j
-                    j2i[j] = buff[-1]
-                    j += 1
-                    buff = []
-                elif sum(lengths1[i] for i in buff) > lengths2[j]:
-                    break
-            else:
-                if buff and sum(lengths1[i] for i in buff) == lengths2[j]:
-                    for i in buff:
-                        i2j[i] = j
-                    j2i[j] = buff[-1]
-    return i2j, j2i
-
-
-def _convert_sequence(seq):
-    if isinstance(seq, numpy.ndarray):
-        return numpy.ascontiguousarray(seq, dtype='uint32_t')
-    cdef np.ndarray output = numpy.zeros((len(seq),), dtype='uint32')
-    cdef bytes item_bytes
-    for i, item in enumerate(seq):
-        if item == "``":
-            item = '"'
-        elif item == "''":
-            item = '"'
-        if isinstance(item, unicode):
-            item_bytes = item.encode('utf8')
-        else:
-            item_bytes = item
-        output[i] = hash32(<void*><char*>item_bytes, len(item_bytes), 0)
-    return output
-
-
-cdef void fill_matrix(int* D, 
-        const int* S, int m, const int* T, int n) nogil:
-    m1 = m+1
-    n1 = n+1
-    for i in range(m1*n1):
-        D[i] = 0
- 
-    for i in range(m1):
-        D[i*n1] = i
- 
-    for j in range(n1):
-        D[j] = j
- 
-    cdef int sub_cost, ins_cost, del_cost
-    for j in range(n):
-        for i in range(m):
-            i_j = i*n1 + j
-            i1_j1 = (i+1)*n1 + j+1
-            i1_j = (i+1)*n1 + j
-            i_j1 = i*n1 + j+1
-            if S[i] != T[j]:
-                sub_cost = D[i_j] + 1
-            else:
-                sub_cost = D[i_j]
-            del_cost = D[i_j1] + 1
-            ins_cost = D[i1_j] + 1
-            best = min(min(sub_cost, ins_cost), del_cost)
-            D[i1_j1] = best
-
-
-cdef void fill_i2j(np.ndarray i2j, np.ndarray D) except *:
-    j = D.shape[1]-2
-    cdef int i = D.shape[0]-2
-    while i >= 0:
-        while D[i+1, j] < D[i+1, j+1]:
-            j -= 1
-        if D[i, j+1] < D[i+1, j+1]:
-            i2j[i] = -1
-        else:
-            i2j[i] = j
-            j -= 1
-        i -= 1
-
-cdef void fill_j2i(np.ndarray j2i, np.ndarray D) except *:
-    i = D.shape[0]-2
-    cdef int j = D.shape[1]-2
-    while j >= 0:
-        while D[i, j+1] < D[i+1, j+1]:
-            i -= 1
-        if D[i+1, j] < D[i+1, j+1]:
-            j2i[j] = -1
-        else:
-            j2i[j] = i
-            i -= 1
-        j -= 1

spacy/_ml.py

@@ -1,985 +0,0 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-import numpy
-from thinc.v2v import Model, Maxout, Softmax, Affine, ReLu
-from thinc.t2t import ExtractWindow, ParametricAttention
-from thinc.t2v import Pooling, sum_pool, mean_pool
-from thinc.i2v import HashEmbed
-from thinc.misc import Residual, FeatureExtracter
-from thinc.misc import LayerNorm as LN
-from thinc.api import add, layerize, chain, clone, concatenate, with_flatten
-from thinc.api import with_getitem, flatten_add_lengths
-from thinc.api import uniqued, wrap, noop
-from thinc.linear.linear import LinearModel
-from thinc.neural.ops import NumpyOps, CupyOps
-from thinc.neural.util import get_array_module, copy_array
-from thinc.neural.optimizers import Adam
-
-from thinc import describe
-from thinc.describe import Dimension, Synapses, Biases, Gradient
-from thinc.neural._classes.affine import _set_dimensions_if_needed
-import thinc.extra.load_nlp
-
-from .attrs import ID, ORTH, LOWER, NORM, PREFIX, SUFFIX, SHAPE
-from .errors import Errors, user_warning, Warnings
-from . import util
-from . import ml as new_ml
-from .ml import _legacy_tok2vec
-
-
-VECTORS_KEY = "spacy_pretrained_vectors"
-# Backwards compatibility with <2.2.2
-USE_MODEL_REGISTRY_TOK2VEC = False
-
-
-def cosine(vec1, vec2):
-    xp = get_array_module(vec1)
-    norm1 = xp.linalg.norm(vec1)
-    norm2 = xp.linalg.norm(vec2)
-    if norm1 == 0.0 or norm2 == 0.0:
-        return 0
-    else:
-        return vec1.dot(vec2) / (norm1 * norm2)
-
-
-def create_default_optimizer(ops, **cfg):
-    learn_rate = util.env_opt("learn_rate", 0.001)
-    beta1 = util.env_opt("optimizer_B1", 0.9)
-    beta2 = util.env_opt("optimizer_B2", 0.999)
-    eps = util.env_opt("optimizer_eps", 1e-8)
-    L2 = util.env_opt("L2_penalty", 1e-6)
-    max_grad_norm = util.env_opt("grad_norm_clip", 1.0)
-    optimizer = Adam(ops, learn_rate, L2=L2, beta1=beta1, beta2=beta2, eps=eps)
-    optimizer.max_grad_norm = max_grad_norm
-    optimizer.device = ops.device
-    return optimizer
-
-
-@layerize
-def _flatten_add_lengths(seqs, pad=0, drop=0.0):
-    ops = Model.ops
-    lengths = ops.asarray([len(seq) for seq in seqs], dtype="i")
-
-    def finish_update(d_X, sgd=None):
-        return ops.unflatten(d_X, lengths, pad=pad)
-
-    X = ops.flatten(seqs, pad=pad)
-    return (X, lengths), finish_update
-
-
-def _zero_init(model):
-    def _zero_init_impl(self, *args, **kwargs):
-        self.W.fill(0)
-
-    model.on_init_hooks.append(_zero_init_impl)
-    if model.W is not None:
-        model.W.fill(0.0)
-    return model
-
-
-def with_cpu(ops, model):
-    """Wrap a model that should run on CPU, transferring inputs and outputs
-    as necessary."""
-    model.to_cpu()
-
-    def with_cpu_forward(inputs, drop=0.0):
-        cpu_outputs, backprop = model.begin_update(_to_cpu(inputs), drop=drop)
-        gpu_outputs = _to_device(ops, cpu_outputs)
-
-        def with_cpu_backprop(d_outputs, sgd=None):
-            cpu_d_outputs = _to_cpu(d_outputs)
-            return backprop(cpu_d_outputs, sgd=sgd)
-
-        return gpu_outputs, with_cpu_backprop
-
-    return wrap(with_cpu_forward, model)
-
-
-def _to_cpu(X):
-    if isinstance(X, numpy.ndarray):
-        return X
-    elif isinstance(X, tuple):
-        return tuple([_to_cpu(x) for x in X])
-    elif isinstance(X, list):
-        return [_to_cpu(x) for x in X]
-    elif hasattr(X, "get"):
-        return X.get()
-    else:
-        return X
-
-
-def _to_device(ops, X):
-    if isinstance(X, tuple):
-        return tuple([_to_device(ops, x) for x in X])
-    elif isinstance(X, list):
-        return [_to_device(ops, x) for x in X]
-    else:
-        return ops.asarray(X)
-
-
-class extract_ngrams(Model):
-    def __init__(self, ngram_size, attr=LOWER):
-        Model.__init__(self)
-        self.ngram_size = ngram_size
-        self.attr = attr
-
-    def begin_update(self, docs, drop=0.0):
-        batch_keys = []
-        batch_vals = []
-        for doc in docs:
-            unigrams = doc.to_array([self.attr])
-            ngrams = [unigrams]
-            for n in range(2, self.ngram_size + 1):
-                ngrams.append(self.ops.ngrams(n, unigrams))
-            keys = self.ops.xp.concatenate(ngrams)
-            keys, vals = self.ops.xp.unique(keys, return_counts=True)
-            batch_keys.append(keys)
-            batch_vals.append(vals)
-        # The dtype here matches what thinc is expecting -- which differs per
-        # platform (by int definition). This should be fixed once the problem
-        # is fixed on Thinc's side.
-        lengths = self.ops.asarray(
-            [arr.shape[0] for arr in batch_keys], dtype=numpy.int_
-        )
-        batch_keys = self.ops.xp.concatenate(batch_keys)
-        batch_vals = self.ops.asarray(self.ops.xp.concatenate(batch_vals), dtype="f")
-        return (batch_keys, batch_vals, lengths), None
-
-
-@describe.on_data(
-    _set_dimensions_if_needed, lambda model, X, y: model.init_weights(model)
-)
-@describe.attributes(
-    nI=Dimension("Input size"),
-    nF=Dimension("Number of features"),
-    nO=Dimension("Output size"),
-    nP=Dimension("Maxout pieces"),
-    W=Synapses("Weights matrix", lambda obj: (obj.nF, obj.nO, obj.nP, obj.nI)),
-    b=Biases("Bias vector", lambda obj: (obj.nO, obj.nP)),
-    pad=Synapses(
-        "Pad",
-        lambda obj: (1, obj.nF, obj.nO, obj.nP),
-        lambda M, ops: ops.normal_init(M, 1.0),
-    ),
-    d_W=Gradient("W"),
-    d_pad=Gradient("pad"),
-    d_b=Gradient("b"),
-)
-class PrecomputableAffine(Model):
-    def __init__(self, nO=None, nI=None, nF=None, nP=None, **kwargs):
-        Model.__init__(self, **kwargs)
-        self.nO = nO
-        self.nP = nP
-        self.nI = nI
-        self.nF = nF
-
-    def begin_update(self, X, drop=0.0):
-        Yf = self.ops.gemm(
-            X, self.W.reshape((self.nF * self.nO * self.nP, self.nI)), trans2=True
-        )
-        Yf = Yf.reshape((Yf.shape[0], self.nF, self.nO, self.nP))
-        Yf = self._add_padding(Yf)
-
-        def backward(dY_ids, sgd=None):
-            dY, ids = dY_ids
-            dY, ids = self._backprop_padding(dY, ids)
-            Xf = X[ids]
-            Xf = Xf.reshape((Xf.shape[0], self.nF * self.nI))
-
-            self.d_b += dY.sum(axis=0)
-            dY = dY.reshape((dY.shape[0], self.nO * self.nP))
-
-            Wopfi = self.W.transpose((1, 2, 0, 3))
-            Wopfi = self.ops.xp.ascontiguousarray(Wopfi)
-            Wopfi = Wopfi.reshape((self.nO * self.nP, self.nF * self.nI))
-            dXf = self.ops.gemm(dY.reshape((dY.shape[0], self.nO * self.nP)), Wopfi)
-
-            # Reuse the buffer
-            dWopfi = Wopfi
-            dWopfi.fill(0.0)
-            self.ops.gemm(dY, Xf, out=dWopfi, trans1=True)
-            dWopfi = dWopfi.reshape((self.nO, self.nP, self.nF, self.nI))
-            # (o, p, f, i) --> (f, o, p, i)
-            self.d_W += dWopfi.transpose((2, 0, 1, 3))
-
-            if sgd is not None:
-                sgd(self._mem.weights, self._mem.gradient, key=self.id)
-            return dXf.reshape((dXf.shape[0], self.nF, self.nI))
-
-        return Yf, backward
-
-    def _add_padding(self, Yf):
-        Yf_padded = self.ops.xp.vstack((self.pad, Yf))
-        return Yf_padded
-
-    def _backprop_padding(self, dY, ids):
-        # (1, nF, nO, nP) += (nN, nF, nO, nP) where IDs (nN, nF) < 0
-        mask = ids < 0.0
-        mask = mask.sum(axis=1)
-        d_pad = dY * mask.reshape((ids.shape[0], 1, 1))
-        self.d_pad += d_pad.sum(axis=0)
-        return dY, ids
-
-    @staticmethod
-    def init_weights(model):
-        """This is like the 'layer sequential unit variance', but instead
-        of taking the actual inputs, we randomly generate whitened data.
-
-        Why's this all so complicated? We have a huge number of inputs,
-        and the maxout unit makes guessing the dynamics tricky. Instead
-        we set the maxout weights to values that empirically result in
-        whitened outputs given whitened inputs.
-        """
-        if (model.W ** 2).sum() != 0.0:
-            return
-        ops = model.ops
-        xp = ops.xp
-        ops.normal_init(model.W, model.nF * model.nI, inplace=True)
-
-        ids = ops.allocate((5000, model.nF), dtype="f")
-        ids += xp.random.uniform(0, 1000, ids.shape)
-        ids = ops.asarray(ids, dtype="i")
-        tokvecs = ops.allocate((5000, model.nI), dtype="f")
-        tokvecs += xp.random.normal(loc=0.0, scale=1.0, size=tokvecs.size).reshape(
-            tokvecs.shape
-        )
-
-        def predict(ids, tokvecs):
-            # nS ids. nW tokvecs. Exclude the padding array.
-            hiddens = model(tokvecs[:-1])  # (nW, f, o, p)
-            vectors = model.ops.allocate((ids.shape[0], model.nO * model.nP), dtype="f")
-            # need nS vectors
-            hiddens = hiddens.reshape(
-                (hiddens.shape[0] * model.nF, model.nO * model.nP)
-            )
-            model.ops.scatter_add(vectors, ids.flatten(), hiddens)
-            vectors = vectors.reshape((vectors.shape[0], model.nO, model.nP))
-            vectors += model.b
-            vectors = model.ops.asarray(vectors)
-            if model.nP >= 2:
-                return model.ops.maxout(vectors)[0]
-            else:
-                return vectors * (vectors >= 0)
-
-        tol_var = 0.01
-        tol_mean = 0.01
-        t_max = 10
-        t_i = 0
-        for t_i in range(t_max):
-            acts1 = predict(ids, tokvecs)
-            var = model.ops.xp.var(acts1)
-            mean = model.ops.xp.mean(acts1)
-            if abs(var - 1.0) >= tol_var:
-                model.W /= model.ops.xp.sqrt(var)
-            elif abs(mean) >= tol_mean:
-                model.b -= mean
-            else:
-                break
-
-
-def link_vectors_to_models(vocab):
-    vectors = vocab.vectors
-    if vectors.name is None:
-        vectors.name = VECTORS_KEY
-        if vectors.data.size != 0:
-            user_warning(Warnings.W020.format(shape=vectors.data.shape))
-    ops = Model.ops
-    for word in vocab:
-        if word.orth in vectors.key2row:
-            word.rank = vectors.key2row[word.orth]
-        else:
-            word.rank = 0
-    data = ops.asarray(vectors.data)
-    # Set an entry here, so that vectors are accessed by StaticVectors
-    # (unideal, I know)
-    key = (ops.device, vectors.name)
-    if key in thinc.extra.load_nlp.VECTORS:
-        if thinc.extra.load_nlp.VECTORS[key].shape != data.shape:
-            # This is a hack to avoid the problem in #3853. Maybe we should
-            # print a warning as well?
-            old_name = vectors.name
-            new_name = vectors.name + "_%d" % data.shape[0]
-            user_warning(Warnings.W019.format(old=old_name, new=new_name))
-            vectors.name = new_name
-            key = (ops.device, vectors.name)
-    thinc.extra.load_nlp.VECTORS[key] = data
-
-
-def PyTorchBiLSTM(nO, nI, depth, dropout=0.2):
-    import torch.nn
-    from thinc.api import with_square_sequences
-    from thinc.extra.wrappers import PyTorchWrapperRNN
-
-    if depth == 0:
-        return layerize(noop())
-    model = torch.nn.LSTM(nI, nO // 2, depth, bidirectional=True, dropout=dropout)
-    return with_square_sequences(PyTorchWrapperRNN(model))
-
-
-def Tok2Vec(width, embed_size, **kwargs):
-    if not USE_MODEL_REGISTRY_TOK2VEC:
-        # Preserve prior tok2vec for backwards compat, in v2.2.2
-        return _legacy_tok2vec.Tok2Vec(width, embed_size, **kwargs)
-    pretrained_vectors = kwargs.get("pretrained_vectors", None)
-    cnn_maxout_pieces = kwargs.get("cnn_maxout_pieces", 3)
-    subword_features = kwargs.get("subword_features", True)
-    char_embed = kwargs.get("char_embed", False)
-    conv_depth = kwargs.get("conv_depth", 4)
-    bilstm_depth = kwargs.get("bilstm_depth", 0)
-    conv_window = kwargs.get("conv_window", 1)
-
-    cols = ["ID", "NORM", "PREFIX", "SUFFIX", "SHAPE", "ORTH"]
-
-    doc2feats_cfg = {"arch": "spacy.Doc2Feats.v1", "config": {"columns": cols}}
-    if char_embed:
-        embed_cfg = {
-            "arch": "spacy.CharacterEmbed.v1",
-            "config": {
-                "width": 64,
-                "chars": 6,
-                "@mix": {
-                    "arch": "spacy.LayerNormalizedMaxout.v1",
-                    "config": {"width": width, "pieces": 3},
-                },
-                "@embed_features": None,
-            },
-        }
-    else:
-        embed_cfg = {
-            "arch": "spacy.MultiHashEmbed.v1",
-            "config": {
-                "width": width,
-                "rows": embed_size,
-                "columns": cols,
-                "use_subwords": subword_features,
-                "@pretrained_vectors": None,
-                "@mix": {
-                    "arch": "spacy.LayerNormalizedMaxout.v1",
-                    "config": {"width": width, "pieces": 3},
-                },
-            },
-        }
-        if pretrained_vectors:
-            embed_cfg["config"]["@pretrained_vectors"] = {
-                "arch": "spacy.PretrainedVectors.v1",
-                "config": {
-                    "vectors_name": pretrained_vectors,
-                    "width": width,
-                    "column": cols.index("ID"),
-                },
-            }
-    if cnn_maxout_pieces >= 2:
-        cnn_cfg = {
-            "arch": "spacy.MaxoutWindowEncoder.v1",
-            "config": {
-                "width": width,
-                "window_size": conv_window,
-                "pieces": cnn_maxout_pieces,
-                "depth": conv_depth,
-            },
-        }
-    else:
-        cnn_cfg = {
-            "arch": "spacy.MishWindowEncoder.v1",
-            "config": {"width": width, "window_size": conv_window, "depth": conv_depth},
-        }
-    bilstm_cfg = {
-        "arch": "spacy.TorchBiLSTMEncoder.v1",
-        "config": {"width": width, "depth": bilstm_depth},
-    }
-    if conv_depth == 0 and bilstm_depth == 0:
-        encode_cfg = {}
-    elif conv_depth >= 1 and bilstm_depth >= 1:
-        encode_cfg = {
-            "arch": "thinc.FeedForward.v1",
-            "config": {"children": [cnn_cfg, bilstm_cfg]},
-        }
-    elif conv_depth >= 1:
-        encode_cfg = cnn_cfg
-    else:
-        encode_cfg = bilstm_cfg
-    config = {"@doc2feats": doc2feats_cfg, "@embed": embed_cfg, "@encode": encode_cfg}
-    return new_ml.Tok2Vec(config)
-
-
-def reapply(layer, n_times):
-    def reapply_fwd(X, drop=0.0):
-        backprops = []
-        for i in range(n_times):
-            Y, backprop = layer.begin_update(X, drop=drop)
-            X = Y
-            backprops.append(backprop)
-
-        def reapply_bwd(dY, sgd=None):
-            dX = None
-            for backprop in reversed(backprops):
-                dY = backprop(dY, sgd=sgd)
-                if dX is None:
-                    dX = dY
-                else:
-                    dX += dY
-            return dX
-
-        return Y, reapply_bwd
-
-    return wrap(reapply_fwd, layer)
-
-
-def asarray(ops, dtype):
-    def forward(X, drop=0.0):
-        return ops.asarray(X, dtype=dtype), None
-
-    return layerize(forward)
-
-
-def _divide_array(X, size):
-    parts = []
-    index = 0
-    while index < len(X):
-        parts.append(X[index : index + size])
-        index += size
-    return parts
-
-
-def get_col(idx):
-    if idx < 0:
-        raise IndexError(Errors.E066.format(value=idx))
-
-    def forward(X, drop=0.0):
-        if isinstance(X, numpy.ndarray):
-            ops = NumpyOps()
-        else:
-            ops = CupyOps()
-        output = ops.xp.ascontiguousarray(X[:, idx], dtype=X.dtype)
-
-        def backward(y, sgd=None):
-            dX = ops.allocate(X.shape)
-            dX[:, idx] += y
-            return dX
-
-        return output, backward
-
-    return layerize(forward)
-
-
-def doc2feats(cols=None):
-    if cols is None:
-        cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
-
-    def forward(docs, drop=0.0):
-        feats = []
-        for doc in docs:
-            feats.append(doc.to_array(cols))
-        return feats, None
-
-    model = layerize(forward)
-    model.cols = cols
-    return model
-
-
-def print_shape(prefix):
-    def forward(X, drop=0.0):
-        return X, lambda dX, **kwargs: dX
-
-    return layerize(forward)
-
-
-@layerize
-def get_token_vectors(tokens_attrs_vectors, drop=0.0):
-    tokens, attrs, vectors = tokens_attrs_vectors
-
-    def backward(d_output, sgd=None):
-        return (tokens, d_output)
-
-    return vectors, backward
-
-
-@layerize
-def logistic(X, drop=0.0):
-    xp = get_array_module(X)
-    if not isinstance(X, xp.ndarray):
-        X = xp.asarray(X)
-    # Clip to range (-10, 10)
-    X = xp.minimum(X, 10.0, X)
-    X = xp.maximum(X, -10.0, X)
-    Y = 1.0 / (1.0 + xp.exp(-X))
-
-    def logistic_bwd(dY, sgd=None):
-        dX = dY * (Y * (1 - Y))
-        return dX
-
-    return Y, logistic_bwd
-
-
-def zero_init(model):
-    def _zero_init_impl(self, X, y):
-        self.W.fill(0)
-
-    model.on_data_hooks.append(_zero_init_impl)
-    return model
-
-
-def getitem(i):
-    def getitem_fwd(X, drop=0.0):
-        return X[i], None
-
-    return layerize(getitem_fwd)
-
-
-@describe.attributes(
-    W=Synapses("Weights matrix", lambda obj: (obj.nO, obj.nI), lambda W, ops: None)
-)
-class MultiSoftmax(Affine):
-    """Neural network layer that predicts several multi-class attributes at once.
-    For instance, we might predict one class with 6 variables, and another with 5.
-    We predict the 11 neurons required for this, and then softmax them such
-    that columns 0-6 make a probability distribution and coumns 6-11 make another.
-    """
-
-    name = "multisoftmax"
-
-    def __init__(self, out_sizes, nI=None, **kwargs):
-        Model.__init__(self, **kwargs)
-        self.out_sizes = out_sizes
-        self.nO = sum(out_sizes)
-        self.nI = nI
-
-    def predict(self, input__BI):
-        output__BO = self.ops.affine(self.W, self.b, input__BI)
-        i = 0
-        for out_size in self.out_sizes:
-            self.ops.softmax(output__BO[:, i : i + out_size], inplace=True)
-            i += out_size
-        return output__BO
-
-    def begin_update(self, input__BI, drop=0.0):
-        output__BO = self.predict(input__BI)
-
-        def finish_update(grad__BO, sgd=None):
-            self.d_W += self.ops.gemm(grad__BO, input__BI, trans1=True)
-            self.d_b += grad__BO.sum(axis=0)
-            grad__BI = self.ops.gemm(grad__BO, self.W)
-            if sgd is not None:
-                sgd(self._mem.weights, self._mem.gradient, key=self.id)
-            return grad__BI
-
-        return output__BO, finish_update
-
-
-def build_tagger_model(nr_class, **cfg):
-    embed_size = util.env_opt("embed_size", 2000)
-    if "token_vector_width" in cfg:
-        token_vector_width = cfg["token_vector_width"]
-    else:
-        token_vector_width = util.env_opt("token_vector_width", 96)
-    pretrained_vectors = cfg.get("pretrained_vectors")
-    subword_features = cfg.get("subword_features", True)
-    with Model.define_operators({">>": chain, "+": add}):
-        if "tok2vec" in cfg:
-            tok2vec = cfg["tok2vec"]
-        else:
-            tok2vec = Tok2Vec(
-                token_vector_width,
-                embed_size,
-                subword_features=subword_features,
-                pretrained_vectors=pretrained_vectors,
-            )
-        softmax = with_flatten(Softmax(nr_class, token_vector_width))
-        model = tok2vec >> softmax
-    model.nI = None
-    model.tok2vec = tok2vec
-    model.softmax = softmax
-    return model
-
-
-def build_morphologizer_model(class_nums, **cfg):
-    embed_size = util.env_opt("embed_size", 7000)
-    if "token_vector_width" in cfg:
-        token_vector_width = cfg["token_vector_width"]
-    else:
-        token_vector_width = util.env_opt("token_vector_width", 128)
-    pretrained_vectors = cfg.get("pretrained_vectors")
-    char_embed = cfg.get("char_embed", True)
-    with Model.define_operators({">>": chain, "+": add, "**": clone}):
-        if "tok2vec" in cfg:
-            tok2vec = cfg["tok2vec"]
-        else:
-            tok2vec = Tok2Vec(
-                token_vector_width,
-                embed_size,
-                char_embed=char_embed,
-                pretrained_vectors=pretrained_vectors,
-            )
-        softmax = with_flatten(MultiSoftmax(class_nums, token_vector_width))
-        softmax.out_sizes = class_nums
-        model = tok2vec >> softmax
-    model.nI = None
-    model.tok2vec = tok2vec
-    model.softmax = softmax
-    return model
-
-
-@layerize
-def SpacyVectors(docs, drop=0.0):
-    batch = []
-    for doc in docs:
-        indices = numpy.zeros((len(doc),), dtype="i")
-        for i, word in enumerate(doc):
-            if word.orth in doc.vocab.vectors.key2row:
-                indices[i] = doc.vocab.vectors.key2row[word.orth]
-            else:
-                indices[i] = 0
-        vectors = doc.vocab.vectors.data[indices]
-        batch.append(vectors)
-    return batch, None
-
-
-def build_text_classifier(nr_class, width=64, **cfg):
-    depth = cfg.get("depth", 2)
-    nr_vector = cfg.get("nr_vector", 5000)
-    pretrained_dims = cfg.get("pretrained_dims", 0)
-    with Model.define_operators({">>": chain, "+": add, "|": concatenate, "**": clone}):
-        if cfg.get("low_data") and pretrained_dims:
-            model = (
-                SpacyVectors
-                >> flatten_add_lengths
-                >> with_getitem(0, Affine(width, pretrained_dims))
-                >> ParametricAttention(width)
-                >> Pooling(sum_pool)
-                >> Residual(ReLu(width, width)) ** 2
-                >> zero_init(Affine(nr_class, width, drop_factor=0.0))
-                >> logistic
-            )
-            return model
-
-        lower = HashEmbed(width, nr_vector, column=1)
-        prefix = HashEmbed(width // 2, nr_vector, column=2)
-        suffix = HashEmbed(width // 2, nr_vector, column=3)
-        shape = HashEmbed(width // 2, nr_vector, column=4)
-
-        trained_vectors = FeatureExtracter(
-            [ORTH, LOWER, PREFIX, SUFFIX, SHAPE, ID]
-        ) >> with_flatten(
-            uniqued(
-                (lower | prefix | suffix | shape)
-                >> LN(Maxout(width, width + (width // 2) * 3)),
-                column=0,
-            )
-        )
-
-        if pretrained_dims:
-            static_vectors = SpacyVectors >> with_flatten(
-                Affine(width, pretrained_dims)
-            )
-            # TODO Make concatenate support lists
-            vectors = concatenate_lists(trained_vectors, static_vectors)
-            vectors_width = width * 2
-        else:
-            vectors = trained_vectors
-            vectors_width = width
-            static_vectors = None
-        tok2vec = vectors >> with_flatten(
-            LN(Maxout(width, vectors_width))
-            >> Residual((ExtractWindow(nW=1) >> LN(Maxout(width, width * 3)))) ** depth,
-            pad=depth,
-        )
-        cnn_model = (
-            tok2vec
-            >> flatten_add_lengths
-            >> ParametricAttention(width)
-            >> Pooling(sum_pool)
-            >> Residual(zero_init(Maxout(width, width)))
-            >> zero_init(Affine(nr_class, width, drop_factor=0.0))
-        )
-
-        linear_model = build_bow_text_classifier(
-            nr_class, ngram_size=cfg.get("ngram_size", 1), exclusive_classes=False
-        )
-        if cfg.get("exclusive_classes"):
-            output_layer = Softmax(nr_class, nr_class * 2)
-        else:
-            output_layer = (
-                zero_init(Affine(nr_class, nr_class * 2, drop_factor=0.0)) >> logistic
-            )
-        model = (linear_model | cnn_model) >> output_layer
-        model.tok2vec = chain(tok2vec, flatten)
-    model.nO = nr_class
-    model.lsuv = False
-    return model
-
-
-def build_bow_text_classifier(
-    nr_class, ngram_size=1, exclusive_classes=False, no_output_layer=False, **cfg
-):
-    with Model.define_operators({">>": chain}):
-        model = with_cpu(
-            Model.ops, extract_ngrams(ngram_size, attr=ORTH) >> LinearModel(nr_class)
-        )
-        if not no_output_layer:
-            model = model >> (cpu_softmax if exclusive_classes else logistic)
-    model.nO = nr_class
-    return model
-
-
-@layerize
-def cpu_softmax(X, drop=0.0):
-    ops = NumpyOps()
-
-    def cpu_softmax_backward(dY, sgd=None):
-        return dY
-
-    return ops.softmax(X), cpu_softmax_backward
-
-
-def build_simple_cnn_text_classifier(tok2vec, nr_class, exclusive_classes=False, **cfg):
-    """
-    Build a simple CNN text classifier, given a token-to-vector model as inputs.
-    If exclusive_classes=True, a softmax non-linearity is applied, so that the
-    outputs sum to 1. If exclusive_classes=False, a logistic non-linearity
-    is applied instead, so that outputs are in the range [0, 1].
-    """
-    with Model.define_operators({">>": chain}):
-        if exclusive_classes:
-            output_layer = Softmax(nr_class, tok2vec.nO)
-        else:
-            output_layer = (
-                zero_init(Affine(nr_class, tok2vec.nO, drop_factor=0.0)) >> logistic
-            )
-        model = tok2vec >> flatten_add_lengths >> Pooling(mean_pool) >> output_layer
-    model.tok2vec = chain(tok2vec, flatten)
-    model.nO = nr_class
-    return model
-
-
-def build_nel_encoder(embed_width, hidden_width, ner_types, **cfg):
-    if "entity_width" not in cfg:
-        raise ValueError(Errors.E144.format(param="entity_width"))
-
-    conv_depth = cfg.get("conv_depth", 2)
-    cnn_maxout_pieces = cfg.get("cnn_maxout_pieces", 3)
-    pretrained_vectors = cfg.get("pretrained_vectors", None)
-    context_width = cfg.get("entity_width")
-
-    with Model.define_operators({">>": chain, "**": clone}):
-        # context encoder
-        tok2vec = Tok2Vec(
-            width=hidden_width,
-            embed_size=embed_width,
-            pretrained_vectors=pretrained_vectors,
-            cnn_maxout_pieces=cnn_maxout_pieces,
-            subword_features=True,
-            conv_depth=conv_depth,
-            bilstm_depth=0,
-        )
-
-        model = (
-            tok2vec
-            >> flatten_add_lengths
-            >> Pooling(mean_pool)
-            >> Residual(zero_init(Maxout(hidden_width, hidden_width)))
-            >> zero_init(Affine(context_width, hidden_width, drop_factor=0.0))
-        )
-
-        model.tok2vec = tok2vec
-        model.nO = context_width
-    return model
-
-
-@layerize
-def flatten(seqs, drop=0.0):
-    ops = Model.ops
-    lengths = ops.asarray([len(seq) for seq in seqs], dtype="i")
-
-    def finish_update(d_X, sgd=None):
-        return ops.unflatten(d_X, lengths, pad=0)
-
-    X = ops.flatten(seqs, pad=0)
-    return X, finish_update
-
-
-def concatenate_lists(*layers, **kwargs):  # pragma: no cover
-    """Compose two or more models `f`, `g`, etc, such that their outputs are
-    concatenated, i.e. `concatenate(f, g)(x)` computes `hstack(f(x), g(x))`
-    """
-    if not layers:
-        return noop()
-    drop_factor = kwargs.get("drop_factor", 1.0)
-    ops = layers[0].ops
-    layers = [chain(layer, flatten) for layer in layers]
-    concat = concatenate(*layers)
-
-    def concatenate_lists_fwd(Xs, drop=0.0):
-        if drop is not None:
-            drop *= drop_factor
-        lengths = ops.asarray([len(X) for X in Xs], dtype="i")
-        flat_y, bp_flat_y = concat.begin_update(Xs, drop=drop)
-        ys = ops.unflatten(flat_y, lengths)
-
-        def concatenate_lists_bwd(d_ys, sgd=None):
-            return bp_flat_y(ops.flatten(d_ys), sgd=sgd)
-
-        return ys, concatenate_lists_bwd
-
-    model = wrap(concatenate_lists_fwd, concat)
-    return model
-
-
-def masked_language_model(vocab, model, mask_prob=0.15):
-    """Convert a model into a BERT-style masked language model"""
-
-    random_words = _RandomWords(vocab)
-
-    def mlm_forward(docs, drop=0.0):
-        mask, docs = _apply_mask(docs, random_words, mask_prob=mask_prob)
-        mask = model.ops.asarray(mask).reshape((mask.shape[0], 1))
-        output, backprop = model.begin_update(docs, drop=drop)
-
-        def mlm_backward(d_output, sgd=None):
-            d_output *= 1 - mask
-            return backprop(d_output, sgd=sgd)
-
-        return output, mlm_backward
-
-    return wrap(mlm_forward, model)
-
-
-class _RandomWords(object):
-    def __init__(self, vocab):
-        self.words = [lex.text for lex in vocab if lex.prob != 0.0]
-        self.probs = [lex.prob for lex in vocab if lex.prob != 0.0]
-        self.words = self.words[:10000]
-        self.probs = self.probs[:10000]
-        self.probs = numpy.exp(numpy.array(self.probs, dtype="f"))
-        self.probs /= self.probs.sum()
-        self._cache = []
-
-    def next(self):
-        if not self._cache:
-            self._cache.extend(
-                numpy.random.choice(len(self.words), 10000, p=self.probs)
-            )
-        index = self._cache.pop()
-        return self.words[index]
-
-
-def _apply_mask(docs, random_words, mask_prob=0.15):
-    # This needs to be here to avoid circular imports
-    from .tokens.doc import Doc
-
-    N = sum(len(doc) for doc in docs)
-    mask = numpy.random.uniform(0.0, 1.0, (N,))
-    mask = mask >= mask_prob
-    i = 0
-    masked_docs = []
-    for doc in docs:
-        words = []
-        for token in doc:
-            if not mask[i]:
-                word = _replace_word(token.text, random_words)
-            else:
-                word = token.text
-            words.append(word)
-            i += 1
-        spaces = [bool(w.whitespace_) for w in doc]
-        # NB: If you change this implementation to instead modify
-        # the docs in place, take care that the IDs reflect the original
-        # words. Currently we use the original docs to make the vectors
-        # for the target, so we don't lose the original tokens. But if
-        # you modified the docs in place here, you would.
-        masked_docs.append(Doc(doc.vocab, words=words, spaces=spaces))
-    return mask, masked_docs
-
-
-def _replace_word(word, random_words, mask="[MASK]"):
-    roll = numpy.random.random()
-    if roll < 0.8:
-        return mask
-    elif roll < 0.9:
-        return random_words.next()
-    else:
-        return word
-
-
-def _uniform_init(lo, hi):
-    def wrapped(W, ops):
-        copy_array(W, ops.xp.random.uniform(lo, hi, W.shape))
-
-    return wrapped
-
-
-@describe.attributes(
-    nM=Dimension("Vector dimensions"),
-    nC=Dimension("Number of characters per word"),
-    vectors=Synapses(
-        "Embed matrix", lambda obj: (obj.nC, obj.nV, obj.nM), _uniform_init(-0.1, 0.1)
-    ),
-    d_vectors=Gradient("vectors"),
-)
-class CharacterEmbed(Model):
-    def __init__(self, nM=None, nC=None, **kwargs):
-        Model.__init__(self, **kwargs)
-        self.nM = nM
-        self.nC = nC
-
-    @property
-    def nO(self):
-        return self.nM * self.nC
-
-    @property
-    def nV(self):
-        return 256
-
-    def begin_update(self, docs, drop=0.0):
-        if not docs:
-            return []
-        ids = []
-        output = []
-        weights = self.vectors
-        # This assists in indexing; it's like looping over this dimension.
-        # Still consider this weird witch craft...But thanks to Mark Neumann
-        # for the tip.
-        nCv = self.ops.xp.arange(self.nC)
-        for doc in docs:
-            doc_ids = doc.to_utf8_array(nr_char=self.nC)
-            doc_vectors = self.ops.allocate((len(doc), self.nC, self.nM))
-            # Let's say I have a 2d array of indices, and a 3d table of data. What numpy
-            # incantation do I chant to get
-            # output[i, j, k] == data[j, ids[i, j], k]?
-            doc_vectors[:, nCv] = weights[nCv, doc_ids[:, nCv]]
-            output.append(doc_vectors.reshape((len(doc), self.nO)))
-            ids.append(doc_ids)
-
-        def backprop_character_embed(d_vectors, sgd=None):
-            gradient = self.d_vectors
-            for doc_ids, d_doc_vectors in zip(ids, d_vectors):
-                d_doc_vectors = d_doc_vectors.reshape((len(doc_ids), self.nC, self.nM))
-                gradient[nCv, doc_ids[:, nCv]] += d_doc_vectors[:, nCv]
-            if sgd is not None:
-                sgd(self._mem.weights, self._mem.gradient, key=self.id)
-            return None
-
-        return output, backprop_character_embed
-
-
-def get_cossim_loss(yh, y, ignore_zeros=False):
-    xp = get_array_module(yh)
-    # Find the zero vectors
-    if ignore_zeros:
-        zero_indices = xp.abs(y).sum(axis=1) == 0
-    # Add a small constant to avoid 0 vectors
-    yh = yh + 1e-8
-    y = y + 1e-8
-    # https://math.stackexchange.com/questions/1923613/partial-derivative-of-cosine-similarity
-    norm_yh = xp.linalg.norm(yh, axis=1, keepdims=True)
-    norm_y = xp.linalg.norm(y, axis=1, keepdims=True)
-    mul_norms = norm_yh * norm_y
-    cosine = (yh * y).sum(axis=1, keepdims=True) / mul_norms
-    d_yh = (y / mul_norms) - (cosine * (yh / norm_yh ** 2))
-    losses = xp.abs(cosine - 1)
-    if ignore_zeros:
-        # If the target was a zero vector, don't count it in the loss.
-        d_yh[zero_indices] = 0
-        losses[zero_indices] = 0
-    loss = losses.sum()
-    return loss, -d_yh

spacy/about.py

@@ -1,6 +1,6 @@
 # fmt: off
 __title__ = "spacy"
-__version__ = "2.2.3"
+__version__ = "3.0.0.dev3"
 __release__ = True
 __download_url__ = "https://github.com/explosion/spacy-models/releases/download"
 __compatibility__ = "https://raw.githubusercontent.com/explosion/spacy-models/master/compatibility.json"

spacy/analysis.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-from collections import OrderedDict
 from wasabi import Printer
 
 from .tokens import Doc, Token, Span
@@ -23,7 +19,7 @@ def analyze_pipes(pipeline, name, pipe, index, warn=True):
     assert pipeline[index][0] == name
     prev_pipes = pipeline[:index]
     pipe_requires = getattr(pipe, "requires", [])
-    requires = OrderedDict([(annot, False) for annot in pipe_requires])
+    requires = {annot: False for annot in pipe_requires}
     if requires:
         for prev_name, prev_pipe in prev_pipes:
             prev_assigns = getattr(prev_pipe, "assigns", [])
@@ -98,15 +94,15 @@ def validate_attrs(values):
                 for ext_attr, ext_value in value.items():
                     # We don't check whether the attribute actually exists
                     if ext_value is not True:  # attr is something like doc._.x.y
-                        good = "{}._.{}".format(obj_key, ext_attr)
+                        good = f"{obj_key}._.{ext_attr}"
-                        bad = "{}.{}".format(good, ".".join(ext_value))
+                        bad = f"{good}.{'.'.join(ext_value)}"
                         raise ValueError(Errors.E183.format(attr=bad, solution=good))
                 continue  # we can't validate those further
             if attr.endswith("_"):  # attr is something like "token.pos_"
                 raise ValueError(Errors.E184.format(attr=attr, solution=attr[:-1]))
             if value is not True:  # attr is something like doc.x.y
-                good = "{}.{}".format(obj_key, attr)
+                good = f"{obj_key}.{attr}"
-                bad = "{}.{}".format(good, ".".join(value))
+                bad = f"{good}.{'.'.join(value)}"
                 raise ValueError(Errors.E183.format(attr=bad, solution=good))
             obj = objs[obj_key]
             if not hasattr(obj, attr):
@@ -168,11 +164,10 @@ def print_summary(nlp, pretty=True, no_print=False):
     msg.table(overview, header=header, divider=True, multiline=True)
     n_problems = sum(len(p) for p in problems.values())
     if any(p for p in problems.values()):
-        msg.divider("Problems ({})".format(n_problems))
+        msg.divider(f"Problems ({n_problems})")
         for name, problem in problems.items():
             if problem:
-                problem = ", ".join(problem)
+                msg.warn(f"'{name}' requirements not met: {', '.join(problem)}")
-                msg.warn("'{}' requirements not met: {}".format(name, problem))
     else:
         msg.good("No problems found.")
     if no_print:

spacy/attrs.pxd

@@ -91,4 +91,5 @@ cdef enum attr_id_t:
 
     LANG
     ENT_KB_ID = symbols.ENT_KB_ID
+    MORPH
     ENT_ID = symbols.ENT_ID

spacy/attrs.pyx

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 
 IDS = {
     "": NULL_ATTR,
@@ -91,6 +88,7 @@ IDS = {
     "SPACY": SPACY,
     "PROB": PROB,
     "LANG": LANG,
+    "MORPH": MORPH,
 }
 
 

spacy/cli/__init__.py

@@ -1,12 +1,21 @@
+from wasabi import msg
+
 from .download import download  # noqa: F401
 from .info import info  # noqa: F401
-from .link import link  # noqa: F401
 from .package import package  # noqa: F401
 from .profile import profile  # noqa: F401
 from .train import train  # noqa: F401
+from .train_from_config import train_from_config_cli  # noqa: F401
 from .pretrain import pretrain  # noqa: F401
 from .debug_data import debug_data  # noqa: F401
 from .evaluate import evaluate  # noqa: F401
 from .convert import convert  # noqa: F401
 from .init_model import init_model  # noqa: F401
 from .validate import validate  # noqa: F401
+
+
+def link(*args, **kwargs):
+    msg.warn(
+        "As of spaCy v3.0, model symlinks are deprecated. You can load models "
+        "using their full names or from a directory path."
+    )

spacy/cli/_schemas.py

@@ -1,220 +0,0 @@
-# coding: utf-8
-from __future__ import unicode_literals
-
-
-# NB: This schema describes the new format of the training data, see #2928
-TRAINING_SCHEMA = {
-    "$schema": "http://json-schema.org/draft-06/schema",
-    "title": "Training data for spaCy models",
-    "type": "array",
-    "items": {
-        "type": "object",
-        "properties": {
-            "text": {
-                "title": "The text of the training example",
-                "type": "string",
-                "minLength": 1,
-            },
-            "ents": {
-                "title": "Named entity spans in the text",
-                "type": "array",
-                "items": {
-                    "type": "object",
-                    "properties": {
-                        "start": {
-                            "title": "Start character offset of the span",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                        "end": {
-                            "title": "End character offset of the span",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                        "label": {
-                            "title": "Entity label",
-                            "type": "string",
-                            "minLength": 1,
-                            "pattern": "^[A-Z0-9]*$",
-                        },
-                    },
-                    "required": ["start", "end", "label"],
-                },
-            },
-            "sents": {
-                "title": "Sentence spans in the text",
-                "type": "array",
-                "items": {
-                    "type": "object",
-                    "properties": {
-                        "start": {
-                            "title": "Start character offset of the span",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                        "end": {
-                            "title": "End character offset of the span",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                    },
-                    "required": ["start", "end"],
-                },
-            },
-            "cats": {
-                "title": "Text categories for the text classifier",
-                "type": "object",
-                "patternProperties": {
-                    "*": {
-                        "title": "A text category",
-                        "oneOf": [
-                            {"type": "boolean"},
-                            {"type": "number", "minimum": 0},
-                        ],
-                    }
-                },
-                "propertyNames": {"pattern": "^[A-Z0-9]*$", "minLength": 1},
-            },
-            "tokens": {
-                "title": "The tokens in the text",
-                "type": "array",
-                "items": {
-                    "type": "object",
-                    "minProperties": 1,
-                    "properties": {
-                        "id": {
-                            "title": "Token ID, usually token index",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                        "start": {
-                            "title": "Start character offset of the token",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                        "end": {
-                            "title": "End character offset of the token",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                        "pos": {
-                            "title": "Coarse-grained part-of-speech tag",
-                            "type": "string",
-                            "minLength": 1,
-                        },
-                        "tag": {
-                            "title": "Fine-grained part-of-speech tag",
-                            "type": "string",
-                            "minLength": 1,
-                        },
-                        "dep": {
-                            "title": "Dependency label",
-                            "type": "string",
-                            "minLength": 1,
-                        },
-                        "head": {
-                            "title": "Index of the token's head",
-                            "type": "integer",
-                            "minimum": 0,
-                        },
-                    },
-                    "required": ["start", "end"],
-                },
-            },
-            "_": {"title": "Custom user space", "type": "object"},
-        },
-        "required": ["text"],
-    },
-}
-
-META_SCHEMA = {
-    "$schema": "http://json-schema.org/draft-06/schema",
-    "type": "object",
-    "properties": {
-        "lang": {
-            "title": "Two-letter language code, e.g. 'en'",
-            "type": "string",
-            "minLength": 2,
-            "maxLength": 2,
-            "pattern": "^[a-z]*$",
-        },
-        "name": {
-            "title": "Model name",
-            "type": "string",
-            "minLength": 1,
-            "pattern": "^[a-z_]*$",
-        },
-        "version": {
-            "title": "Model version",
-            "type": "string",
-            "minLength": 1,
-            "pattern": "^[0-9a-z.-]*$",
-        },
-        "spacy_version": {
-            "title": "Compatible spaCy version identifier",
-            "type": "string",
-            "minLength": 1,
-            "pattern": "^[0-9a-z.-><=]*$",
-        },
-        "parent_package": {
-            "title": "Name of parent spaCy package, e.g. spacy or spacy-nightly",
-            "type": "string",
-            "minLength": 1,
-            "default": "spacy",
-        },
-        "pipeline": {
-            "title": "Names of pipeline components",
-            "type": "array",
-            "items": {"type": "string", "minLength": 1},
-        },
-        "description": {"title": "Model description", "type": "string"},
-        "license": {"title": "Model license", "type": "string"},
-        "author": {"title": "Model author name", "type": "string"},
-        "email": {"title": "Model author email", "type": "string", "format": "email"},
-        "url": {"title": "Model author URL", "type": "string", "format": "uri"},
-        "sources": {
-            "title": "Training data sources",
-            "type": "array",
-            "items": {"type": "string"},
-        },
-        "vectors": {
-            "title": "Included word vectors",
-            "type": "object",
-            "properties": {
-                "keys": {
-                    "title": "Number of unique keys",
-                    "type": "integer",
-                    "minimum": 0,
-                },
-                "vectors": {
-                    "title": "Number of unique vectors",
-                    "type": "integer",
-                    "minimum": 0,
-                },
-                "width": {
-                    "title": "Number of dimensions",
-                    "type": "integer",
-                    "minimum": 0,
-                },
-            },
-        },
-        "accuracy": {
-            "title": "Accuracy numbers",
-            "type": "object",
-            "patternProperties": {"*": {"type": "number", "minimum": 0.0}},
-        },
-        "speed": {
-            "title": "Speed evaluation numbers",
-            "type": "object",
-            "patternProperties": {
-                "*": {
-                    "oneOf": [
-                        {"type": "number", "minimum": 0.0},
-                        {"type": "integer", "minimum": 0},
-                    ]
-                }
-            },
-        },
-    },
-    "required": ["lang", "name", "version"],
-}

spacy/cli/convert.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-import plac
 from pathlib import Path
 from wasabi import Printer
 import srsly
@@ -29,27 +25,20 @@ FILE_TYPES = ("json", "jsonl", "msg")
 FILE_TYPES_STDOUT = ("json", "jsonl")
 
 
-@plac.annotations(
-    input_file=("Input file", "positional", None, str),
-    output_dir=("Output directory. '-' for stdout.", "positional", None, str),
-    file_type=("Type of data to produce: {}".format(FILE_TYPES), "option", "t", str),
-    n_sents=("Number of sentences per doc (0 to disable)", "option", "n", int),
-    seg_sents=("Segment sentences (for -c ner)", "flag", "s"),
-    model=("Model for sentence segmentation (for -s)", "option", "b", str),
-    converter=("Converter: {}".format(tuple(CONVERTERS.keys())), "option", "c", str),
-    lang=("Language (if tokenizer required)", "option", "l", str),
-    morphology=("Enable appending morphology to tags", "flag", "m", bool),
-)
 def convert(
-    input_file,
+    # fmt: off
-    output_dir="-",
+    input_file: ("Input file", "positional", None, str),
-    file_type="json",
+    output_dir: ("Output directory. '-' for stdout.", "positional", None, str) = "-",
-    n_sents=1,
+    file_type: (f"Type of data to produce: {FILE_TYPES}", "option", "t", str, FILE_TYPES) = "json",
-    seg_sents=False,
+    n_sents: ("Number of sentences per doc (0 to disable)", "option", "n", int) = 1,
-    model=None,
+    seg_sents: ("Segment sentences (for -c ner)", "flag", "s") = False,
-    morphology=False,
+    model: ("Model for sentence segmentation (for -s)", "option", "b", str) = None,
-    converter="auto",
+    morphology: ("Enable appending morphology to tags", "flag", "m", bool) = False,
-    lang=None,
+    merge_subtokens: ("Merge CoNLL-U subtokens", "flag", "T", bool) = False,
+    converter: (f"Converter: {tuple(CONVERTERS.keys())}", "option", "c", str) = "auto",
+    ner_map_path: ("NER tag mapping (as JSON-encoded dict of entity types)", "option", "N", Path) = None,
+    lang: ("Language (if tokenizer required)", "option", "l", str) = None,
+    # fmt: on
 ):
     """
     Convert files into JSON format for use with train command and other
@@ -60,16 +49,10 @@ def convert(
     no_print = output_dir == "-"
     msg = Printer(no_print=no_print)
     input_path = Path(input_file)
-    if file_type not in FILE_TYPES:
-        msg.fail(
-            "Unknown file type: '{}'".format(file_type),
-            "Supported file types: '{}'".format(", ".join(FILE_TYPES)),
-            exits=1,
-        )
     if file_type not in FILE_TYPES_STDOUT and output_dir == "-":
         # TODO: support msgpack via stdout in srsly?
         msg.fail(
-            "Can't write .{} data to stdout.".format(file_type),
+            f"Can't write .{file_type} data to stdout",
             "Please specify an output directory.",
             exits=1,
         )
@@ -93,21 +76,26 @@ def convert(
                 "Can't automatically detect NER format. Conversion may not succeed. See https://spacy.io/api/cli#convert"
             )
     if converter not in CONVERTERS:
-        msg.fail("Can't find converter for {}".format(converter), exits=1)
+        msg.fail(f"Can't find converter for {converter}", exits=1)
+    ner_map = None
+    if ner_map_path is not None:
+        ner_map = srsly.read_json(ner_map_path)
     # Use converter function to convert data
     func = CONVERTERS[converter]
     data = func(
         input_data,
         n_sents=n_sents,
         seg_sents=seg_sents,
-        use_morphology=morphology,
+        append_morphology=morphology,
+        merge_subtokens=merge_subtokens,
         lang=lang,
         model=model,
         no_print=no_print,
+        ner_map=ner_map,
     )
     if output_dir != "-":
         # Export data to a file
-        suffix = ".{}".format(file_type)
+        suffix = f".{file_type}"
         output_file = Path(output_dir) / Path(input_path.parts[-1]).with_suffix(suffix)
         if file_type == "json":
             srsly.write_json(output_file, data)
@@ -115,9 +103,7 @@ def convert(
             srsly.write_jsonl(output_file, data)
         elif file_type == "msg":
             srsly.write_msgpack(output_file, data)
-        msg.good(
+        msg.good(f"Generated output file ({len(data)} documents): {output_file}")
-            "Generated output file ({} documents): {}".format(len(data), output_file)
-        )
     else:
         # Print to stdout
         if file_type == "json":

spacy/cli/converters/conll_ner2json.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from wasabi import Printer
 
 from ...gold import iob_to_biluo
@@ -64,9 +61,9 @@ def conll_ner2json(
         # sentence segmentation required for document segmentation
         if n_sents > 0 and not seg_sents:
             msg.warn(
-                "No sentence boundaries found to use with option `-n {}`. "
+                f"No sentence boundaries found to use with option `-n {n_sents}`. "
-                "Use `-s` to automatically segment sentences or `-n 0` "
+                f"Use `-s` to automatically segment sentences or `-n 0` "
-                "to disable.".format(n_sents)
+                f"to disable."
             )
         else:
             n_sents_info(msg, n_sents)
@@ -129,7 +126,7 @@ def segment_sents_and_docs(doc, n_sents, doc_delimiter, model=None, msg=None):
     if model:
         nlp = load_model(model)
         if "parser" in nlp.pipe_names:
-            msg.info("Segmenting sentences with parser from model '{}'.".format(model))
+            msg.info(f"Segmenting sentences with parser from model '{model}'.")
             sentencizer = nlp.get_pipe("parser")
     if not sentencizer:
         msg.info(
@@ -166,7 +163,7 @@ def segment_docs(input_data, n_sents, doc_delimiter):
 
 
 def n_sents_info(msg, n_sents):
-    msg.info("Grouping every {} sentences into a document.".format(n_sents))
+    msg.info(f"Grouping every {n_sents} sentences into a document.")
     if n_sents == 1:
         msg.warn(
             "To generate better training data, you may want to group "

spacy/cli/converters/conllu2json.py

@@ -1,141 +1,348 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 import re
 
-from ...gold import iob_to_biluo
+from ...gold import Example
+from ...gold import iob_to_biluo, spans_from_biluo_tags, biluo_tags_from_offsets
+from ...language import Language
+from ...tokens import Doc, Token
+from .conll_ner2json import n_sents_info
+from wasabi import Printer
 
 
-def conllu2json(input_data, n_sents=10, use_morphology=False, lang=None, **_):
+def conllu2json(
+    input_data,
+    n_sents=10,
+    append_morphology=False,
+    lang=None,
+    ner_map=None,
+    merge_subtokens=False,
+    no_print=False,
+    **_
+):
     """
     Convert conllu files into JSON format for use with train cli.
-    use_morphology parameter enables appending morphology to tags, which is
+    append_morphology parameter enables appending morphology to tags, which is
     useful for languages such as Spanish, where UD tags are not so rich.
 
     Extract NER tags if available and convert them so that they follow
     BILUO and the Wikipedia scheme
     """
-    # by @dvsrepo, via #11 explosion/spacy-dev-resources
+    MISC_NER_PATTERN = "\|?(?:name=)?(([A-Z_]+)-([A-Z_]+)|O)\|?"
-    # by @katarkor
+    msg = Printer(no_print=no_print)
+    n_sents_info(msg, n_sents)
     docs = []
+    raw = ""
     sentences = []
-    conll_tuples = read_conllx(input_data, use_morphology=use_morphology)
+    conll_data = read_conllx(
-    checked_for_ner = False
+        input_data,
-    has_ner_tags = False
+        append_morphology=append_morphology,
-    for i, (raw_text, tokens) in enumerate(conll_tuples):
+        ner_tag_pattern=MISC_NER_PATTERN,
-        sentence, brackets = tokens[0]
+        ner_map=ner_map,
-        if not checked_for_ner:
+        merge_subtokens=merge_subtokens,
-            has_ner_tags = is_ner(sentence[5][0])
+    )
-            checked_for_ner = True
+    has_ner_tags = has_ner(input_data, ner_tag_pattern=MISC_NER_PATTERN)
-        sentences.append(generate_sentence(sentence, has_ner_tags))
+    for i, example in enumerate(conll_data):
+        raw += example.text
+        sentences.append(
+            generate_sentence(
+                example.token_annotation,
+                has_ner_tags,
+                MISC_NER_PATTERN,
+                ner_map=ner_map,
+            )
+        )
         # Real-sized documents could be extracted using the comments on the
-        # conluu document
+        # conllu document
         if len(sentences) % n_sents == 0:
-            doc = create_doc(sentences, i)
+            doc = create_json_doc(raw, sentences, i)
             docs.append(doc)
+            raw = ""
             sentences = []
     if sentences:
-        doc = create_doc(sentences, i)
+        doc = create_json_doc(raw, sentences, i)
         docs.append(doc)
     return docs
 
 
-def is_ner(tag):
+def has_ner(input_data, ner_tag_pattern):
     """
     Check the 10th column of the first token to determine if the file contains
     NER tags
     """
-    tag_match = re.match("([A-Z_]+)-([A-Z_]+)", tag)
-    if tag_match:
-        return True
-    elif tag == "O":
-        return True
-    else:
-        return False
-
-
-def read_conllx(input_data, use_morphology=False, n=0):
-    i = 0
     for sent in input_data.strip().split("\n\n"):
         lines = sent.strip().split("\n")
         if lines:
             while lines[0].startswith("#"):
                 lines.pop(0)
-            tokens = []
+            if lines:
-            for line in lines:
+                parts = lines[0].split("\t")
+                id_, word, lemma, pos, tag, morph, head, dep, _1, misc = parts
+                if re.search(ner_tag_pattern, misc):
+                    return True
+                else:
+                    return False
 
+
+def read_conllx(
+    input_data,
+    append_morphology=False,
+    merge_subtokens=False,
+    ner_tag_pattern="",
+    ner_map=None,
+):
+    """ Yield examples, one for each sentence """
+    vocab = Language.Defaults.create_vocab()  # need vocab to make a minimal Doc
+    for sent in input_data.strip().split("\n\n"):
+        lines = sent.strip().split("\n")
+        if lines:
+            while lines[0].startswith("#"):
+                lines.pop(0)
+            example = example_from_conllu_sentence(
+                vocab,
+                lines,
+                ner_tag_pattern,
+                merge_subtokens=merge_subtokens,
+                append_morphology=append_morphology,
+                ner_map=ner_map,
+            )
+            yield example
+
+
+def get_entities(lines, tag_pattern, ner_map=None):
+    """Find entities in the MISC column according to the pattern and map to
+    final entity type with `ner_map` if mapping present. Entity tag is 'O' if
+    the pattern is not matched.
+
+    lines (unicode): CONLL-U lines for one sentences
+    tag_pattern (unicode): Regex pattern for entity tag
+    ner_map (dict): Map old NER tag names to new ones, '' maps to O.
+    RETURNS (list): List of BILUO entity tags
+    """
+    miscs = []
+    for line in lines:
         parts = line.split("\t")
-                id_, word, lemma, pos, tag, morph, head, dep, _1, iob = parts
+        id_, word, lemma, pos, tag, morph, head, dep, _1, misc = parts
         if "-" in id_ or "." in id_:
             continue
-                try:
+        miscs.append(misc)
-                    id_ = int(id_) - 1
-                    head = (int(head) - 1) if head not in ["0", "_"] else id_
-                    dep = "ROOT" if dep == "root" else dep
-                    tag = pos if tag == "_" else tag
-                    tag = tag + "__" + morph if use_morphology else tag
-                    iob = iob if iob else "O"
-                    tokens.append((id_, word, tag, head, dep, iob))
-                except:  # noqa: E722
-                    print(line)
-                    raise
-            tuples = [list(t) for t in zip(*tokens)]
-            yield (None, [[tuples, []]])
-            i += 1
-            if n >= 1 and i >= n:
-                break
 
-
+    iob = []
-def simplify_tags(iob):
+    for misc in miscs:
-    """
+        tag_match = re.search(tag_pattern, misc)
-    Simplify tags obtained from the dataset in order to follow Wikipedia
+        iob_tag = "O"
-    scheme (PER, LOC, ORG, MISC). 'PER', 'LOC' and 'ORG' keep their tags, while
-    'GPE_LOC' is simplified to 'LOC', 'GPE_ORG' to 'ORG' and all remaining tags to
-    'MISC'.
-    """
-    new_iob = []
-    for tag in iob:
-        tag_match = re.match("([A-Z_]+)-([A-Z_]+)", tag)
         if tag_match:
-            prefix = tag_match.group(1)
+            prefix = tag_match.group(2)
-            suffix = tag_match.group(2)
+            suffix = tag_match.group(3)
-            if suffix == "GPE_LOC":
+            if prefix and suffix:
-                suffix = "LOC"
+                iob_tag = prefix + "-" + suffix
-            elif suffix == "GPE_ORG":
+                if ner_map:
-                suffix = "ORG"
+                    suffix = ner_map.get(suffix, suffix)
-            elif suffix != "PER" and suffix != "LOC" and suffix != "ORG":
+                    if suffix == "":
-                suffix = "MISC"
+                        iob_tag = "O"
-            tag = prefix + "-" + suffix
+                    else:
-        new_iob.append(tag)
+                        iob_tag = prefix + "-" + suffix
-    return new_iob
+        iob.append(iob_tag)
+    return iob_to_biluo(iob)
 
 
-def generate_sentence(sent, has_ner_tags):
+def generate_sentence(token_annotation, has_ner_tags, tag_pattern, ner_map=None):
-    (id_, word, tag, head, dep, iob) = sent
     sentence = {}
     tokens = []
-    if has_ner_tags:
+    for i, id_ in enumerate(token_annotation.ids):
-        iob = simplify_tags(iob)
-        biluo = iob_to_biluo(iob)
-    for i, id in enumerate(id_):
         token = {}
-        token["id"] = id
+        token["id"] = id_
-        token["orth"] = word[i]
+        token["orth"] = token_annotation.get_word(i)
-        token["tag"] = tag[i]
+        token["tag"] = token_annotation.get_tag(i)
-        token["head"] = head[i] - id
+        token["pos"] = token_annotation.get_pos(i)
-        token["dep"] = dep[i]
+        token["lemma"] = token_annotation.get_lemma(i)
+        token["morph"] = token_annotation.get_morph(i)
+        token["head"] = token_annotation.get_head(i) - id_
+        token["dep"] = token_annotation.get_dep(i)
         if has_ner_tags:
-            token["ner"] = biluo[i]
+            token["ner"] = token_annotation.get_entity(i)
         tokens.append(token)
     sentence["tokens"] = tokens
     return sentence
 
 
-def create_doc(sentences, id):
+def create_json_doc(raw, sentences, id_):
     doc = {}
     paragraph = {}
-    doc["id"] = id
+    doc["id"] = id_
     doc["paragraphs"] = []
+    paragraph["raw"] = raw.strip()
     paragraph["sentences"] = sentences
     doc["paragraphs"].append(paragraph)
     return doc
+
+
+def example_from_conllu_sentence(
+    vocab,
+    lines,
+    ner_tag_pattern,
+    merge_subtokens=False,
+    append_morphology=False,
+    ner_map=None,
+):
+    """Create an Example from the lines for one CoNLL-U sentence, merging
+    subtokens and appending morphology to tags if required.
+
+    lines (unicode): The non-comment lines for a CoNLL-U sentence
+    ner_tag_pattern (unicode): The regex pattern for matching NER in MISC col
+    RETURNS (Example): An example containing the annotation
+    """
+    # create a Doc with each subtoken as its own token
+    # if merging subtokens, each subtoken orth is the merged subtoken form
+    if not Token.has_extension("merged_orth"):
+        Token.set_extension("merged_orth", default="")
+    if not Token.has_extension("merged_lemma"):
+        Token.set_extension("merged_lemma", default="")
+    if not Token.has_extension("merged_morph"):
+        Token.set_extension("merged_morph", default="")
+    if not Token.has_extension("merged_spaceafter"):
+        Token.set_extension("merged_spaceafter", default="")
+    words, spaces, tags, poses, morphs, lemmas = [], [], [], [], [], []
+    heads, deps = [], []
+    subtok_word = ""
+    in_subtok = False
+    for i in range(len(lines)):
+        line = lines[i]
+        parts = line.split("\t")
+        id_, word, lemma, pos, tag, morph, head, dep, _1, misc = parts
+        if "." in id_:
+            continue
+        if "-" in id_:
+            in_subtok = True
+        if "-" in id_:
+            in_subtok = True
+            subtok_word = word
+            subtok_start, subtok_end = id_.split("-")
+            subtok_spaceafter = "SpaceAfter=No" not in misc
+            continue
+        if merge_subtokens and in_subtok:
+            words.append(subtok_word)
+        else:
+            words.append(word)
+        if in_subtok:
+            if id_ == subtok_end:
+                spaces.append(subtok_spaceafter)
+            else:
+                spaces.append(False)
+        elif "SpaceAfter=No" in misc:
+            spaces.append(False)
+        else:
+            spaces.append(True)
+        if in_subtok and id_ == subtok_end:
+            subtok_word = ""
+            in_subtok = False
+        id_ = int(id_) - 1
+        head = (int(head) - 1) if head not in ("0", "_") else id_
+        tag = pos if tag == "_" else tag
+        morph = morph if morph != "_" else ""
+        dep = "ROOT" if dep == "root" else dep
+        lemmas.append(lemma)
+        poses.append(pos)
+        tags.append(tag)
+        morphs.append(morph)
+        heads.append(head)
+        deps.append(dep)
+
+    doc = Doc(vocab, words=words, spaces=spaces)
+    for i in range(len(doc)):
+        doc[i].tag_ = tags[i]
+        doc[i].pos_ = poses[i]
+        doc[i].dep_ = deps[i]
+        doc[i].lemma_ = lemmas[i]
+        doc[i].head = doc[heads[i]]
+        doc[i]._.merged_orth = words[i]
+        doc[i]._.merged_morph = morphs[i]
+        doc[i]._.merged_lemma = lemmas[i]
+        doc[i]._.merged_spaceafter = spaces[i]
+    ents = get_entities(lines, ner_tag_pattern, ner_map)
+    doc.ents = spans_from_biluo_tags(doc, ents)
+    doc.is_parsed = True
+    doc.is_tagged = True
+
+    if merge_subtokens:
+        doc = merge_conllu_subtokens(lines, doc)
+
+    # create Example from custom Doc annotation
+    ids, words, tags, heads, deps = [], [], [], [], []
+    pos, lemmas, morphs, spaces = [], [], [], []
+    for i, t in enumerate(doc):
+        ids.append(i)
+        words.append(t._.merged_orth)
+        if append_morphology and t._.merged_morph:
+            tags.append(t.tag_ + "__" + t._.merged_morph)
+        else:
+            tags.append(t.tag_)
+        pos.append(t.pos_)
+        morphs.append(t._.merged_morph)
+        lemmas.append(t._.merged_lemma)
+        heads.append(t.head.i)
+        deps.append(t.dep_)
+        spaces.append(t._.merged_spaceafter)
+    ent_offsets = [(e.start_char, e.end_char, e.label_) for e in doc.ents]
+    ents = biluo_tags_from_offsets(doc, ent_offsets)
+    raw = ""
+    for word, space in zip(words, spaces):
+        raw += word
+        if space:
+            raw += " "
+    example = Example(doc=raw)
+    example.set_token_annotation(
+        ids=ids,
+        words=words,
+        tags=tags,
+        pos=pos,
+        morphs=morphs,
+        lemmas=lemmas,
+        heads=heads,
+        deps=deps,
+        entities=ents,
+    )
+    return example
+
+
+def merge_conllu_subtokens(lines, doc):
+    # identify and process all subtoken spans to prepare attrs for merging
+    subtok_spans = []
+    for line in lines:
+        parts = line.split("\t")
+        id_, word, lemma, pos, tag, morph, head, dep, _1, misc = parts
+        if "-" in id_:
+            subtok_start, subtok_end = id_.split("-")
+            subtok_span = doc[int(subtok_start) - 1 : int(subtok_end)]
+            subtok_spans.append(subtok_span)
+            # create merged tag, morph, and lemma values
+            tags = []
+            morphs = {}
+            lemmas = []
+            for token in subtok_span:
+                tags.append(token.tag_)
+                lemmas.append(token.lemma_)
+                if token._.merged_morph:
+                    for feature in token._.merged_morph.split("|"):
+                        field, values = feature.split("=", 1)
+                        if field not in morphs:
+                            morphs[field] = set()
+                        for value in values.split(","):
+                            morphs[field].add(value)
+            # create merged features for each morph field
+            for field, values in morphs.items():
+                morphs[field] = field + "=" + ",".join(sorted(values))
+            # set the same attrs on all subtok tokens so that whatever head the
+            # retokenizer chooses, the final attrs are available on that token
+            for token in subtok_span:
+                token._.merged_orth = token.orth_
+                token._.merged_lemma = " ".join(lemmas)
+                token.tag_ = "_".join(tags)
+                token._.merged_morph = "|".join(sorted(morphs.values()))
+                token._.merged_spaceafter = (
+                    True if subtok_span[-1].whitespace_ else False
+                )
+
+    with doc.retokenize() as retokenizer:
+        for span in subtok_spans:
+            retokenizer.merge(span)
+
+    return doc

spacy/cli/converters/iob2json.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from wasabi import Printer
 
 from ...gold import iob_to_biluo

spacy/cli/converters/jsonl2json.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 import srsly
 
 from ...gold import docs_to_json

spacy/cli/debug_data.py

@@ -1,9 +1,5 @@
-# coding: utf8
-from __future__ import unicode_literals, print_function
-
 from pathlib import Path
 from collections import Counter
-import plac
 import sys
 import srsly
 from wasabi import Printer, MESSAGES
@@ -22,30 +18,18 @@ BLANK_MODEL_MIN_THRESHOLD = 100
 BLANK_MODEL_THRESHOLD = 2000
 
 
-@plac.annotations(
-    lang=("model language", "positional", None, str),
-    train_path=("location of JSON-formatted training data", "positional", None, Path),
-    dev_path=("location of JSON-formatted development data", "positional", None, Path),
-    base_model=("name of model to update (optional)", "option", "b", str),
-    pipeline=(
-        "Comma-separated names of pipeline components to train",
-        "option",
-        "p",
-        str,
-    ),
-    ignore_warnings=("Ignore warnings, only show stats and errors", "flag", "IW", bool),
-    verbose=("Print additional information and explanations", "flag", "V", bool),
-    no_format=("Don't pretty-print the results", "flag", "NF", bool),
-)
 def debug_data(
-    lang,
+    # fmt: off
-    train_path,
+    lang: ("Model language", "positional", None, str),
-    dev_path,
+    train_path: ("Location of JSON-formatted training data", "positional", None, Path),
-    base_model=None,
+    dev_path: ("Location of JSON-formatted development data", "positional", None, Path),
-    pipeline="tagger,parser,ner",
+    tag_map_path: ("Location of JSON-formatted tag map", "option", "tm", Path) = None,
-    ignore_warnings=False,
+    base_model: ("Name of model to update (optional)", "option", "b", str) = None,
-    verbose=False,
+    pipeline: ("Comma-separated names of pipeline components to train", "option", "p", str) = "tagger,parser,ner",
-    no_format=False,
+    ignore_warnings: ("Ignore warnings, only show stats and errors", "flag", "IW", bool) = False,
+    verbose: ("Print additional information and explanations", "flag", "V", bool) = False,
+    no_format: ("Don't pretty-print the results", "flag", "NF", bool) = False,
+    # fmt: on
 ):
     """
     Analyze, debug and validate your training and development data, get useful
@@ -60,6 +44,10 @@ def debug_data(
     if not dev_path.exists():
         msg.fail("Development data not found", dev_path, exits=1)
 
+    tag_map = {}
+    if tag_map_path is not None:
+        tag_map = srsly.read_json(tag_map_path)
+
     # Initialize the model and pipeline
     pipeline = [p.strip() for p in pipeline.split(",")]
     if base_model:
@@ -67,6 +55,8 @@ def debug_data(
     else:
         lang_cls = get_lang_class(lang)
         nlp = lang_cls()
+    # Update tag map with provided mapping
+    nlp.vocab.morphology.tag_map.update(tag_map)
 
     msg.divider("Data format validation")
 
@@ -80,20 +70,16 @@ def debug_data(
     with msg.loading("Loading corpus..."):
         corpus = GoldCorpus(train_path, dev_path)
         try:
-            train_docs = list(corpus.train_docs(nlp))
+            train_dataset = list(corpus.train_dataset(nlp))
-            train_docs_unpreprocessed = list(
+            train_dataset_unpreprocessed = list(
-                corpus.train_docs_without_preprocessing(nlp)
+                corpus.train_dataset_without_preprocessing(nlp)
             )
         except ValueError as e:
-            loading_train_error_message = "Training data cannot be loaded: {}".format(
+            loading_train_error_message = f"Training data cannot be loaded: {e}"
-                str(e)
-            )
         try:
-            dev_docs = list(corpus.dev_docs(nlp))
+            dev_dataset = list(corpus.dev_dataset(nlp))
         except ValueError as e:
-            loading_dev_error_message = "Development data cannot be loaded: {}".format(
+            loading_dev_error_message = f"Development data cannot be loaded: {e}"
-                str(e)
-            )
     if loading_train_error_message or loading_dev_error_message:
         if loading_train_error_message:
             msg.fail(loading_train_error_message)
@@ -102,80 +88,68 @@ def debug_data(
         sys.exit(1)
     msg.good("Corpus is loadable")
 
-    # Create all gold data here to avoid iterating over the train_docs constantly
+    # Create all gold data here to avoid iterating over the train_dataset constantly
-    gold_train_data = _compile_gold(train_docs, pipeline)
+    gold_train_data = _compile_gold(train_dataset, pipeline)
-    gold_train_unpreprocessed_data = _compile_gold(train_docs_unpreprocessed, pipeline)
+    gold_train_unpreprocessed_data = _compile_gold(
-    gold_dev_data = _compile_gold(dev_docs, pipeline)
+        train_dataset_unpreprocessed, pipeline
+    )
+    gold_dev_data = _compile_gold(dev_dataset, pipeline)
 
     train_texts = gold_train_data["texts"]
     dev_texts = gold_dev_data["texts"]
 
     msg.divider("Training stats")
-    msg.text("Training pipeline: {}".format(", ".join(pipeline)))
+    msg.text(f"Training pipeline: {', '.join(pipeline)}")
     for pipe in [p for p in pipeline if p not in nlp.factories]:
-        msg.fail("Pipeline component '{}' not available in factories".format(pipe))
+        msg.fail(f"Pipeline component '{pipe}' not available in factories")
     if base_model:
-        msg.text("Starting with base model '{}'".format(base_model))
+        msg.text(f"Starting with base model '{base_model}'")
     else:
-        msg.text("Starting with blank model '{}'".format(lang))
+        msg.text(f"Starting with blank model '{lang}'")
-    msg.text("{} training docs".format(len(train_docs)))
+    msg.text(f"{len(train_dataset)} training docs")
-    msg.text("{} evaluation docs".format(len(dev_docs)))
+    msg.text(f"{len(dev_dataset)} evaluation docs")
 
-    if not len(dev_docs):
+    if not len(gold_dev_data):
         msg.fail("No evaluation docs")
     overlap = len(train_texts.intersection(dev_texts))
     if overlap:
-        msg.warn("{} training examples also in evaluation data".format(overlap))
+        msg.warn(f"{overlap} training examples also in evaluation data")
     else:
         msg.good("No overlap between training and evaluation data")
-    if not base_model and len(train_docs) < BLANK_MODEL_THRESHOLD:
+    if not base_model and len(train_dataset) < BLANK_MODEL_THRESHOLD:
-        text = "Low number of examples to train from a blank model ({})".format(
+        text = (
-            len(train_docs)
+            f"Low number of examples to train from a blank model ({len(train_dataset)})"
         )
-        if len(train_docs) < BLANK_MODEL_MIN_THRESHOLD:
+        if len(train_dataset) < BLANK_MODEL_MIN_THRESHOLD:
             msg.fail(text)
         else:
             msg.warn(text)
         msg.text(
-            "It's recommended to use at least {} examples (minimum {})".format(
+            f"It's recommended to use at least {BLANK_MODEL_THRESHOLD} examples "
-                BLANK_MODEL_THRESHOLD, BLANK_MODEL_MIN_THRESHOLD
+            f"(minimum {BLANK_MODEL_MIN_THRESHOLD})",
-            ),
             show=verbose,
         )
 
     msg.divider("Vocab & Vectors")
     n_words = gold_train_data["n_words"]
     msg.info(
-        "{} total {} in the data ({} unique)".format(
+        f"{n_words} total word(s) in the data ({len(gold_train_data['words'])} unique)"
-            n_words, "word" if n_words == 1 else "words", len(gold_train_data["words"])
-        )
     )
     if gold_train_data["n_misaligned_words"] > 0:
-        msg.warn(
+        n_misaligned = gold_train_data["n_misaligned_words"]
-            "{} misaligned tokens in the training data".format(
+        msg.warn(f"{n_misaligned} misaligned tokens in the training data")
-                gold_train_data["n_misaligned_words"]
-            )
-        )
     if gold_dev_data["n_misaligned_words"] > 0:
-        msg.warn(
+        n_misaligned = gold_dev_data["n_misaligned_words"]
-            "{} misaligned tokens in the dev data".format(
+        msg.warn(f"{n_misaligned} misaligned tokens in the dev data")
-                gold_dev_data["n_misaligned_words"]
-            )
-        )
     most_common_words = gold_train_data["words"].most_common(10)
     msg.text(
-        "10 most common words: {}".format(
+        f"10 most common words: {_format_labels(most_common_words, counts=True)}",
-            _format_labels(most_common_words, counts=True)
-        ),
         show=verbose,
     )
     if len(nlp.vocab.vectors):
         msg.info(
-            "{} vectors ({} unique keys, {} dimensions)".format(
+            f"{len(nlp.vocab.vectors)} vectors ({nlp.vocab.vectors.n_keys} "
-                len(nlp.vocab.vectors),
+            f"unique keys, {nlp.vocab.vectors_length} dimensions)"
-                nlp.vocab.vectors.n_keys,
-                nlp.vocab.vectors_length,
-            )
         )
     else:
         msg.info("No word vectors present in the model")
@@ -183,7 +157,7 @@ def debug_data(
     if "ner" in pipeline:
         # Get all unique NER labels present in the data
         labels = set(
-            label for label in gold_train_data["ner"] if label not in ("O", "-")
+            label for label in gold_train_data["ner"] if label not in ("O", "-", None)
         )
         label_counts = gold_train_data["ner"]
         model_labels = _get_labels_from_model(nlp, "ner")
@@ -196,19 +170,10 @@ def debug_data(
 
         msg.divider("Named Entity Recognition")
         msg.info(
-            "{} new {}, {} existing {}".format(
+            f"{len(new_labels)} new label(s), {len(existing_labels)} existing label(s)"
-                len(new_labels),
-                "label" if len(new_labels) == 1 else "labels",
-                len(existing_labels),
-                "label" if len(existing_labels) == 1 else "labels",
-            )
         )
         missing_values = label_counts["-"]
-        msg.text(
+        msg.text(f"{missing_values} missing value(s) (tokens with '-' label)")
-            "{} missing {} (tokens with '-' label)".format(
-                missing_values, "value" if missing_values == 1 else "values"
-            )
-        )
         for label in new_labels:
             if len(label) == 0:
                 msg.fail("Empty label found in new labels")
@@ -219,39 +184,28 @@ def debug_data(
                 if label != "-"
             ]
             labels_with_counts = _format_labels(labels_with_counts, counts=True)
-            msg.text("New: {}".format(labels_with_counts), show=verbose)
+            msg.text(f"New: {labels_with_counts}", show=verbose)
         if existing_labels:
-            msg.text(
+            msg.text(f"Existing: {_format_labels(existing_labels)}", show=verbose)
-                "Existing: {}".format(_format_labels(existing_labels)), show=verbose
-            )
-
         if gold_train_data["ws_ents"]:
-            msg.fail(
+            msg.fail(f"{gold_train_data['ws_ents']} invalid whitespace entity spans")
-                "{} invalid whitespace entity span(s)".format(gold_train_data["ws_ents"])
-            )
             has_ws_ents_error = True
 
         if gold_train_data["punct_ents"]:
-            msg.warn(
+            msg.warn(f"{gold_train_data['punct_ents']} entity span(s) with punctuation")
-                "{} entity span(s) with punctuation".format(gold_train_data["punct_ents"])
-            )
             has_punct_ents_warning = True
 
         for label in new_labels:
             if label_counts[label] <= NEW_LABEL_THRESHOLD:
                 msg.warn(
-                    "Low number of examples for new label '{}' ({})".format(
+                    f"Low number of examples for new label '{label}' ({label_counts[label]})"
-                        label, label_counts[label]
-                    )
                 )
                 has_low_data_warning = True
 
                 with msg.loading("Analyzing label distribution..."):
-                    neg_docs = _get_examples_without_label(train_docs, label)
+                    neg_docs = _get_examples_without_label(train_dataset, label)
                 if neg_docs == 0:
-                    msg.warn(
+                    msg.warn(f"No examples for texts WITHOUT new label '{label}'")
-                        "No examples for texts WITHOUT new label '{}'".format(label)
-                    )
                     has_no_neg_warning = True
 
         if not has_low_data_warning:
@@ -265,8 +219,8 @@ def debug_data(
 
         if has_low_data_warning:
             msg.text(
-                "To train a new entity type, your data should include at "
+                f"To train a new entity type, your data should include at "
-                "least {} instances of the new label".format(NEW_LABEL_THRESHOLD),
+                f"least {NEW_LABEL_THRESHOLD} instances of the new label",
                 show=verbose,
             )
         if has_no_neg_warning:
@@ -295,27 +249,21 @@ def debug_data(
         new_labels = [l for l in labels if l not in model_labels]
         existing_labels = [l for l in labels if l in model_labels]
         msg.info(
-            "Text Classification: {} new label(s), {} existing label(s)".format(
+            f"Text Classification: {len(new_labels)} new label(s), "
-                len(new_labels), len(existing_labels)
+            f"{len(existing_labels)} existing label(s)"
-            )
         )
         if new_labels:
             labels_with_counts = _format_labels(
                 gold_train_data["cats"].most_common(), counts=True
             )
-            msg.text("New: {}".format(labels_with_counts), show=verbose)
+            msg.text(f"New: {labels_with_counts}", show=verbose)
         if existing_labels:
-            msg.text(
+            msg.text(f"Existing: {_format_labels(existing_labels)}", show=verbose)
-                "Existing: {}".format(_format_labels(existing_labels)), show=verbose
-            )
         if set(gold_train_data["cats"]) != set(gold_dev_data["cats"]):
             msg.fail(
-                "The train and dev labels are not the same. "
+                f"The train and dev labels are not the same. "
-                "Train labels: {}. "
+                f"Train labels: {_format_labels(gold_train_data['cats'])}. "
-                "Dev labels: {}.".format(
+                f"Dev labels: {_format_labels(gold_dev_data['cats'])}."
-                    _format_labels(gold_train_data["cats"]),
-                    _format_labels(gold_dev_data["cats"]),
-                )
             )
         if gold_train_data["n_cats_multilabel"] > 0:
             msg.info(
@@ -344,28 +292,17 @@ def debug_data(
     if "tagger" in pipeline:
         msg.divider("Part-of-speech Tagging")
         labels = [label for label in gold_train_data["tags"]]
-        tag_map = nlp.Defaults.tag_map
+        tag_map = nlp.vocab.morphology.tag_map
-        msg.info(
+        msg.info(f"{len(labels)} label(s) in data ({len(tag_map)} label(s) in tag map)")
-            "{} {} in data ({} {} in tag map)".format(
-                len(labels),
-                "label" if len(labels) == 1 else "labels",
-                len(tag_map),
-                "label" if len(tag_map) == 1 else "labels",
-            )
-        )
         labels_with_counts = _format_labels(
             gold_train_data["tags"].most_common(), counts=True
         )
         msg.text(labels_with_counts, show=verbose)
         non_tagmap = [l for l in labels if l not in tag_map]
         if not non_tagmap:
-            msg.good("All labels present in tag map for language '{}'".format(nlp.lang))
+            msg.good(f"All labels present in tag map for language '{nlp.lang}'")
         for label in non_tagmap:
-            msg.fail(
+            msg.fail(f"Label '{label}' not found in tag map for language '{nlp.lang}'")
-                "Label '{}' not found in tag map for language '{}'".format(
-                    label, nlp.lang
-                )
-            )
 
     if "parser" in pipeline:
         has_low_data_warning = False
@@ -373,21 +310,18 @@ def debug_data(
 
         # profile sentence length
         msg.info(
-            "Found {} sentence{} with an average length of {:.1f} words.".format(
+            f"Found {gold_train_data['n_sents']} sentence(s) with an average "
-                gold_train_data["n_sents"],
+            f"length of {gold_train_data['n_words'] / gold_train_data['n_sents']:.1f} words."
-                "s" if len(train_docs) > 1 else "",
-                gold_train_data["n_words"] / gold_train_data["n_sents"],
-            )
         )
 
         # check for documents with multiple sentences
         sents_per_doc = gold_train_data["n_sents"] / len(gold_train_data["texts"])
         if sents_per_doc < 1.1:
             msg.warn(
-                "The training data contains {:.2f} sentences per "
+                f"The training data contains {sents_per_doc:.2f} sentences per "
-                "document. When there are very few documents containing more "
+                f"document. When there are very few documents containing more "
-                "than one sentence, the parser will not learn how to segment "
+                f"than one sentence, the parser will not learn how to segment "
-                "longer texts into sentences.".format(sents_per_doc)
+                f"longer texts into sentences."
             )
 
         # profile labels
@@ -398,32 +332,13 @@ def debug_data(
         labels_dev = [label for label in gold_dev_data["deps"]]
 
         if gold_train_unpreprocessed_data["n_nonproj"] > 0:
-            msg.info(
+            n_nonproj = gold_train_unpreprocessed_data["n_nonproj"]
-                "Found {} nonprojective train sentence{}".format(
+            msg.info(f"Found {n_nonproj} nonprojective train sentence(s)")
-                    gold_train_unpreprocessed_data["n_nonproj"],
-                    "s" if gold_train_unpreprocessed_data["n_nonproj"] > 1 else "",
-                )
-            )
         if gold_dev_data["n_nonproj"] > 0:
-            msg.info(
+            n_nonproj = gold_dev_data["n_nonproj"]
-                "Found {} nonprojective dev sentence{}".format(
+            msg.info(f"Found {n_nonproj} nonprojective dev sentence(s)")
-                    gold_dev_data["n_nonproj"],
+        msg.info(f"{labels_train_unpreprocessed} label(s) in train data")
-                    "s" if gold_dev_data["n_nonproj"] > 1 else "",
+        msg.info(f"{len(labels_train)} label(s) in projectivized train data")
-                )
-            )
-
-        msg.info(
-            "{} {} in train data".format(
-                len(labels_train_unpreprocessed),
-                "label" if len(labels_train) == 1 else "labels",
-            )
-        )
-        msg.info(
-            "{} {} in projectivized train data".format(
-                len(labels_train), "label" if len(labels_train) == 1 else "labels"
-            )
-        )
-
         labels_with_counts = _format_labels(
             gold_train_unpreprocessed_data["deps"].most_common(), counts=True
         )
@@ -433,9 +348,8 @@ def debug_data(
         for label in gold_train_unpreprocessed_data["deps"]:
             if gold_train_unpreprocessed_data["deps"][label] <= DEP_LABEL_THRESHOLD:
                 msg.warn(
-                    "Low number of examples for label '{}' ({})".format(
+                    f"Low number of examples for label '{label}' "
-                        label, gold_train_unpreprocessed_data["deps"][label]
+                    f"({gold_train_unpreprocessed_data['deps'][label]})"
-                    )
                 )
                 has_low_data_warning = True
 
@@ -444,22 +358,19 @@ def debug_data(
         for label in gold_train_data["deps"]:
             if gold_train_data["deps"][label] <= DEP_LABEL_THRESHOLD and "||" in label:
                 rare_projectivized_labels.append(
-                    "{}: {}".format(label, str(gold_train_data["deps"][label]))
+                    f"{label}: {gold_train_data['deps'][label]}"
                 )
 
         if len(rare_projectivized_labels) > 0:
             msg.warn(
-                "Low number of examples for {} label{} in the "
+                f"Low number of examples for {len(rare_projectivized_labels)} "
-                "projectivized dependency trees used for training. You may "
+                "label(s) in the projectivized dependency trees used for "
-                "want to projectivize labels such as punct before "
+                "training. You may want to projectivize labels such as punct "
-                "training in order to improve parser performance.".format(
+                "before training in order to improve parser performance."
-                    len(rare_projectivized_labels),
-                    "s" if len(rare_projectivized_labels) > 1 else "",
-                )
             )
             msg.warn(
-                "Projectivized labels with low numbers of examples: "
+                f"Projectivized labels with low numbers of examples: ",
-                "{}".format("\n".join(rare_projectivized_labels)),
+                ", ".join(rare_projectivized_labels),
                 show=verbose,
             )
             has_low_data_warning = True
@@ -467,50 +378,44 @@ def debug_data(
         # labels only in train
         if set(labels_train) - set(labels_dev):
             msg.warn(
-                "The following labels were found only in the train data: "
+                "The following labels were found only in the train data:",
-                "{}".format(", ".join(set(labels_train) - set(labels_dev))),
+                ", ".join(set(labels_train) - set(labels_dev)),
                 show=verbose,
             )
 
         # labels only in dev
         if set(labels_dev) - set(labels_train):
             msg.warn(
-                "The following labels were found only in the dev data: "
+                "The following labels were found only in the dev data:",
-                + ", ".join(set(labels_dev) - set(labels_train)),
+                ", ".join(set(labels_dev) - set(labels_train)),
                 show=verbose,
             )
 
         if has_low_data_warning:
             msg.text(
-                "To train a parser, your data should include at "
+                f"To train a parser, your data should include at "
-                "least {} instances of each label.".format(DEP_LABEL_THRESHOLD),
+                f"least {DEP_LABEL_THRESHOLD} instances of each label.",
                 show=verbose,
             )
 
         # multiple root labels
         if len(gold_train_unpreprocessed_data["roots"]) > 1:
             msg.warn(
-                "Multiple root labels ({}) ".format(
+                f"Multiple root labels "
-                    ", ".join(gold_train_unpreprocessed_data["roots"])
+                f"({', '.join(gold_train_unpreprocessed_data['roots'])}) "
-                )
+                f"found in training data. spaCy's parser uses a single root "
-                + "found in training data. spaCy's parser uses a single root "
+                f"label ROOT so this distinction will not be available."
-                "label ROOT so this distinction will not be available."
             )
 
         # these should not happen, but just in case
         if gold_train_data["n_nonproj"] > 0:
             msg.fail(
-                "Found {} nonprojective projectivized train sentence{}".format(
+                f"Found {gold_train_data['n_nonproj']} nonprojective "
-                    gold_train_data["n_nonproj"],
+                f"projectivized train sentence(s)"
-                    "s" if gold_train_data["n_nonproj"] > 1 else "",
-                )
             )
         if gold_train_data["n_cycles"] > 0:
             msg.fail(
-                "Found {} projectivized train sentence{} with cycles".format(
+                f"Found {gold_train_data['n_cycles']} projectivized train sentence(s) with cycles"
-                    gold_train_data["n_cycles"],
-                    "s" if gold_train_data["n_cycles"] > 1 else "",
-                )
             )
 
     msg.divider("Summary")
@@ -518,42 +423,34 @@ def debug_data(
     warn_counts = msg.counts[MESSAGES.WARN]
     fail_counts = msg.counts[MESSAGES.FAIL]
     if good_counts:
-        msg.good(
+        msg.good(f"{good_counts} {'check' if good_counts == 1 else 'checks'} passed")
-            "{} {} passed".format(
-                good_counts, "check" if good_counts == 1 else "checks"
-            )
-        )
     if warn_counts:
-        msg.warn(
+        msg.warn(f"{warn_counts} {'warning' if warn_counts == 1 else 'warnings'}")
-            "{} {}".format(warn_counts, "warning" if warn_counts == 1 else "warnings")
-        )
-    if fail_counts:
-        msg.fail("{} {}".format(fail_counts, "error" if fail_counts == 1 else "errors"))
-
     if fail_counts:
+        msg.fail(f"{fail_counts} {'error' if fail_counts == 1 else 'errors'}")
         sys.exit(1)
 
 
 def _load_file(file_path, msg):
     file_name = file_path.parts[-1]
     if file_path.suffix == ".json":
-        with msg.loading("Loading {}...".format(file_name)):
+        with msg.loading(f"Loading {file_name}..."):
             data = srsly.read_json(file_path)
-        msg.good("Loaded {}".format(file_name))
+        msg.good(f"Loaded {file_name}")
         return data
     elif file_path.suffix == ".jsonl":
-        with msg.loading("Loading {}...".format(file_name)):
+        with msg.loading(f"Loading {file_name}..."):
             data = srsly.read_jsonl(file_path)
-        msg.good("Loaded {}".format(file_name))
+        msg.good(f"Loaded {file_name}")
         return data
     msg.fail(
-        "Can't load file extension {}".format(file_path.suffix),
+        f"Can't load file extension {file_path.suffix}",
         "Expected .json or .jsonl",
         exits=1,
     )
 
 
-def _compile_gold(train_docs, pipeline):
+def _compile_gold(examples, pipeline):
     data = {
         "ner": Counter(),
         "cats": Counter(),
@@ -571,7 +468,9 @@ def _compile_gold(train_docs, pipeline):
         "n_cats_multilabel": 0,
         "texts": set(),
     }
-    for doc, gold in train_docs:
+    for example in examples:
+        gold = example.gold
+        doc = example.doc
         valid_words = [x for x in gold.words if x is not None]
         data["words"].update(valid_words)
         data["n_words"] += len(valid_words)
@@ -584,7 +483,13 @@ def _compile_gold(train_docs, pipeline):
                 if label.startswith(("B-", "U-", "L-")) and doc[i].is_space:
                     # "Illegal" whitespace entity
                     data["ws_ents"] += 1
-                if label.startswith(("B-", "U-", "L-")) and doc[i].text in [".", "'", "!", "?", ","]:
+                if label.startswith(("B-", "U-", "L-")) and doc[i].text in [
+                    ".",
+                    "'",
+                    "!",
+                    "?",
+                    ",",
+                ]:
                     # punctuation entity: could be replaced by whitespace when training with noise,
                     # so add a warning to alert the user to this unexpected side effect.
                     data["punct_ents"] += 1
@@ -614,14 +519,18 @@ def _compile_gold(train_docs, pipeline):
 
 def _format_labels(labels, counts=False):
     if counts:
-        return ", ".join(["'{}' ({})".format(l, c) for l, c in labels])
+        return ", ".join([f"'{l}' ({c})" for l, c in labels])
-    return ", ".join(["'{}'".format(l) for l in labels])
+    return ", ".join([f"'{l}'" for l in labels])
 
 
 def _get_examples_without_label(data, label):
     count = 0
-    for doc, gold in data:
+    for ex in data:
-        labels = [label.split("-")[1] for label in gold.ner if label not in ("O", "-")]
+        labels = [
+            label.split("-")[1]
+            for label in ex.gold.ner
+            if label not in ("O", "-", None)
+        ]
         if label not in labels:
             count += 1
     return count

spacy/cli/download.py

@@ -1,28 +1,21 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-import plac
 import requests
 import os
 import subprocess
 import sys
 from wasabi import msg
 
-from .link import link
-from ..util import get_package_path
 from .. import about
 
 
-@plac.annotations(
+def download(
-    model=("Model to download (shortcut or name)", "positional", None, str),
+    model: ("Model to download (shortcut or name)", "positional", None, str),
-    direct=("Force direct download of name + version", "flag", "d", bool),
+    direct: ("Force direct download of name + version", "flag", "d", bool) = False,
-    pip_args=("Additional arguments to be passed to `pip install` on model install"),
+    *pip_args: ("Additional arguments to be passed to `pip install` on model install"),
-)
+):
-def download(model, direct=False, *pip_args):
     """
-    Download compatible model from default download path using pip. Model
+    Download compatible model from default download path using pip. If --direct
-    can be shortcut, model name or, if --direct flag is set, full model name
+    flag is set, the command expects the full model name with version.
-    with version. For direct downloads, the compatibility check will be skipped.
+    For direct downloads, the compatibility check will be skipped.
     """
     if not require_package("spacy") and "--no-deps" not in pip_args:
         msg.warn(
@@ -50,29 +43,7 @@ def download(model, direct=False, *pip_args):
             sys.exit(dl)
         msg.good(
             "Download and installation successful",
-            "You can now load the model via spacy.load('{}')".format(model_name),
+            f"You can now load the model via spacy.load('{model_name}')",
-        )
-        # Only create symlink if the model is installed via a shortcut like 'en'.
-        # There's no real advantage over an additional symlink for en_core_web_sm
-        # and if anything, it's more error prone and causes more confusion.
-        if model in shortcuts:
-            try:
-                # Get package path here because link uses
-                # pip.get_installed_distributions() to check if model is a
-                # package, which fails if model was just installed via
-                # subprocess
-                package_path = get_package_path(model_name)
-                link(model_name, model, force=True, model_path=package_path)
-            except:  # noqa: E722
-                # Dirty, but since spacy.download and the auto-linking is
-                # mostly a convenience wrapper, it's best to show a success
-                # message and loading instructions, even if linking fails.
-                msg.warn(
-                    "Download successful but linking failed",
-                    "Creating a shortcut link for '{}' didn't work (maybe you "
-                    "don't have admin permissions?), but you can still load "
-                    "the model via its full package name: "
-                    "nlp = spacy.load('{}')".format(model, model_name),
         )
         # If a model is downloaded and then loaded within the same process, our
         # is_package check currently fails, because pkg_resources.working_set
@@ -95,11 +66,11 @@ def get_json(url, desc):
     r = requests.get(url)
     if r.status_code != 200:
         msg.fail(
-            "Server error ({})".format(r.status_code),
+            f"Server error ({r.status_code})",
-            "Couldn't fetch {}. Please find a model for your spaCy "
+            f"Couldn't fetch {desc}. Please find a model for your spaCy "
-            "installation (v{}), and download it manually. For more "
+            f"installation (v{about.__version__}), and download it manually. "
-            "details, see the documentation: "
+            f"For more details, see the documentation: "
-            "https://spacy.io/usage/models".format(desc, about.__version__),
+            f"https://spacy.io/usage/models",
             exits=1,
         )
     return r.json()
@@ -111,7 +82,7 @@ def get_compatibility():
     comp_table = get_json(about.__compatibility__, "compatibility table")
     comp = comp_table["spacy"]
     if version not in comp:
-        msg.fail("No compatible models found for v{} of spaCy".format(version), exits=1)
+        msg.fail(f"No compatible models found for v{version} of spaCy", exits=1)
     return comp[version]
 
 
@@ -119,8 +90,7 @@ def get_version(model, comp):
     model = model.rsplit(".dev", 1)[0]
     if model not in comp:
         msg.fail(
-            "No compatible model found for '{}' "
+            f"No compatible model found for '{model}' (spaCy v{about.__version__})",
-            "(spaCy v{}).".format(model, about.__version__),
             exits=1,
         )
     return comp[model][0]

spacy/cli/evaluate.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals, division, print_function
-
-import plac
 from timeit import default_timer as timer
 from wasabi import msg
 
@@ -10,23 +6,16 @@ from .. import util
 from .. import displacy
 
 
-@plac.annotations(
-    model=("Model name or path", "positional", None, str),
-    data_path=("Location of JSON-formatted evaluation data", "positional", None, str),
-    gold_preproc=("Use gold preprocessing", "flag", "G", bool),
-    gpu_id=("Use GPU", "option", "g", int),
-    displacy_path=("Directory to output rendered parses as HTML", "option", "dp", str),
-    displacy_limit=("Limit of parses to render as HTML", "option", "dl", int),
-    return_scores=("Return dict containing model scores", "flag", "R", bool),
-)
 def evaluate(
-    model,
+    # fmt: off
-    data_path,
+    model: ("Model name or path", "positional", None, str),
-    gpu_id=-1,
+    data_path: ("Location of JSON-formatted evaluation data", "positional", None, str),
-    gold_preproc=False,
+    gpu_id: ("Use GPU", "option", "g", int) = -1,
-    displacy_path=None,
+    gold_preproc: ("Use gold preprocessing", "flag", "G", bool) = False,
-    displacy_limit=25,
+    displacy_path: ("Directory to output rendered parses as HTML", "option", "dp", str) = None,
-    return_scores=False,
+    displacy_limit: ("Limit of parses to render as HTML", "option", "dl", int) = 25,
+    return_scores: ("Return dict containing model scores", "flag", "R", bool) = False,
+    # fmt: on
 ):
     """
     Evaluate a model. To render a sample of parses in a HTML file, set an
@@ -44,28 +33,31 @@ def evaluate(
         msg.fail("Visualization output directory not found", displacy_path, exits=1)
     corpus = GoldCorpus(data_path, data_path)
     nlp = util.load_model(model)
-    dev_docs = list(corpus.dev_docs(nlp, gold_preproc=gold_preproc))
+    dev_dataset = list(corpus.dev_dataset(nlp, gold_preproc=gold_preproc))
     begin = timer()
-    scorer = nlp.evaluate(dev_docs, verbose=False)
+    scorer = nlp.evaluate(dev_dataset, verbose=False)
     end = timer()
-    nwords = sum(len(doc_gold[0]) for doc_gold in dev_docs)
+    nwords = sum(len(ex.doc) for ex in dev_dataset)
     results = {
-        "Time": "%.2f s" % (end - begin),
+        "Time": f"{end - begin:.2f} s",
         "Words": nwords,
-        "Words/s": "%.0f" % (nwords / (end - begin)),
+        "Words/s": f"{nwords / (end - begin):.0f}",
-        "TOK": "%.2f" % scorer.token_acc,
+        "TOK": f"{scorer.token_acc:.2f}",
-        "POS": "%.2f" % scorer.tags_acc,
+        "POS": f"{scorer.tags_acc:.2f}",
-        "UAS": "%.2f" % scorer.uas,
+        "UAS": f"{scorer.uas:.2f}",
-        "LAS": "%.2f" % scorer.las,
+        "LAS": f"{scorer.las:.2f}",
-        "NER P": "%.2f" % scorer.ents_p,
+        "NER P": f"{scorer.ents_p:.2f}",
-        "NER R": "%.2f" % scorer.ents_r,
+        "NER R": f"{scorer.ents_r:.2f}",
-        "NER F": "%.2f" % scorer.ents_f,
+        "NER F": f"{scorer.ents_f:.2f}",
-        "Textcat": "%.2f" % scorer.textcat_score,
+        "Textcat": f"{scorer.textcat_score:.2f}",
+        "Sent P": f"{scorer.sent_p:.2f}",
+        "Sent R": f"{scorer.sent_r:.2f}",
+        "Sent F": f"{scorer.sent_f:.2f}",
     }
     msg.table(results, title="Results")
 
     if displacy_path:
-        docs, golds = zip(*dev_docs)
+        docs = [ex.doc for ex in dev_dataset]
         render_deps = "parser" in nlp.meta.get("pipeline", [])
         render_ents = "ner" in nlp.meta.get("pipeline", [])
         render_parses(
@@ -76,7 +68,7 @@ def evaluate(
             deps=render_deps,
             ents=render_ents,
         )
-        msg.good("Generated {} parses as HTML".format(displacy_limit), displacy_path)
+        msg.good(f"Generated {displacy_limit} parses as HTML", displacy_path)
     if return_scores:
         return scorer.scores
 

spacy/cli/info.py

@@ -1,44 +1,39 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-import plac
 import platform
 from pathlib import Path
 from wasabi import msg
 import srsly
 
-from ..compat import path2str, basestring_, unicode_
+from .validate import get_model_pkgs
 from .. import util
 from .. import about
 
 
-@plac.annotations(
+def info(
-    model=("Optional shortcut link of model", "positional", None, str),
+    model: ("Optional model name", "positional", None, str) = None,
-    markdown=("Generate Markdown for GitHub issues", "flag", "md", str),
+    markdown: ("Generate Markdown for GitHub issues", "flag", "md", str) = False,
-    silent=("Don't print anything (just return)", "flag", "s"),
+    silent: ("Don't print anything (just return)", "flag", "s") = False,
-)
+):
-def info(model=None, markdown=False, silent=False):
     """
-    Print info about spaCy installation. If a model shortcut link is
+    Print info about spaCy installation. If a model is speficied as an argument,
-    speficied as an argument, print model information. Flag --markdown
+    print model information. Flag --markdown prints details in Markdown for easy
-    prints details in Markdown for easy copy-pasting to GitHub issues.
+    copy-pasting to GitHub issues.
     """
     if model:
         if util.is_package(model):
             model_path = util.get_package_path(model)
         else:
-            model_path = util.get_data_path() / model
+            model_path = model
         meta_path = model_path / "meta.json"
         if not meta_path.is_file():
             msg.fail("Can't find model meta.json", meta_path, exits=1)
         meta = srsly.read_json(meta_path)
         if model_path.resolve() != model_path:
-            meta["link"] = path2str(model_path)
+            meta["link"] = str(model_path)
-            meta["source"] = path2str(model_path.resolve())
+            meta["source"] = str(model_path.resolve())
         else:
-            meta["source"] = path2str(model_path)
+            meta["source"] = str(model_path)
         if not silent:
-            title = "Info about model '{}'".format(model)
+            title = f"Info about model '{model}'"
             model_meta = {
                 k: v for k, v in meta.items() if k not in ("accuracy", "speed")
             }
@@ -47,12 +42,13 @@ def info(model=None, markdown=False, silent=False):
             else:
                 msg.table(model_meta, title=title)
         return meta
+    all_models, _ = get_model_pkgs()
     data = {
         "spaCy version": about.__version__,
-        "Location": path2str(Path(__file__).parent.parent),
+        "Location": str(Path(__file__).parent.parent),
         "Platform": platform.platform(),
         "Python version": platform.python_version(),
-        "Models": list_models(),
+        "Models": ", ".join(model["name"] for model in all_models.values()),
     }
     if not silent:
         title = "Info about spaCy"
@@ -63,19 +59,6 @@ def info(model=None, markdown=False, silent=False):
     return data
 
 
-def list_models():
-    def exclude_dir(dir_name):
-        # exclude common cache directories and hidden directories
-        exclude = ("cache", "pycache", "__pycache__")
-        return dir_name in exclude or dir_name.startswith(".")
-
-    data_path = util.get_data_path()
-    if data_path:
-        models = [f.parts[-1] for f in data_path.iterdir() if f.is_dir()]
-        return ", ".join([m for m in models if not exclude_dir(m)])
-    return "-"
-
-
 def print_markdown(data, title=None):
     """Print data in GitHub-flavoured Markdown format for issues etc.
 
@@ -84,9 +67,9 @@ def print_markdown(data, title=None):
     """
     markdown = []
     for key, value in data.items():
-        if isinstance(value, basestring_) and Path(value).exists():
+        if isinstance(value, str) and Path(value).exists():
             continue
-        markdown.append("* **{}:** {}".format(key, unicode_(value)))
+        markdown.append(f"* **{key}:** {value}")
     if title:
-        print("\n## {}".format(title))
+        print(f"\n## {title}")
     print("\n{}\n".format("\n".join(markdown)))

spacy/cli/init_model.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-import plac
 import math
 from tqdm import tqdm
 import numpy
@@ -27,32 +23,18 @@ except ImportError:
 DEFAULT_OOV_PROB = -20
 
 
-@plac.annotations(
-    lang=("Model language", "positional", None, str),
-    output_dir=("Model output directory", "positional", None, Path),
-    freqs_loc=("Location of words frequencies file", "option", "f", Path),
-    jsonl_loc=("Location of JSONL-formatted attributes file", "option", "j", Path),
-    clusters_loc=("Optional location of brown clusters data", "option", "c", str),
-    vectors_loc=("Optional vectors file in Word2Vec format", "option", "v", str),
-    prune_vectors=("Optional number of vectors to prune to", "option", "V", int),
-    vectors_name=(
-        "Optional name for the word vectors, e.g. en_core_web_lg.vectors",
-        "option",
-        "vn",
-        str,
-    ),
-    model_name=("Optional name for the model meta", "option", "mn", str),
-)
 def init_model(
-    lang,
+    # fmt: off
-    output_dir,
+    lang: ("Model language", "positional", None, str),
-    freqs_loc=None,
+    output_dir: ("Model output directory", "positional", None, Path),
-    clusters_loc=None,
+    freqs_loc: ("Location of words frequencies file", "option", "f", Path) = None,
-    jsonl_loc=None,
+    clusters_loc: ("Optional location of brown clusters data", "option", "c", str) = None,
-    vectors_loc=None,
+    jsonl_loc: ("Location of JSONL-formatted attributes file", "option", "j", Path) = None,
-    prune_vectors=-1,
+    vectors_loc: ("Optional vectors file in Word2Vec format", "option", "v", str) = None,
-    vectors_name=None,
+    prune_vectors: ("Optional number of vectors to prune to", "option", "V", int) = -1,
-    model_name=None,
+    vectors_name: ("Optional name for the word vectors, e.g. en_core_web_lg.vectors", "option", "vn", str) = None,
+    model_name: ("Optional name for the model meta", "option", "mn", str) = None,
+    # fmt: on
 ):
     """
     Create a new model from raw data, like word frequencies, Brown clusters
@@ -91,8 +73,7 @@ def init_model(
     vec_added = len(nlp.vocab.vectors)
     lex_added = len(nlp.vocab)
     msg.good(
-        "Sucessfully compiled vocab",
+        "Sucessfully compiled vocab", f"{lex_added} entries, {vec_added} vectors",
-        "{} entries, {} vectors".format(lex_added, vec_added),
     )
     if not output_dir.exists():
         output_dir.mkdir()
@@ -177,9 +158,9 @@ def add_vectors(nlp, vectors_loc, prune_vectors, name=None):
                 nlp.vocab.vectors.add(lex.orth, row=lex.rank)
     else:
         if vectors_loc:
-            with msg.loading("Reading vectors from {}".format(vectors_loc)):
+            with msg.loading(f"Reading vectors from {vectors_loc}"):
                 vectors_data, vector_keys = read_vectors(vectors_loc)
-            msg.good("Loaded vectors from {}".format(vectors_loc))
+            msg.good(f"Loaded vectors from {vectors_loc}")
         else:
             vectors_data, vector_keys = (None, None)
         if vector_keys is not None:
@@ -190,7 +171,7 @@ def add_vectors(nlp, vectors_loc, prune_vectors, name=None):
         if vectors_data is not None:
             nlp.vocab.vectors = Vectors(data=vectors_data, keys=vector_keys)
     if name is None:
-        nlp.vocab.vectors.name = "%s_model.vectors" % nlp.meta["lang"]
+        nlp.vocab.vectors.name = f"{nlp.meta['lang']}_model.vectors"
     else:
         nlp.vocab.vectors.name = name
     nlp.meta["vectors"]["name"] = nlp.vocab.vectors.name
@@ -236,7 +217,7 @@ def read_freqs(freqs_loc, max_length=100, min_doc_freq=5, min_freq=50):
                     word = literal_eval(key)
                 except SyntaxError:
                     # Take odd strings literally.
-                    word = literal_eval("'%s'" % key)
+                    word = literal_eval(f"'{key}'")
                 smooth_count = counts.smoother(int(freq))
                 probs[word] = math.log(smooth_count) - log_total
     oov_prob = math.log(counts.smoother(0)) - log_total

spacy/cli/link.py

@@ -1,77 +0,0 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-import plac
-from pathlib import Path
-from wasabi import msg
-
-from ..compat import symlink_to, path2str
-from .. import util
-
-
-@plac.annotations(
-    origin=("package name or local path to model", "positional", None, str),
-    link_name=("name of shortuct link to create", "positional", None, str),
-    force=("force overwriting of existing link", "flag", "f", bool),
-)
-def link(origin, link_name, force=False, model_path=None):
-    """
-    Create a symlink for models within the spacy/data directory. Accepts
-    either the name of a pip package, or the local path to the model data
-    directory. Linking models allows loading them via spacy.load(link_name).
-    """
-    if util.is_package(origin):
-        model_path = util.get_package_path(origin)
-    else:
-        model_path = Path(origin) if model_path is None else Path(model_path)
-    if not model_path.exists():
-        msg.fail(
-            "Can't locate model data",
-            "The data should be located in {}".format(path2str(model_path)),
-            exits=1,
-        )
-    data_path = util.get_data_path()
-    if not data_path or not data_path.exists():
-        spacy_loc = Path(__file__).parent.parent
-        msg.fail(
-            "Can't find the spaCy data path to create model symlink",
-            "Make sure a directory `/data` exists within your spaCy "
-            "installation and try again. The data directory should be located "
-            "here:".format(path=spacy_loc),
-            exits=1,
-        )
-    link_path = util.get_data_path() / link_name
-    if link_path.is_symlink() and not force:
-        msg.fail(
-            "Link '{}' already exists".format(link_name),
-            "To overwrite an existing link, use the --force flag",
-            exits=1,
-        )
-    elif link_path.is_symlink():  # does a symlink exist?
-        # NB: It's important to check for is_symlink here and not for exists,
-        # because invalid/outdated symlinks would return False otherwise.
-        link_path.unlink()
-    elif link_path.exists():  # does it exist otherwise?
-        # NB: Check this last because valid symlinks also "exist".
-        msg.fail(
-            "Can't overwrite symlink '{}'".format(link_name),
-            "This can happen if your data directory contains a directory or "
-            "file of the same name.",
-            exits=1,
-        )
-    details = "%s --> %s" % (path2str(model_path), path2str(link_path))
-    try:
-        symlink_to(link_path, model_path)
-    except:  # noqa: E722
-        # This is quite dirty, but just making sure other errors are caught.
-        msg.fail(
-            "Couldn't link model to '{}'".format(link_name),
-            "Creating a symlink in spacy/data failed. Make sure you have the "
-            "required permissions and try re-running the command as admin, or "
-            "use a virtualenv. You can still import the model as a module and "
-            "call its load() method, or create the symlink manually.",
-        )
-        msg.text(details)
-        raise
-    msg.good("Linking successful", details)
-    msg.text("You can now load the model via spacy.load('{}')".format(link_name))

spacy/cli/package.py

@@ -1,25 +1,21 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-import plac
 import shutil
 from pathlib import Path
 from wasabi import msg, get_raw_input
 import srsly
 
-from ..compat import path2str
 from .. import util
 from .. import about
 
 
-@plac.annotations(
+def package(
-    input_dir=("Directory with model data", "positional", None, str),
+    # fmt: off
-    output_dir=("Output parent directory", "positional", None, str),
+    input_dir: ("Directory with model data", "positional", None, str),
-    meta_path=("Path to meta.json", "option", "m", str),
+    output_dir: ("Output parent directory", "positional", None, str),
-    create_meta=("Create meta.json, even if one exists", "flag", "c", bool),
+    meta_path: ("Path to meta.json", "option", "m", str) = None,
-    force=("Force overwriting existing model in output directory", "flag", "f", bool),
+    create_meta: ("Create meta.json, even if one exists", "flag", "c", bool) = False,
-)
+    force: ("Force overwriting existing model in output directory", "flag", "f", bool) = False,
-def package(input_dir, output_dir, meta_path=None, create_meta=False, force=False):
+    # fmt: on
+):
     """
     Generate Python package for model data, including meta and required
     installation files. A new directory will be created in the specified
@@ -47,7 +43,7 @@ def package(input_dir, output_dir, meta_path=None, create_meta=False, force=Fals
     for key in ("lang", "name", "version"):
         if key not in meta or meta[key] == "":
             msg.fail(
-                "No '{}' setting found in meta.json".format(key),
+                f"No '{key}' setting found in meta.json",
                 "This setting is required to build your package.",
                 exits=1,
             )
@@ -58,22 +54,21 @@ def package(input_dir, output_dir, meta_path=None, create_meta=False, force=Fals
 
     if package_path.exists():
         if force:
-            shutil.rmtree(path2str(package_path))
+            shutil.rmtree(str(package_path))
         else:
             msg.fail(
                 "Package directory already exists",
                 "Please delete the directory and try again, or use the "
-                "`--force` flag to overwrite existing "
+                "`--force` flag to overwrite existing directories.",
-                "directories.".format(path=path2str(package_path)),
                 exits=1,
             )
     Path.mkdir(package_path, parents=True)
-    shutil.copytree(path2str(input_path), path2str(package_path / model_name_v))
+    shutil.copytree(str(input_path), str(package_path / model_name_v))
     create_file(main_path / "meta.json", srsly.json_dumps(meta, indent=2))
     create_file(main_path / "setup.py", TEMPLATE_SETUP)
     create_file(main_path / "MANIFEST.in", TEMPLATE_MANIFEST)
     create_file(package_path / "__init__.py", TEMPLATE_INIT)
-    msg.good("Successfully created package '{}'".format(model_name_v), main_path)
+    msg.good(f"Successfully created package '{model_name_v}'", main_path)
     msg.text("To build the package, run `python setup.py sdist` in this directory.")
 
 
@@ -88,7 +83,7 @@ def generate_meta(model_path, existing_meta, msg):
         ("lang", "Model language", meta.get("lang", "en")),
         ("name", "Model name", meta.get("name", "model")),
         ("version", "Model version", meta.get("version", "0.0.0")),
-        ("spacy_version", "Required spaCy version", ">=%s,<3.0.0" % about.__version__),
+        ("spacy_version", "Required spaCy version", f">={about.__version__},<3.0.0"),
         ("description", "Model description", meta.get("description", False)),
         ("author", "Author", meta.get("author", False)),
         ("email", "Author email", meta.get("email", False)),
@@ -118,9 +113,6 @@ def generate_meta(model_path, existing_meta, msg):
 
 TEMPLATE_SETUP = """
 #!/usr/bin/env python
-# coding: utf8
-from __future__ import unicode_literals
-
 import io
 import json
 from os import path, walk
@@ -190,9 +182,6 @@ include meta.json
 
 
 TEMPLATE_INIT = """
-# coding: utf8
-from __future__ import unicode_literals
-
 from pathlib import Path
 from spacy.util import load_model_from_init_py, get_model_meta
 

spacy/cli/pretrain.py

@@ -1,107 +1,50 @@
-# coding: utf8
-from __future__ import print_function, unicode_literals
-
-import plac
 import random
 import numpy
 import time
 import re
 from collections import Counter
 from pathlib import Path
-from thinc.v2v import Affine, Maxout
+from thinc.api import Linear, Maxout, chain, list2array, prefer_gpu
-from thinc.misc import LayerNorm as LN
+from thinc.api import CosineDistance, L2Distance
-from thinc.neural.util import prefer_gpu
 from wasabi import msg
 import srsly
 
+from ..gold import Example
 from ..errors import Errors
 from ..tokens import Doc
 from ..attrs import ID, HEAD
-from .._ml import Tok2Vec, flatten, chain, create_default_optimizer
+from ..ml.component_models import Tok2Vec
-from .._ml import masked_language_model, get_cossim_loss
+from ..ml.component_models import masked_language_model
 from .. import util
+from ..util import create_default_optimizer
 from .train import _load_pretrained_tok2vec
 
 
-@plac.annotations(
-    texts_loc=(
-        "Path to JSONL file with raw texts to learn from, with text provided as the key 'text' or tokens as the "
-        "key 'tokens'",
-        "positional",
-        None,
-        str,
-    ),
-    vectors_model=("Name or path to spaCy model with vectors to learn from"),
-    output_dir=("Directory to write models to on each epoch", "positional", None, str),
-    width=("Width of CNN layers", "option", "cw", int),
-    depth=("Depth of CNN layers", "option", "cd", int),
-    cnn_window=("Window size for CNN layers", "option", "cW", int),
-    cnn_pieces=("Maxout size for CNN layers. 1 for Mish", "option", "cP", int),
-    use_chars=("Whether to use character-based embedding", "flag", "chr", bool),
-    sa_depth=("Depth of self-attention layers", "option", "sa", int),
-    bilstm_depth=("Depth of BiLSTM layers (requires PyTorch)", "option", "lstm", int),
-    embed_rows=("Number of embedding rows", "option", "er", int),
-    loss_func=(
-        "Loss function to use for the objective. Either 'L2' or 'cosine'",
-        "option",
-        "L",
-        str,
-    ),
-    use_vectors=("Whether to use the static vectors as input features", "flag", "uv"),
-    dropout=("Dropout rate", "option", "d", float),
-    batch_size=("Number of words per training batch", "option", "bs", int),
-    max_length=(
-        "Max words per example. Longer examples are discarded",
-        "option",
-        "xw",
-        int,
-    ),
-    min_length=(
-        "Min words per example. Shorter examples are discarded",
-        "option",
-        "nw",
-        int,
-    ),
-    seed=("Seed for random number generators", "option", "s", int),
-    n_iter=("Number of iterations to pretrain", "option", "i", int),
-    n_save_every=("Save model every X batches.", "option", "se", int),
-    init_tok2vec=(
-        "Path to pretrained weights for the token-to-vector parts of the models. See 'spacy pretrain'. Experimental.",
-        "option",
-        "t2v",
-        Path,
-    ),
-    epoch_start=(
-        "The epoch to start counting at. Only relevant when using '--init-tok2vec' and the given weight file has been "
-        "renamed. Prevents unintended overwriting of existing weight files.",
-        "option",
-        "es",
-        int,
-    ),
-)
 def pretrain(
-    texts_loc,
+    # fmt: off
-    vectors_model,
+    texts_loc: ("Path to JSONL file with raw texts to learn from, with text provided as the key 'text' or tokens as the key 'tokens'", "positional", None, str),
-    output_dir,
+    vectors_model: ("Name or path to spaCy model with vectors to learn from", "positional", None, str),
-    width=96,
+    output_dir: ("Directory to write models to on each epoch", "positional", None, str),
-    depth=4,
+    width: ("Width of CNN layers", "option", "cw", int) = 96,
-    bilstm_depth=0,
+    conv_depth: ("Depth of CNN layers", "option", "cd", int) = 4,
-    cnn_pieces=3,
+    bilstm_depth: ("Depth of BiLSTM layers (requires PyTorch)", "option", "lstm", int) = 0,
-    sa_depth=0,
+    cnn_pieces: ("Maxout size for CNN layers. 1 for Mish", "option", "cP", int) = 3,
-    use_chars=False,
+    sa_depth: ("Depth of self-attention layers", "option", "sa", int) = 0,
-    cnn_window=1,
+    use_chars: ("Whether to use character-based embedding", "flag", "chr", bool) = False,
-    embed_rows=2000,
+    cnn_window: ("Window size for CNN layers", "option", "cW", int) = 1,
-    loss_func="cosine",
+    embed_rows: ("Number of embedding rows", "option", "er", int) = 2000,
-    use_vectors=False,
+    loss_func: ("Loss function to use for the objective. Either 'L2' or 'cosine'", "option", "L", str) = "cosine",
-    dropout=0.2,
+    use_vectors: ("Whether to use the static vectors as input features", "flag", "uv") = False,
-    n_iter=1000,
+    dropout: ("Dropout rate", "option", "d", float) = 0.2,
-    batch_size=3000,
+    n_iter: ("Number of iterations to pretrain", "option", "i", int) = 1000,
-    max_length=500,
+    batch_size: ("Number of words per training batch", "option", "bs", int) = 3000,
-    min_length=5,
+    max_length: ("Max words per example. Longer examples are discarded", "option", "xw", int) = 500,
-    seed=0,
+    min_length: ("Min words per example. Shorter examples are discarded", "option", "nw", int) = 5,
-    n_save_every=None,
+    seed: ("Seed for random number generators", "option", "s", int) = 0,
-    init_tok2vec=None,
+    n_save_every: ("Save model every X batches.", "option", "se", int) = None,
-    epoch_start=None,
+    init_tok2vec: ("Path to pretrained weights for the token-to-vector parts of the models. See 'spacy pretrain'. Experimental.", "option", "t2v", Path) = None,
+    epoch_start: ("The epoch to start counting at. Only relevant when using '--init-tok2vec' and the given weight file has been renamed. Prevents unintended overwriting of existing weight files.", "option", "es", int) = None,
+    # fmt: on
 ):
     """
     Pre-train the 'token-to-vector' (tok2vec) layer of pipeline components,
@@ -132,9 +75,15 @@ def pretrain(
     msg.info("Using GPU" if has_gpu else "Not using GPU")
 
     output_dir = Path(output_dir)
+    if output_dir.exists() and [p for p in output_dir.iterdir()]:
+        msg.warn(
+            "Output directory is not empty",
+            "It is better to use an empty directory or refer to a new output path, "
+            "then the new directory will be created for you.",
+        )
     if not output_dir.exists():
         output_dir.mkdir()
-        msg.good("Created output directory")
+        msg.good(f"Created output directory: {output_dir}")
     srsly.write_json(output_dir / "config.json", config)
     msg.good("Saved settings to config.json")
 
@@ -153,16 +102,16 @@ def pretrain(
         msg.text("Reading input text from stdin...")
         texts = srsly.read_jsonl("-")
 
-    with msg.loading("Loading model '{}'...".format(vectors_model)):
+    with msg.loading(f"Loading model '{vectors_model}'..."):
         nlp = util.load_model(vectors_model)
-    msg.good("Loaded model '{}'".format(vectors_model))
+    msg.good(f"Loaded model '{vectors_model}'")
-    pretrained_vectors = None if not use_vectors else nlp.vocab.vectors.name
+    pretrained_vectors = None if not use_vectors else nlp.vocab.vectors
     model = create_pretraining_model(
         nlp,
         Tok2Vec(
             width,
             embed_rows,
-            conv_depth=depth,
+            conv_depth=conv_depth,
             pretrained_vectors=pretrained_vectors,
             bilstm_depth=bilstm_depth,  # Requires PyTorch. Experimental.
             subword_features=not use_chars,  # Set to False for Chinese etc
@@ -172,7 +121,7 @@ def pretrain(
     # Load in pretrained weights
     if init_tok2vec is not None:
         components = _load_pretrained_tok2vec(nlp, init_tok2vec)
-        msg.text("Loaded pretrained tok2vec for: {}".format(components))
+        msg.text(f"Loaded pretrained tok2vec for: {components}")
         # Parse the epoch number from the given weight file
         model_name = re.search(r"model\d+\.bin", str(init_tok2vec))
         if model_name:
@@ -181,32 +130,28 @@ def pretrain(
         else:
             if not epoch_start:
                 msg.fail(
-                    "You have to use the '--epoch-start' argument when using a renamed weight file for "
+                    "You have to use the --epoch-start argument when using a renamed weight file for --init-tok2vec",
-                    "'--init-tok2vec'",
                     exits=True,
                 )
             elif epoch_start < 0:
                 msg.fail(
-                    "The argument '--epoch-start' has to be greater or equal to 0. '%d' is invalid"
+                    f"The argument --epoch-start has to be greater or equal to 0. {epoch_start} is invalid",
-                    % epoch_start,
                     exits=True,
                 )
     else:
         # Without '--init-tok2vec' the '--epoch-start' argument is ignored
         epoch_start = 0
 
-    optimizer = create_default_optimizer(model.ops)
+    optimizer = create_default_optimizer()
     tracker = ProgressTracker(frequency=10000)
-    msg.divider("Pre-training tok2vec layer - starting at epoch %d" % epoch_start)
+    msg.divider(f"Pre-training tok2vec layer - starting at epoch {epoch_start}")
     row_settings = {"widths": (3, 10, 10, 6, 4), "aligns": ("r", "r", "r", "r", "r")}
     msg.row(("#", "# Words", "Total Loss", "Loss", "w/s"), **row_settings)
 
     def _save_model(epoch, is_temp=False):
         is_temp_str = ".temp" if is_temp else ""
         with model.use_params(optimizer.averages):
-            with (output_dir / ("model%d%s.bin" % (epoch, is_temp_str))).open(
+            with (output_dir / f"model{epoch}{is_temp_str}.bin").open("wb") as file_:
-                "wb"
-            ) as file_:
                 file_.write(model.tok2vec.to_bytes())
             log = {
                 "nr_word": tracker.nr_word,
@@ -220,7 +165,9 @@ def pretrain(
     skip_counter = 0
     for epoch in range(epoch_start, n_iter + epoch_start):
         for batch_id, batch in enumerate(
-            util.minibatch_by_words(((text, None) for text in texts), size=batch_size)
+            util.minibatch_by_words(
+                (Example(doc=text) for text in texts), size=batch_size
+            )
         ):
             docs, count = make_docs(
                 nlp,
@@ -245,7 +192,7 @@ def pretrain(
             # Reshuffle the texts if texts were loaded from a file
             random.shuffle(texts)
     if skip_counter > 0:
-        msg.warn("Skipped {count} empty values".format(count=str(skip_counter)))
+        msg.warn(f"Skipped {skip_counter} empty values")
     msg.good("Successfully finished pretrain")
 
 
@@ -310,13 +257,14 @@ def get_vectors_loss(ops, docs, prediction, objective="L2"):
     # and look them up all at once. This prevents data copying.
     ids = ops.flatten([doc.to_array(ID).ravel() for doc in docs])
     target = docs[0].vocab.vectors.data[ids]
+    # TODO: this code originally didn't normalize, but shouldn't normalize=True ?
     if objective == "L2":
-        d_target = prediction - target
+        distance = L2Distance(normalize=False)
-        loss = (d_target ** 2).sum()
     elif objective == "cosine":
-        loss, d_target = get_cossim_loss(prediction, target)
+        distance = CosineDistance(normalize=False)
     else:
         raise ValueError(Errors.E142.format(loss_func=objective))
+    d_target, loss = distance(prediction, target)
     return loss, d_target
 
 
@@ -328,18 +276,18 @@ def create_pretraining_model(nlp, tok2vec):
     """
     output_size = nlp.vocab.vectors.data.shape[1]
     output_layer = chain(
-        LN(Maxout(300, pieces=3)), Affine(output_size, drop_factor=0.0)
+        Maxout(300, pieces=3, normalize=True, dropout=0.0), Linear(output_size)
     )
     # This is annoying, but the parser etc have the flatten step after
     # the tok2vec. To load the weights in cleanly, we need to match
     # the shape of the models' components exactly. So what we cann
     # "tok2vec" has to be the same set of processes as what the components do.
-    tok2vec = chain(tok2vec, flatten)
+    tok2vec = chain(tok2vec, list2array())
     model = chain(tok2vec, output_layer)
     model = masked_language_model(nlp.vocab, model)
-    model.tok2vec = tok2vec
+    model.set_ref("tok2vec", tok2vec)
-    model.output_layer = output_layer
+    model.set_ref("output_layer", output_layer)
-    model.begin_training([nlp.make_doc("Give it a doc to infer shapes")])
+    model.initialize(X=[nlp.make_doc("Give it a doc to infer shapes")])
     return model
 
 

spacy/cli/profile.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals, division, print_function
-
-import plac
 import tqdm
 from pathlib import Path
 import srsly
@@ -9,18 +5,19 @@ import cProfile
 import pstats
 import sys
 import itertools
-import thinc.extra.datasets
+import ml_datasets
 from wasabi import msg
 
 from ..util import load_model
 
 
-@plac.annotations(
+def profile(
-    model=("Model to load", "positional", None, str),
+    # fmt: off
-    inputs=("Location of input file. '-' for stdin.", "positional", None, str),
+    model: ("Model to load", "positional", None, str),
-    n_texts=("Maximum number of texts to use if available", "option", "n", int),
+    inputs: ("Location of input file. '-' for stdin.", "positional", None, str) = None,
-)
+    n_texts: ("Maximum number of texts to use if available", "option", "n", int) = 10000,
-def profile(model, inputs=None, n_texts=10000):
+    # fmt: on
+):
     """
     Profile a spaCy pipeline, to find out which functions take the most time.
     Input should be formatted as one JSON object per line with a key "text".
@@ -32,13 +29,13 @@ def profile(model, inputs=None, n_texts=10000):
     if inputs is None:
         n_inputs = 25000
         with msg.loading("Loading IMDB dataset via Thinc..."):
-            imdb_train, _ = thinc.extra.datasets.imdb()
+            imdb_train, _ = ml_datasets.imdb()
             inputs, _ = zip(*imdb_train)
-        msg.info("Loaded IMDB dataset and using {} examples".format(n_inputs))
+        msg.info(f"Loaded IMDB dataset and using {n_inputs} examples")
         inputs = inputs[:n_inputs]
-    with msg.loading("Loading model '{}'...".format(model)):
+    with msg.loading(f"Loading model '{model}'..."):
         nlp = load_model(model)
-    msg.good("Loaded model '{}'".format(model))
+    msg.good(f"Loaded model '{model}'")
     texts = list(itertools.islice(inputs, n_texts))
     cProfile.runctx("parse_texts(nlp, texts)", globals(), locals(), "Profile.prof")
     s = pstats.Stats("Profile.prof")
@@ -60,7 +57,7 @@ def _read_inputs(loc, msg):
         input_path = Path(loc)
         if not input_path.exists() or not input_path.is_file():
             msg.fail("Not a valid input data file", loc, exits=1)
-        msg.info("Using data from {}".format(input_path.parts[-1]))
+        msg.info(f"Using data from {input_path.parts[-1]}")
         file_ = input_path.open()
     for line in file_:
         data = srsly.json_loads(line)

spacy/cli/train.py

@@ -1,11 +1,7 @@
-# coding: utf8
-from __future__ import unicode_literals, division, print_function
-
-import plac
 import os
 import tqdm
 from pathlib import Path
-from thinc.neural._classes.model import Model
+from thinc.api import use_ops
 from timeit import default_timer as timer
 import shutil
 import srsly
@@ -13,76 +9,53 @@ from wasabi import msg
 import contextlib
 import random
 
-from .._ml import create_default_optimizer
+from ..util import create_default_optimizer
 from ..util import use_gpu as set_gpu
 from ..attrs import PROB, IS_OOV, CLUSTER, LANG
 from ..gold import GoldCorpus
-from ..compat import path2str
 from .. import util
 from .. import about
 
 
-@plac.annotations(
-    # fmt: off
-    lang=("Model language", "positional", None, str),
-    output_path=("Output directory to store model in", "positional", None, Path),
-    train_path=("Location of JSON-formatted training data", "positional", None, Path),
-    dev_path=("Location of JSON-formatted development data", "positional", None, Path),
-    raw_text=("Path to jsonl file with unlabelled text documents.", "option", "rt", Path),
-    base_model=("Name of model to update (optional)", "option", "b", str),
-    pipeline=("Comma-separated names of pipeline components", "option", "p", str),
-    replace_components=("Replace components from base model", "flag", "R", bool),
-    vectors=("Model to load vectors from", "option", "v", str),
-    n_iter=("Number of iterations", "option", "n", int),
-    n_early_stopping=("Maximum number of training epochs without dev accuracy improvement", "option", "ne", int),
-    n_examples=("Number of examples", "option", "ns", int),
-    use_gpu=("Use GPU", "option", "g", int),
-    version=("Model version", "option", "V", str),
-    meta_path=("Optional path to meta.json to use as base.", "option", "m", Path),
-    init_tok2vec=("Path to pretrained weights for the token-to-vector parts of the models. See 'spacy pretrain'. Experimental.", "option", "t2v", Path),
-    parser_multitasks=("Side objectives for parser CNN, e.g. 'dep' or 'dep,tag'", "option", "pt", str),
-    entity_multitasks=("Side objectives for NER CNN, e.g. 'dep' or 'dep,tag'", "option", "et", str),
-    noise_level=("Amount of corruption for data augmentation", "option", "nl", float),
-    orth_variant_level=("Amount of orthography variation for data augmentation", "option", "ovl", float),
-    eval_beam_widths=("Beam widths to evaluate, e.g. 4,8", "option", "bw", str),
-    gold_preproc=("Use gold preprocessing", "flag", "G", bool),
-    learn_tokens=("Make parser learn gold-standard tokenization", "flag", "T", bool),
-    textcat_multilabel=("Textcat classes aren't mutually exclusive (multilabel)", "flag", "TML", bool),
-    textcat_arch=("Textcat model architecture", "option", "ta", str),
-    textcat_positive_label=("Textcat positive label for binary classes with two labels", "option", "tpl", str),
-    verbose=("Display more information for debug", "flag", "VV", bool),
-    debug=("Run data diagnostics before training", "flag", "D", bool),
-    # fmt: on
-)
 def train(
-    lang,
+    # fmt: off
-    output_path,
+    lang: ("Model language", "positional", None, str),
-    train_path,
+    output_path: ("Output directory to store model in", "positional", None, Path),
-    dev_path,
+    train_path: ("Location of JSON-formatted training data", "positional", None, Path),
-    raw_text=None,
+    dev_path: ("Location of JSON-formatted development data", "positional", None, Path),
-    base_model=None,
+    raw_text: ("Path to jsonl file with unlabelled text documents.", "option", "rt", Path) = None,
-    pipeline="tagger,parser,ner",
+    base_model: ("Name of model to update (optional)", "option", "b", str) = None,
-    replace_components=False,
+    pipeline: ("Comma-separated names of pipeline components", "option", "p", str) = "tagger,parser,ner",
-    vectors=None,
+    vectors: ("Model to load vectors from", "option", "v", str) = None,
-    n_iter=30,
+    replace_components: ("Replace components from base model", "flag", "R", bool) = False,
-    n_early_stopping=None,
+    width: ("Width of CNN layers of Tok2Vec component", "option", "cw", int) = 96,
-    n_examples=0,
+    conv_depth: ("Depth of CNN layers of Tok2Vec component", "option", "cd", int) = 4,
-    use_gpu=-1,
+    cnn_window: ("Window size for CNN layers of Tok2Vec component", "option", "cW", int) = 1,
-    version="0.0.0",
+    cnn_pieces: ("Maxout size for CNN layers of Tok2Vec component. 1 for Mish", "option", "cP", int) = 3,
-    meta_path=None,
+    use_chars: ("Whether to use character-based embedding of Tok2Vec component", "flag", "chr", bool) = False,
-    init_tok2vec=None,
+    bilstm_depth: ("Depth of BiLSTM layers of Tok2Vec component (requires PyTorch)", "option", "lstm", int) = 0,
-    parser_multitasks="",
+    embed_rows: ("Number of embedding rows of Tok2Vec component", "option", "er", int) = 2000,
-    entity_multitasks="",
+    n_iter: ("Number of iterations", "option", "n", int) = 30,
-    noise_level=0.0,
+    n_early_stopping: ("Maximum number of training epochs without dev accuracy improvement", "option", "ne", int) = None,
-    orth_variant_level=0.0,
+    n_examples: ("Number of examples", "option", "ns", int) = 0,
-    eval_beam_widths="",
+    use_gpu: ("Use GPU", "option", "g", int) = -1,
-    gold_preproc=False,
+    version: ("Model version", "option", "V", str) = "0.0.0",
-    learn_tokens=False,
+    meta_path: ("Optional path to meta.json to use as base.", "option", "m", Path) = None,
-    textcat_multilabel=False,
+    init_tok2vec: ("Path to pretrained weights for the token-to-vector parts of the models. See 'spacy pretrain'. Experimental.", "option", "t2v", Path) = None,
-    textcat_arch="bow",
+    parser_multitasks: ("Side objectives for parser CNN, e.g. 'dep' or 'dep,tag'", "option", "pt", str) = "",
-    textcat_positive_label=None,
+    entity_multitasks: ("Side objectives for NER CNN, e.g. 'dep' or 'dep,tag'", "option", "et", str) = "",
-    verbose=False,
+    noise_level: ("Amount of corruption for data augmentation", "option", "nl", float) = 0.0,
-    debug=False,
+    orth_variant_level: ("Amount of orthography variation for data augmentation", "option", "ovl", float) = 0.0,
+    eval_beam_widths: ("Beam widths to evaluate, e.g. 4,8", "option", "bw", str) = "",
+    gold_preproc: ("Use gold preprocessing", "flag", "G", bool) = False,
+    learn_tokens: ("Make parser learn gold-standard tokenization", "flag", "T", bool) = False,
+    textcat_multilabel: ("Textcat classes aren't mutually exclusive (multilabel)", "flag", "TML", bool) = False,
+    textcat_arch: ("Textcat model architecture", "option", "ta", str) = "bow",
+    textcat_positive_label: ("Textcat positive label for binary classes with two labels", "option", "tpl", str) = None,
+    tag_map_path: ("Location of JSON-formatted tag map", "option", "tm", Path) = None,
+    verbose: ("Display more information for debug", "flag", "VV", bool) = False,
+    debug: ("Run data diagnostics before training", "flag", "D", bool) = False,
+    # fmt: on
 ):
     """
     Train or update a spaCy model. Requires data to be formatted in spaCy's
@@ -116,7 +89,11 @@ def train(
         )
     if not output_path.exists():
         output_path.mkdir()
+        msg.good(f"Created output directory: {output_path}")
 
+    tag_map = {}
+    if tag_map_path is not None:
+        tag_map = srsly.read_json(tag_map_path)
     # Take dropout and batch size as generators of values -- dropout
     # starts high and decays sharply, to force the optimizer to explore.
     # Batch size starts at 1 and grows, so that we make updates quickly
@@ -145,28 +122,29 @@ def train(
     # the model and make sure the pipeline matches the pipeline setting. If
     # training starts from a blank model, intitalize the language class.
     pipeline = [p.strip() for p in pipeline.split(",")]
+    msg.text(f"Training pipeline: {pipeline}")
     disabled_pipes = None
     pipes_added = False
-    msg.text("Training pipeline: {}".format(pipeline))
+    msg.text(f"Training pipeline: {pipeline}")
     if use_gpu >= 0:
         activated_gpu = None
         try:
             activated_gpu = set_gpu(use_gpu)
         except Exception as e:
-            msg.warn("Exception: {}".format(e))
+            msg.warn(f"Exception: {e}")
         if activated_gpu is not None:
-            msg.text("Using GPU: {}".format(use_gpu))
+            msg.text(f"Using GPU: {use_gpu}")
         else:
-            msg.warn("Unable to activate GPU: {}".format(use_gpu))
+            msg.warn(f"Unable to activate GPU: {use_gpu}")
             msg.text("Using CPU only")
             use_gpu = -1
     if base_model:
-        msg.text("Starting with base model '{}'".format(base_model))
+        msg.text(f"Starting with base model '{base_model}'")
         nlp = util.load_model(base_model)
         if nlp.lang != lang:
             msg.fail(
-                "Model language ('{}') doesn't match language specified as "
+                f"Model language ('{nlp.lang}') doesn't match language "
-                "`lang` argument ('{}') ".format(nlp.lang, lang),
+                f"specified as `lang` argument ('{lang}') ",
                 exits=1,
             )
         for pipe in pipeline:
@@ -180,11 +158,11 @@ def train(
                     "positive_label": textcat_positive_label,
                 }
             if pipe not in nlp.pipe_names:
-                msg.text("Adding component to base model '{}'".format(pipe))
+                msg.text(f"Adding component to base model '{pipe}'")
                 nlp.add_pipe(nlp.create_pipe(pipe, config=pipe_cfg))
                 pipes_added = True
             elif replace_components:
-                msg.text("Replacing component from base model '{}'".format(pipe))
+                msg.text(f"Replacing component from base model '{pipe}'")
                 nlp.replace_pipe(pipe, nlp.create_pipe(pipe, config=pipe_cfg))
                 pipes_added = True
             else:
@@ -197,17 +175,17 @@ def train(
                     }
                     if base_cfg != pipe_cfg:
                         msg.fail(
-                            "The base textcat model configuration does"
+                            f"The base textcat model configuration does"
-                            "not match the provided training options. "
+                            f"not match the provided training options. "
-                            "Existing cfg: {}, provided cfg: {}".format(
+                            f"Existing cfg: {base_cfg}, provided cfg: {pipe_cfg}",
-                                base_cfg, pipe_cfg
-                            ),
                             exits=1,
                         )
-                msg.text("Extending component from base model '{}'".format(pipe))
+                msg.text(f"Extending component from base model '{pipe}'")
-        disabled_pipes = nlp.disable_pipes([p for p in nlp.pipe_names if p not in pipeline])
+        disabled_pipes = nlp.disable_pipes(
+            [p for p in nlp.pipe_names if p not in pipeline]
+        )
     else:
-        msg.text("Starting with blank model '{}'".format(lang))
+        msg.text(f"Starting with blank model '{lang}'")
         lang_cls = util.get_lang_class(lang)
         nlp = lang_cls()
         for pipe in pipeline:
@@ -223,8 +201,11 @@ def train(
                 pipe_cfg = {}
             nlp.add_pipe(nlp.create_pipe(pipe, config=pipe_cfg))
 
+    # Update tag map with provided mapping
+    nlp.vocab.morphology.tag_map.update(tag_map)
+
     if vectors:
-        msg.text("Loading vector from model '{}'".format(vectors))
+        msg.text(f"Loading vector from model '{vectors}'")
         _load_vectors(nlp, vectors)
 
     # Multitask objectives
@@ -233,49 +214,56 @@ def train(
         if multitasks:
             if pipe_name not in pipeline:
                 msg.fail(
-                    "Can't use multitask objective without '{}' in the "
+                    f"Can't use multitask objective without '{pipe_name}' in "
-                    "pipeline".format(pipe_name)
+                    f"the pipeline"
                 )
             pipe = nlp.get_pipe(pipe_name)
             for objective in multitasks.split(","):
                 pipe.add_multitask_objective(objective)
 
     # Prepare training corpus
-    msg.text("Counting training words (limit={})".format(n_examples))
+    msg.text(f"Counting training words (limit={n_examples})")
     corpus = GoldCorpus(train_path, dev_path, limit=n_examples)
     n_train_words = corpus.count_train()
 
     if base_model and not pipes_added:
         # Start with an existing model, use default optimizer
-        optimizer = create_default_optimizer(Model.ops)
+        optimizer = create_default_optimizer()
     else:
         # Start with a blank model, call begin_training
-        optimizer = nlp.begin_training(lambda: corpus.train_tuples, device=use_gpu)
+        cfg = {"device": use_gpu}
-
+        cfg["conv_depth"] = conv_depth
+        cfg["token_vector_width"] = width
+        cfg["bilstm_depth"] = bilstm_depth
+        cfg["cnn_maxout_pieces"] = cnn_pieces
+        cfg["embed_size"] = embed_rows
+        cfg["conv_window"] = cnn_window
+        cfg["subword_features"] = not use_chars
+        optimizer = nlp.begin_training(lambda: corpus.train_tuples, **cfg)
     nlp._optimizer = None
 
     # Load in pretrained weights
     if init_tok2vec is not None:
         components = _load_pretrained_tok2vec(nlp, init_tok2vec)
-        msg.text("Loaded pretrained tok2vec for: {}".format(components))
+        msg.text(f"Loaded pretrained tok2vec for: {components}")
 
     # Verify textcat config
     if "textcat" in pipeline:
         textcat_labels = nlp.get_pipe("textcat").cfg.get("labels", [])
         if textcat_positive_label and textcat_positive_label not in textcat_labels:
             msg.fail(
-                "The textcat_positive_label (tpl) '{}' does not match any "
+                f"The textcat_positive_label (tpl) '{textcat_positive_label}' "
-                "label in the training data.".format(textcat_positive_label),
+                f"does not match any label in the training data.",
                 exits=1,
             )
         if textcat_positive_label and len(textcat_labels) != 2:
             msg.fail(
-                "A textcat_positive_label (tpl) '{}' was provided for training "
+                "A textcat_positive_label (tpl) '{textcat_positive_label}' was "
-                "data that does not appear to be a binary classification "
+                "provided for training data that does not appear to be a "
-                "problem with two labels.".format(textcat_positive_label),
+                "binary classification problem with two labels.",
                 exits=1,
             )
-        train_docs = corpus.train_docs(
+        train_data = corpus.train_data(
             nlp,
             noise_level=noise_level,
             gold_preproc=gold_preproc,
@@ -285,9 +273,9 @@ def train(
         train_labels = set()
         if textcat_multilabel:
             multilabel_found = False
-            for text, gold in train_docs:
+            for ex in train_data:
-                train_labels.update(gold.cats.keys())
+                train_labels.update(ex.gold.cats.keys())
-                if list(gold.cats.values()).count(1.0) != 1:
+                if list(ex.gold.cats.values()).count(1.0) != 1:
                     multilabel_found = True
             if not multilabel_found and not base_model:
                 msg.warn(
@@ -297,9 +285,9 @@ def train(
                     "mutually-exclusive classes."
                 )
         if not textcat_multilabel:
-            for text, gold in train_docs:
+            for ex in train_data:
-                train_labels.update(gold.cats.keys())
+                train_labels.update(ex.gold.cats.keys())
-                if list(gold.cats.values()).count(1.0) != 1 and not base_model:
+                if list(ex.gold.cats.values()).count(1.0) != 1 and not base_model:
                     msg.warn(
                         "Some textcat training instances do not have exactly "
                         "one positive label. Modifying training options to "
@@ -311,20 +299,20 @@ def train(
                     break
         if base_model and set(textcat_labels) != train_labels:
             msg.fail(
-                "Cannot extend textcat model using data with different "
+                f"Cannot extend textcat model using data with different "
-                "labels. Base model labels: {}, training data labels: "
+                f"labels. Base model labels: {textcat_labels}, training data "
-                "{}.".format(textcat_labels, list(train_labels)),
+                f"labels: {list(train_labels)}",
                 exits=1,
             )
         if textcat_multilabel:
             msg.text(
-                "Textcat evaluation score: ROC AUC score macro-averaged across "
+                f"Textcat evaluation score: ROC AUC score macro-averaged across "
-                "the labels '{}'".format(", ".join(textcat_labels))
+                f"the labels '{', '.join(textcat_labels)}'"
             )
         elif textcat_positive_label and len(textcat_labels) == 2:
             msg.text(
-                "Textcat evaluation score: F1-score for the "
+                f"Textcat evaluation score: F1-score for the "
-                "label '{}'".format(textcat_positive_label)
+                f"label '{textcat_positive_label}'"
             )
         elif len(textcat_labels) > 1:
             if len(textcat_labels) == 2:
@@ -334,8 +322,8 @@ def train(
                     "an evaluation on the positive class."
                 )
             msg.text(
-                "Textcat evaluation score: F1-score macro-averaged across "
+                f"Textcat evaluation score: F1-score macro-averaged across "
-                "the labels '{}'".format(", ".join(textcat_labels))
+                f"the labels '{', '.join(textcat_labels)}'"
             )
         else:
             msg.fail(
@@ -355,7 +343,7 @@ def train(
         iter_since_best = 0
         best_score = 0.0
         for i in range(n_iter):
-            train_docs = corpus.train_docs(
+            train_data = corpus.train_dataset(
                 nlp,
                 noise_level=noise_level,
                 orth_variant_level=orth_variant_level,
@@ -371,10 +359,11 @@ def train(
             words_seen = 0
             with tqdm.tqdm(total=n_train_words, leave=False) as pbar:
                 losses = {}
-                for batch in util.minibatch_by_words(train_docs, size=batch_sizes):
+                for batch in util.minibatch_by_words(train_data, size=batch_sizes):
                     if not batch:
                         continue
                     docs, golds = zip(*batch)
+                    try:
                         nlp.update(
                             docs,
                             golds,
@@ -382,62 +371,70 @@ def train(
                             drop=next(dropout_rates),
                             losses=losses,
                         )
+                    except ValueError as e:
+                        msg.warn("Error during training")
+                        if init_tok2vec:
+                            msg.warn(
+                                "Did you provide the same parameters during 'train' as during 'pretrain'?"
+                            )
+                        msg.fail(f"Original error message: {e}", exits=1)
                     if raw_text:
                         # If raw text is available, perform 'rehearsal' updates,
                         # which use unlabelled data to reduce overfitting.
                         raw_batch = list(next(raw_batches))
                         nlp.rehearse(raw_batch, sgd=optimizer, losses=losses)
+                    docs = [ex.doc for ex in batch]
                     if not int(os.environ.get("LOG_FRIENDLY", 0)):
                         pbar.update(sum(len(doc) for doc in docs))
                     words_seen += sum(len(doc) for doc in docs)
             with nlp.use_params(optimizer.averages):
                 util.set_env_log(False)
-                epoch_model_path = output_path / ("model%d" % i)
+                epoch_model_path = output_path / f"model{i}"
                 nlp.to_disk(epoch_model_path)
                 nlp_loaded = util.load_model_from_path(epoch_model_path)
                 for beam_width in eval_beam_widths:
                     for name, component in nlp_loaded.pipeline:
                         if hasattr(component, "cfg"):
                             component.cfg["beam_width"] = beam_width
-                    dev_docs = list(
+                    dev_dataset = list(
-                        corpus.dev_docs(
+                        corpus.dev_dataset(
                             nlp_loaded,
                             gold_preproc=gold_preproc,
                             ignore_misaligned=True,
                         )
                     )
-                    nwords = sum(len(doc_gold[0]) for doc_gold in dev_docs)
+                    nwords = sum(len(ex.doc) for ex in dev_dataset)
                     start_time = timer()
-                    scorer = nlp_loaded.evaluate(dev_docs, verbose=verbose)
+                    scorer = nlp_loaded.evaluate(dev_dataset, verbose=verbose)
                     end_time = timer()
                     if use_gpu < 0:
                         gpu_wps = None
                         cpu_wps = nwords / (end_time - start_time)
                     else:
                         gpu_wps = nwords / (end_time - start_time)
-                        with Model.use_device("cpu"):
+                        with use_ops("numpy"):
                             nlp_loaded = util.load_model_from_path(epoch_model_path)
                             for name, component in nlp_loaded.pipeline:
                                 if hasattr(component, "cfg"):
                                     component.cfg["beam_width"] = beam_width
-                            dev_docs = list(
+                            dev_dataset = list(
-                                corpus.dev_docs(
+                                corpus.dev_dataset(
                                     nlp_loaded,
                                     gold_preproc=gold_preproc,
                                     ignore_misaligned=True,
                                 )
                             )
                             start_time = timer()
-                            scorer = nlp_loaded.evaluate(dev_docs, verbose=verbose)
+                            scorer = nlp_loaded.evaluate(dev_dataset, verbose=verbose)
                             end_time = timer()
                             cpu_wps = nwords / (end_time - start_time)
-                    acc_loc = output_path / ("model%d" % i) / "accuracy.json"
+                    acc_loc = output_path / f"model{i}" / "accuracy.json"
                     srsly.write_json(acc_loc, scorer.scores)
 
                     # Update model meta.json
                     meta["lang"] = nlp.lang
                     meta["pipeline"] = nlp.pipe_names
-                    meta["spacy_version"] = ">=%s" % about.__version__
+                    meta["spacy_version"] = f">={about.__version__}"
                     if beam_width == 1:
                         meta["speed"] = {
                             "nwords": nwords,
@@ -465,10 +462,10 @@ def train(
                         "keys": nlp.vocab.vectors.n_keys,
                         "name": nlp.vocab.vectors.name,
                     }
-                    meta.setdefault("name", "model%d" % i)
+                    meta.setdefault("name", f"model{i}")
                     meta.setdefault("version", version)
                     meta["labels"] = nlp.meta["labels"]
-                    meta_loc = output_path / ("model%d" % i) / "meta.json"
+                    meta_loc = output_path / f"model{i}" / "meta.json"
                     srsly.write_json(meta_loc, meta)
                     util.set_env_log(verbose)
 
@@ -486,8 +483,8 @@ def train(
                         for cat, cat_score in textcats_per_cat.items():
                             if cat_score.get("roc_auc_score", 0) < 0:
                                 msg.warn(
-                                    "Textcat ROC AUC score is undefined due to "
+                                    f"Textcat ROC AUC score is undefined due to "
-                                    "only one value in label '{}'.".format(cat)
+                                    f"only one value in label '{cat}'."
                                 )
                     msg.row(progress, **row_settings)
                 # Early stopping
@@ -500,14 +497,14 @@ def train(
                         best_score = current_score
                     if iter_since_best >= n_early_stopping:
                         msg.text(
-                            "Early stopping, best iteration "
+                            f"Early stopping, best iteration is: {i - iter_since_best}"
-                            "is: {}".format(i - iter_since_best)
                         )
                         msg.text(
-                            "Best score = {}; Final iteration "
+                            f"Best score = {best_score}; Final iteration score = {current_score}"
-                            "score = {}".format(best_score, current_score)
                         )
                         break
+    except Exception as e:
+        msg.warn(f"Aborting and saving final best model. Encountered exception: {e}")
     finally:
         best_pipes = nlp.pipe_names
         if disabled_pipes:
@@ -535,6 +532,8 @@ def _score_for_model(meta):
         mean_acc.append((acc["ents_p"] + acc["ents_r"] + acc["ents_f"]) / 3)
     if "textcat" in pipes:
         mean_acc.append(acc["textcat_score"])
+    if "sentrec" in pipes:
+        mean_acc.append((acc["sent_p"] + acc["sent_r"] + acc["sent_f"]) / 3)
     return sum(mean_acc) / len(mean_acc)
 
 
@@ -580,12 +579,10 @@ def _collate_best_model(meta, output_path, components):
     for component in components:
         bests[component] = _find_best(output_path, component)
     best_dest = output_path / "model-best"
-    shutil.copytree(path2str(output_path / "model-final"), path2str(best_dest))
+    shutil.copytree(str(output_path / "model-final"), str(best_dest))
     for component, best_component_src in bests.items():
-        shutil.rmtree(path2str(best_dest / component))
+        shutil.rmtree(str(best_dest / component))
-        shutil.copytree(
+        shutil.copytree(str(best_component_src / component), str(best_dest / component))
-            path2str(best_component_src / component), path2str(best_dest / component)
-        )
         accs = srsly.read_json(best_component_src / "accuracy.json")
         for metric in _get_metrics(component):
             meta["accuracy"][metric] = accs[metric]
@@ -608,11 +605,13 @@ def _find_best(experiment_dir, component):
 
 def _get_metrics(component):
     if component == "parser":
-        return ("las", "uas", "las_per_type", "token_acc")
+        return ("las", "uas", "las_per_type", "token_acc", "sent_f")
     elif component == "tagger":
         return ("tags_acc",)
     elif component == "ner":
-        return ("ents_f", "ents_p", "ents_r", "ents_per_type")
+        return ("ents_f", "ents_p", "ents_r", "enty_per_type")
+    elif component == "sentrec":
+        return ("sent_f", "sent_p", "sent_r")
     elif component == "textcat":
         return ("textcat_score",)
     return ("token_acc",)
@@ -626,14 +625,21 @@ def _configure_training_output(pipeline, use_gpu, has_beam_widths):
             row_head.extend(["Tag Loss ", " Tag %  "])
             output_stats.extend(["tag_loss", "tags_acc"])
         elif pipe == "parser":
-            row_head.extend(["Dep Loss ", " UAS  ", " LAS  "])
+            row_head.extend(
-            output_stats.extend(["dep_loss", "uas", "las"])
+                ["Dep Loss ", " UAS  ", " LAS  ", "Sent P", "Sent R", "Sent F"]
+            )
+            output_stats.extend(
+                ["dep_loss", "uas", "las", "sent_p", "sent_r", "sent_f"]
+            )
         elif pipe == "ner":
             row_head.extend(["NER Loss ", "NER P ", "NER R ", "NER F "])
             output_stats.extend(["ner_loss", "ents_p", "ents_r", "ents_f"])
         elif pipe == "textcat":
             row_head.extend(["Textcat Loss", "Textcat"])
             output_stats.extend(["textcat_loss", "textcat_score"])
+        elif pipe == "sentrec":
+            row_head.extend(["Sentrec Loss", "Sent P", "Sent R", "Sent F"])
+            output_stats.extend(["sentrec_loss", "sent_p", "sent_r", "sent_f"])
     row_head.extend(["Token %", "CPU WPS"])
     output_stats.extend(["token_acc", "cpu_wps"])
 
@@ -643,7 +649,10 @@ def _configure_training_output(pipeline, use_gpu, has_beam_widths):
 
     if has_beam_widths:
         row_head.insert(1, "Beam W.")
-    return row_head, output_stats
+    # remove duplicates
+    row_head_dict = {k: 1 for k in row_head}
+    output_stats_dict = {k: 1 for k in output_stats}
+    return row_head_dict.keys(), output_stats_dict.keys()
 
 
 def _get_progress(
@@ -656,6 +665,7 @@ def _get_progress(
     scores["ner_loss"] = losses.get("ner", 0.0)
     scores["tag_loss"] = losses.get("tagger", 0.0)
     scores["textcat_loss"] = losses.get("textcat", 0.0)
+    scores["sentrec_loss"] = losses.get("sentrec", 0.0)
     scores["cpu_wps"] = cpu_wps
     scores["gpu_wps"] = gpu_wps or 0.0
     scores.update(dev_scores)

spacy/cli/train_from_config.py

@@ -0,0 +1,439 @@
+from typing import Optional, Dict, List, Union, Sequence
+import plac
+from wasabi import msg
+from pathlib import Path
+import thinc
+import thinc.schedules
+from thinc.api import Model
+from pydantic import BaseModel, FilePath, StrictInt
+import tqdm
+
+# TODO: relative imports?
+import spacy
+from spacy.gold import GoldCorpus
+from spacy.pipeline.tok2vec import Tok2VecListener
+from spacy.ml import component_models
+from spacy import util
+
+
+registry = util.registry
+
+CONFIG_STR = """
+[training]
+patience = 10
+eval_frequency = 10
+dropout = 0.2
+init_tok2vec = null
+vectors = null
+max_epochs = 100
+orth_variant_level = 0.0
+gold_preproc = false
+max_length = 0
+use_gpu = 0
+scores = ["ents_p",  "ents_r", "ents_f"]
+score_weights = {"ents_f": 1.0}
+limit = 0
+
+[training.batch_size]
+@schedules = "compounding.v1"
+start = 100
+stop = 1000
+compound = 1.001
+
+[optimizer]
+@optimizers = "Adam.v1"
+learn_rate = 0.001
+beta1 = 0.9
+beta2 = 0.999
+
+[nlp]
+lang = "en"
+vectors = ${training:vectors}
+
+[nlp.pipeline.tok2vec]
+factory = "tok2vec"
+
+[nlp.pipeline.ner]
+factory = "ner"
+
+[nlp.pipeline.ner.model]
+@architectures = "transition_based_ner.v1"
+nr_feature_tokens = 3
+hidden_width = 64
+maxout_pieces = 3
+
+[nlp.pipeline.ner.model.tok2vec]
+@architectures = "tok2vec_tensors.v1"
+width = ${nlp.pipeline.tok2vec.model:width}
+
+[nlp.pipeline.tok2vec.model]
+@architectures = "hash_embed_cnn.v1"
+pretrained_vectors = ${nlp:vectors}
+width = 128
+depth = 4
+window_size = 1
+embed_size = 10000
+maxout_pieces = 3
+"""
+
+
+class PipelineComponent(BaseModel):
+    factory: str
+    model: Model
+
+    class Config:
+        arbitrary_types_allowed = True
+
+
+class ConfigSchema(BaseModel):
+    optimizer: Optional["Optimizer"]
+
+    class training(BaseModel):
+        patience: int = 10
+        eval_frequency: int = 100
+        dropout: float = 0.2
+        init_tok2vec: Optional[FilePath] = None
+        vectors: Optional[str] = None
+        max_epochs: int = 100
+        orth_variant_level: float = 0.0
+        gold_preproc: bool = False
+        max_length: int = 0
+        use_gpu: int = 0
+        scores: List[str] = ["ents_p", "ents_r", "ents_f"]
+        score_weights: Dict[str, Union[int, float]] = {"ents_f": 1.0}
+        limit: int = 0
+        batch_size: Union[Sequence[int], int]
+
+    class nlp(BaseModel):
+        lang: str
+        vectors: Optional[str]
+        pipeline: Optional[Dict[str, PipelineComponent]]
+
+    class Config:
+        extra = "allow"
+
+
+# Of course, these would normally decorate the functions where they're defined.
+# But for now...
+@registry.architectures.register("hash_embed_cnn.v1")
+def hash_embed_cnn(
+    pretrained_vectors, width, depth, embed_size, maxout_pieces, window_size
+):
+    return component_models.Tok2Vec(
+        width=width,
+        embed_size=embed_size,
+        pretrained_vectors=pretrained_vectors,
+        conv_depth=depth,
+        cnn_maxout_pieces=maxout_pieces,
+        bilstm_depth=0,
+        window_size=window_size,
+    )
+
+
+@registry.architectures.register("hash_embed_bilstm.v1")
+def hash_embed_bilstm_v1(pretrained_vectors, width, depth, embed_size):
+    return component_models.Tok2Vec(
+        width=width,
+        embed_size=embed_size,
+        pretrained_vectors=pretrained_vectors,
+        bilstm_depth=depth,
+        conv_depth=0,
+        cnn_maxout_pieces=0,
+    )
+
+
+@registry.architectures.register("tagger_model.v1")
+def build_tagger_model_v1(tok2vec):
+    return component_models.build_tagger_model(nr_class=None, tok2vec=tok2vec)
+
+
+@registry.architectures.register("transition_based_parser.v1")
+def create_tb_parser_model(
+    tok2vec: Model,
+    nr_feature_tokens: StrictInt = 3,
+    hidden_width: StrictInt = 64,
+    maxout_pieces: StrictInt = 3,
+):
+    from thinc.api import Linear, chain, list2array, use_ops, zero_init
+    from spacy.ml._layers import PrecomputableAffine
+    from spacy.syntax._parser_model import ParserModel
+
+    token_vector_width = tok2vec.get_dim("nO")
+    tok2vec = chain(tok2vec, list2array())
+    tok2vec.set_dim("nO", token_vector_width)
+
+    lower = PrecomputableAffine(
+        hidden_width, nF=nr_feature_tokens, nI=tok2vec.get_dim("nO"), nP=maxout_pieces
+    )
+    lower.set_dim("nP", maxout_pieces)
+    with use_ops("numpy"):
+        # Initialize weights at zero, as it's a classification layer.
+        upper = Linear(init_W=zero_init)
+    return ParserModel(tok2vec, lower, upper)
+
+
+@plac.annotations(
+    # fmt: off
+    train_path=("Location of JSON-formatted training data", "positional", None, Path),
+    dev_path=("Location of JSON-formatted development data", "positional", None, Path),
+    config_path=("Path to config file", "positional", None, Path),
+    output_path=("Output directory to store model in", "option", "o", Path),
+    meta_path=("Optional path to meta.json to use as base.", "option", "m", Path),
+    raw_text=("Path to jsonl file with unlabelled text documents.", "option", "rt", Path),
+    # fmt: on
+)
+def train_from_config_cli(
+    train_path,
+    dev_path,
+    config_path,
+    output_path=None,
+    meta_path=None,
+    raw_text=None,
+    debug=False,
+    verbose=False,
+):
+    """
+    Train or update a spaCy model. Requires data to be formatted in spaCy's
+    JSON format. To convert data from other formats, use the `spacy convert`
+    command.
+    """
+    if not config_path or not config_path.exists():
+        msg.fail("Config file not found", config_path, exits=1)
+    if not train_path or not train_path.exists():
+        msg.fail("Training data not found", train_path, exits=1)
+    if not dev_path or not dev_path.exists():
+        msg.fail("Development data not found", dev_path, exits=1)
+    if meta_path is not None and not meta_path.exists():
+        msg.fail("Can't find model meta.json", meta_path, exits=1)
+    if output_path is not None and not output_path.exists():
+        output_path.mkdir()
+
+    try:
+        train_from_config(
+            config_path,
+            {"train": train_path, "dev": dev_path},
+            output_path=output_path,
+            meta_path=meta_path,
+            raw_text=raw_text,
+        )
+    except KeyboardInterrupt:
+        msg.warn("Cancelled.")
+
+
+def train_from_config(
+    config_path, data_paths, raw_text=None, meta_path=None, output_path=None,
+):
+    msg.info(f"Loading config from: {config_path}")
+    config = util.load_from_config(config_path, create_objects=True)
+    use_gpu = config["training"]["use_gpu"]
+    if use_gpu >= 0:
+        msg.info("Using GPU")
+    else:
+        msg.info("Using CPU")
+    msg.info("Creating nlp from config")
+    nlp = create_nlp_from_config(**config["nlp"])
+    optimizer = config["optimizer"]
+    limit = config["training"]["limit"]
+    msg.info("Loading training corpus")
+    corpus = GoldCorpus(data_paths["train"], data_paths["dev"], limit=limit)
+    msg.info("Initializing the nlp pipeline")
+    nlp.begin_training(lambda: corpus.train_examples, device=use_gpu)
+
+    train_batches = create_train_batches(nlp, corpus, config["training"])
+    evaluate = create_evaluation_callback(nlp, optimizer, corpus, config["training"])
+
+    # Create iterator, which yields out info after each optimization step.
+    msg.info("Start training")
+    training_step_iterator = train_while_improving(
+        nlp,
+        optimizer,
+        train_batches,
+        evaluate,
+        config["training"]["dropout"],
+        config["training"]["patience"],
+        config["training"]["eval_frequency"],
+    )
+
+    msg.info(f"Training. Initial learn rate: {optimizer.learn_rate}")
+    print_row = setup_printer(config)
+
+    try:
+        progress = tqdm.tqdm(total=config["training"]["eval_frequency"], leave=False)
+        for batch, info, is_best_checkpoint in training_step_iterator:
+            progress.update(1)
+            if is_best_checkpoint is not None:
+                progress.close()
+                print_row(info)
+                if is_best_checkpoint and output_path is not None:
+                    nlp.to_disk(output_path)
+                progress = tqdm.tqdm(
+                    total=config["training"]["eval_frequency"], leave=False
+                )
+    finally:
+        if output_path is not None:
+            with nlp.use_params(optimizer.averages):
+                final_model_path = output_path / "model-final"
+                nlp.to_disk(final_model_path)
+            msg.good("Saved model to output directory", final_model_path)
+        # with msg.loading("Creating best model..."):
+        #     best_model_path = _collate_best_model(meta, output_path, nlp.pipe_names)
+        # msg.good("Created best model", best_model_path)
+
+
+def create_nlp_from_config(lang, vectors, pipeline):
+    lang_class = spacy.util.get_lang_class(lang)
+    nlp = lang_class()
+    if vectors is not None:
+        spacy.cli.train._load_vectors(nlp, vectors)
+    for name, component_cfg in pipeline.items():
+        factory = component_cfg.pop("factory")
+        component = nlp.create_pipe(factory, config=component_cfg)
+        nlp.add_pipe(component, name=name)
+    return nlp
+
+
+def create_train_batches(nlp, corpus, cfg):
+    while True:
+        train_examples = corpus.train_dataset(
+            nlp,
+            noise_level=0.0,
+            orth_variant_level=cfg["orth_variant_level"],
+            gold_preproc=cfg["gold_preproc"],
+            max_length=cfg["max_length"],
+            ignore_misaligned=True,
+        )
+        for batch in util.minibatch_by_words(train_examples, size=cfg["batch_size"]):
+            yield batch
+
+
+def create_evaluation_callback(nlp, optimizer, corpus, cfg):
+    def evaluate():
+        with nlp.use_params(optimizer.averages):
+            dev_examples = list(
+                corpus.dev_dataset(
+                    nlp, gold_preproc=cfg["gold_preproc"], ignore_misaligned=True
+                )
+            )
+            scorer = nlp.evaluate(dev_examples)
+            scores = scorer.scores
+            # Calculate a weighted sum based on score_weights for the main score
+            weights = cfg["score_weights"]
+            weighted_score = sum(scores[s] * weights.get(s, 0.0) for s in weights)
+        return weighted_score, scorer.scores
+
+    return evaluate
+
+
+def train_while_improving(
+    nlp, optimizer, train_data, evaluate, dropout, patience, eval_frequency
+):
+    """Train until an evaluation stops improving. Works as a generator,
+    with each iteration yielding a tuple `(batch, info, is_best_checkpoint)`,
+    where info is a dict, and is_best_checkpoint is in [True, False, None] --
+    None indicating that the iteration was not evaluated as a checkpoint.
+    The evaluation is conducted by calling the evaluate callback, which should
+
+    Positional arguments:
+        nlp: The spaCy pipeline to evaluate.
+        train_data (Iterable[Batch]): A generator of batches, with the training
+            data. Each batch should be a Sized[Tuple[Input, Annot]]. The training
+            data iterable needs to take care of iterating over the epochs and
+            shuffling.
+        evaluate (Callable[[], Tuple[float, Any]]): A callback to perform evaluation.
+            The callback should take no arguments and return a tuple
+            `(main_score, other_scores)`. The main_score should be a float where
+            higher is better. other_scores can be any object.
+
+    Every iteration, the function yields out a tuple with:
+
+    * batch: A zipped sequence of Tuple[Doc, GoldParse] pairs.
+    * info: A dict with various information about the last update (see below).
+    * is_best_checkpoint: A value in None, False, True, indicating whether this
+        was the best evaluation so far. You should use this to save the model
+        checkpoints during training. If None, evaluation was not conducted on
+        that iteration. False means evaluation was conducted, but a previous
+        evaluation was better.
+
+    The info dict provides the following information:
+
+        epoch (int): How many passes over the data have been completed.
+        step (int): How many steps have been completed.
+        score (float): The main score form the last evaluation.
+        other_scores: : The other scores from the last evaluation.
+        loss: The accumulated losses throughout training.
+        checkpoints: A list of previous results, where each result is a
+            (score, step, epoch) tuple.
+    """
+    if isinstance(dropout, float):
+        dropouts = thinc.schedules.constant(dropout)
+    else:
+        dropouts = dropout
+    results = []
+    losses = {}
+    for step, batch in enumerate(train_data):
+        dropout = next(dropouts)
+        for subbatch in subdivide_batch(batch):
+            nlp.update(subbatch, drop=dropout, losses=losses, sgd=False)
+        for name, proc in nlp.pipeline:
+            if hasattr(proc, "model"):
+                proc.model.finish_update(optimizer)
+        optimizer.step_schedules()
+        if not (step % eval_frequency):
+            score, other_scores = evaluate()
+            results.append((score, step))
+            is_best_checkpoint = score == max(results)[0]
+        else:
+            score, other_scores = (None, None)
+            is_best_checkpoint = None
+        info = {
+            "step": step,
+            "score": score,
+            "other_scores": other_scores,
+            "losses": losses,
+            "checkpoints": results,
+        }
+        yield batch, info, is_best_checkpoint
+        if is_best_checkpoint is not None:
+            losses = {}
+        # Stop if no improvement in `patience` updates
+        best_score, best_step = max(results)
+        if (step - best_step) >= patience:
+            break
+
+
+def subdivide_batch(batch):
+    return [batch]
+
+
+def setup_printer(config):
+    score_cols = config["training"]["scores"]
+    score_widths = [max(len(col), 6) for col in score_cols]
+    loss_cols = [f"Loss {pipe}" for pipe in config["nlp"]["pipeline"]]
+    loss_widths = [max(len(col), 8) for col in loss_cols]
+    table_header = ["#"] + loss_cols + score_cols + ["Score"]
+    table_header = [col.upper() for col in table_header]
+    table_widths = [6] + loss_widths + score_widths + [6]
+    table_aligns = ["r" for _ in table_widths]
+
+    msg.row(table_header, widths=table_widths)
+    msg.row(["-" * width for width in table_widths])
+
+    def print_row(info):
+        losses = [
+            "{0:.2f}".format(info["losses"].get(col, 0.0))
+            for col in config["nlp"]["pipeline"]
+        ]
+        scores = [
+            "{0:.2f}".format(info["other_scores"].get(col, 0.0))
+            for col in config["training"]["scores"]
+        ]
+        data = [info["step"]] + losses + scores + ["{0:.2f}".format(info["score"])]
+        msg.row(data, widths=table_widths, aligns=table_aligns)
+
+    return print_row
+
+
+@registry.architectures.register("tok2vec_tensors.v1")
+def tok2vec_tensors_v1(width):
+    tok2vec = Tok2VecListener("tok2vec", width=width)
+    return tok2vec

spacy/cli/validate.py

@@ -1,14 +1,8 @@
-# coding: utf8
-from __future__ import unicode_literals, print_function
-
 from pathlib import Path
 import sys
 import requests
-import srsly
 from wasabi import msg
 
-from ..compat import path2str
-from ..util import get_data_path
 from .. import about
 
 
@@ -17,51 +11,30 @@ def validate():
     Validate that the currently installed version of spaCy is compatible
     with the installed models. Should be run after `pip install -U spacy`.
     """
-    with msg.loading("Loading compatibility table..."):
+    model_pkgs, compat = get_model_pkgs()
-        r = requests.get(about.__compatibility__)
+    spacy_version = about.__version__.rsplit(".dev", 1)[0]
-        if r.status_code != 200:
+    current_compat = compat.get(spacy_version, {})
-            msg.fail(
-                "Server error ({})".format(r.status_code),
-                "Couldn't fetch compatibility table.",
-                exits=1,
-            )
-    msg.good("Loaded compatibility table")
-    compat = r.json()["spacy"]
-    version = about.__version__
-    version = version.rsplit(".dev", 1)[0]
-    current_compat = compat.get(version)
     if not current_compat:
-        msg.fail(
+        msg.warn(f"No compatible models found for v{spacy_version} of spaCy")
-            "Can't find spaCy v{} in compatibility table".format(version),
-            about.__compatibility__,
-            exits=1,
-        )
-    all_models = set()
-    for spacy_v, models in dict(compat).items():
-        all_models.update(models.keys())
-        for model, model_vs in models.items():
-            compat[spacy_v][model] = [reformat_version(v) for v in model_vs]
-    model_links = get_model_links(current_compat)
-    model_pkgs = get_model_pkgs(current_compat, all_models)
-    incompat_links = {l for l, d in model_links.items() if not d["compat"]}
     incompat_models = {d["name"] for _, d in model_pkgs.items() if not d["compat"]}
-    incompat_models.update(
-        [d["name"] for _, d in model_links.items() if not d["compat"]]
-    )
     na_models = [m for m in incompat_models if m not in current_compat]
     update_models = [m for m in incompat_models if m in current_compat]
     spacy_dir = Path(__file__).parent.parent
 
-    msg.divider("Installed models (spaCy v{})".format(about.__version__))
+    msg.divider(f"Installed models (spaCy v{about.__version__})")
-    msg.info("spaCy installation: {}".format(path2str(spacy_dir)))
+    msg.info(f"spaCy installation: {spacy_dir}")
 
-    if model_links or model_pkgs:
+    if model_pkgs:
-        header = ("TYPE", "NAME", "MODEL", "VERSION", "")
+        header = ("NAME", "VERSION", "")
         rows = []
         for name, data in model_pkgs.items():
-            rows.append(get_model_row(current_compat, name, data, msg))
+            if data["compat"]:
-        for name, data in model_links.items():
+                comp = msg.text("", color="green", icon="good", no_print=True)
-            rows.append(get_model_row(current_compat, name, data, msg, "link"))
+                version = msg.text(data["version"], color="green", no_print=True)
+            else:
+                version = msg.text(data["version"], color="red", no_print=True)
+                comp = f"--> {compat.get(data['name'], ['n/a'])[0]}"
+            rows.append((data["name"], version, comp))
         msg.table(rows, header=header)
     else:
         msg.text("No models found in your current environment.", exits=0)
@@ -71,44 +44,32 @@ def validate():
         cmd = "python -m spacy download {}"
         print("\n".join([cmd.format(pkg) for pkg in update_models]) + "\n")
     if na_models:
-        msg.text(
+        msg.warn(
-            "The following models are not available for spaCy "
+            f"The following models are not available for spaCy v{about.__version__}:",
-            "v{}: {}".format(about.__version__, ", ".join(na_models))
+            ", ".join(na_models),
         )
-    if incompat_links:
+    if incompat_models:
-        msg.text(
-            "You may also want to overwrite the incompatible links using the "
-            "`python -m spacy link` command with `--force`, or remove them "
-            "from the data directory. "
-            "Data path: {path}".format(path=path2str(get_data_path()))
-        )
-    if incompat_models or incompat_links:
         sys.exit(1)
 
 
-def get_model_links(compat):
+def get_model_pkgs():
-    links = {}
-    data_path = get_data_path()
-    if data_path:
-        models = [p for p in data_path.iterdir() if is_model_path(p)]
-        for model in models:
-            meta_path = Path(model) / "meta.json"
-            if not meta_path.exists():
-                continue
-            meta = srsly.read_json(meta_path)
-            link = model.parts[-1]
-            name = meta["lang"] + "_" + meta["name"]
-            links[link] = {
-                "name": name,
-                "version": meta["version"],
-                "compat": is_compat(compat, name, meta["version"]),
-            }
-    return links
-
-
-def get_model_pkgs(compat, all_models):
     import pkg_resources
 
+    with msg.loading("Loading compatibility table..."):
+        r = requests.get(about.__compatibility__)
+        if r.status_code != 200:
+            msg.fail(
+                f"Server error ({r.status_code})",
+                "Couldn't fetch compatibility table.",
+                exits=1,
+            )
+    msg.good("Loaded compatibility table")
+    compat = r.json()["spacy"]
+    all_models = set()
+    for spacy_v, models in dict(compat).items():
+        all_models.update(models.keys())
+        for model, model_vs in models.items():
+            compat[spacy_v][model] = [reformat_version(v) for v in model_vs]
     pkgs = {}
     for pkg_name, pkg_data in pkg_resources.working_set.by_key.items():
         package = pkg_name.replace("-", "_")
@@ -117,29 +78,9 @@ def get_model_pkgs(compat, all_models):
             pkgs[pkg_name] = {
                 "name": package,
                 "version": version,
-                "compat": is_compat(compat, package, version),
+                "compat": package in compat and version in compat[package],
             }
-    return pkgs
+    return pkgs, compat
-
-
-def get_model_row(compat, name, data, msg, model_type="package"):
-    if data["compat"]:
-        comp = msg.text("", color="green", icon="good", no_print=True)
-        version = msg.text(data["version"], color="green", no_print=True)
-    else:
-        version = msg.text(data["version"], color="red", no_print=True)
-        comp = "--> {}".format(compat.get(data["name"], ["n/a"])[0])
-    return (model_type, name, data["name"], version, comp)
-
-
-def is_model_path(model_path):
-    exclude = ["cache", "pycache", "__pycache__"]
-    name = model_path.parts[-1]
-    return model_path.is_dir() and name not in exclude and not name.startswith(".")
-
-
-def is_compat(compat, name, version):
-    return name in compat and version in compat[name]
 
 
 def reformat_version(version):

spacy/compat.py

@@ -1,4 +1,3 @@
-# coding: utf8
 """
 Helpers for Python and platform compatibility. To distinguish them from
 the builtin functions, replacement functions are suffixed with an underscore,
@@ -6,15 +5,9 @@ e.g. `unicode_`.
 
 DOCS: https://spacy.io/api/top-level#compat
 """
-from __future__ import unicode_literals
-
-import os
 import sys
-import itertools
-import ast
-import types
 
-from thinc.neural.util import copy_array
+from thinc.util import copy_array
 
 try:
     import cPickle as pickle
@@ -36,91 +29,23 @@ try:
 except ImportError:
     cupy = None
 
-try:
+from thinc.api import Optimizer  # noqa: F401
-    from thinc.neural.optimizers import Optimizer  # noqa: F401
-except ImportError:
-    from thinc.neural.optimizers import Adam as Optimizer  # noqa: F401
 
 pickle = pickle
 copy_reg = copy_reg
 CudaStream = CudaStream
 cupy = cupy
 copy_array = copy_array
-izip = getattr(itertools, "izip", zip)
 
 is_windows = sys.platform.startswith("win")
 is_linux = sys.platform.startswith("linux")
 is_osx = sys.platform == "darwin"
 
-# See: https://github.com/benjaminp/six/blob/master/six.py
-is_python2 = sys.version_info[0] == 2
-is_python3 = sys.version_info[0] == 3
-is_python_pre_3_5 = is_python2 or (is_python3 and sys.version_info[1] < 5)
 
-if is_python2:
+def is_config(windows=None, linux=None, osx=None, **kwargs):
-    bytes_ = str
-    unicode_ = unicode  # noqa: F821
-    basestring_ = basestring  # noqa: F821
-    input_ = raw_input  # noqa: F821
-    path2str = lambda path: str(path).decode("utf8")
-    class_types = (type, types.ClassType)
-
-elif is_python3:
-    bytes_ = bytes
-    unicode_ = str
-    basestring_ = str
-    input_ = input
-    path2str = lambda path: str(path)
-    class_types = (type, types.ClassType) if is_python_pre_3_5 else type
-
-
-def b_to_str(b_str):
-    """Convert a bytes object to a string.
-
-    b_str (bytes): The object to convert.
-    RETURNS (unicode): The converted string.
-    """
-    if is_python2:
-        return b_str
-    # Important: if no encoding is set, string becomes "b'...'"
-    return str(b_str, encoding="utf8")
-
-
-def symlink_to(orig, dest):
-    """Create a symlink. Used for model shortcut links.
-
-    orig (unicode / Path): The origin path.
-    dest (unicode / Path): The destination path of the symlink.
-    """
-    if is_windows:
-        import subprocess
-
-        subprocess.check_call(
-            ["mklink", "/d", path2str(orig), path2str(dest)], shell=True
-        )
-    else:
-        orig.symlink_to(dest)
-
-
-def symlink_remove(link):
-    """Remove a symlink. Used for model shortcut links.
-
-    link (unicode / Path): The path to the symlink.
-    """
-    # https://stackoverflow.com/q/26554135/6400719
-    if os.path.isdir(path2str(link)) and is_windows:
-        # this should only be on Py2.7 and windows
-        os.rmdir(path2str(link))
-    else:
-        os.unlink(path2str(link))
-
-
-def is_config(python2=None, python3=None, windows=None, linux=None, osx=None):
     """Check if a specific configuration of Python version and operating system
     matches the user's setup. Mostly used to display targeted error messages.
 
-    python2 (bool): spaCy is executed with Python 2.x.
-    python3 (bool): spaCy is executed with Python 3.x.
     windows (bool): spaCy is executed on Windows.
     linux (bool): spaCy is executed on Linux.
     osx (bool): spaCy is executed on OS X or macOS.
@@ -129,53 +54,7 @@ def is_config(python2=None, python3=None, windows=None, linux=None, osx=None):
     DOCS: https://spacy.io/api/top-level#compat.is_config
     """
     return (
-        python2 in (None, is_python2)
+        windows in (None, is_windows)
-        and python3 in (None, is_python3)
-        and windows in (None, is_windows)
         and linux in (None, is_linux)
         and osx in (None, is_osx)
     )
-
-
-def import_file(name, loc):
-    """Import module from a file. Used to load models from a directory.
-
-    name (unicode): Name of module to load.
-    loc (unicode / Path): Path to the file.
-    RETURNS: The loaded module.
-    """
-    loc = path2str(loc)
-    if is_python_pre_3_5:
-        import imp
-
-        return imp.load_source(name, loc)
-    else:
-        import importlib.util
-
-        spec = importlib.util.spec_from_file_location(name, str(loc))
-        module = importlib.util.module_from_spec(spec)
-        spec.loader.exec_module(module)
-        return module
-
-
-def unescape_unicode(string):
-    """Python2.7's re module chokes when compiling patterns that have ranges
-    between escaped unicode codepoints if the two codepoints are unrecognised
-    in the unicode database. For instance:
-
-        re.compile('[\\uAA77-\\uAA79]').findall("hello")
-
-    Ends up matching every character (on Python 2). This problem doesn't occur
-    if we're dealing with unicode literals.
-    """
-    if string is None:
-        return string
-    # We only want to unescape the unicode, so we first must protect the other
-    # backslashes.
-    string = string.replace("\\", "\\\\")
-    # Now we remove that protection for the unicode.
-    string = string.replace("\\\\u", "\\u")
-    string = string.replace("\\\\U", "\\U")
-    # Now we unescape by evaling the string with the AST. This can't execute
-    # code -- it only does the representational level.
-    return ast.literal_eval("u'''" + string + "'''")

spacy/displacy/__init__.py

@@ -1,15 +1,11 @@
-# coding: utf8
 """
 spaCy's built in visualization suite for dependencies and named entities.
 
 DOCS: https://spacy.io/api/top-level#displacy
 USAGE: https://spacy.io/usage/visualizers
 """
-from __future__ import unicode_literals
-
 from .render import DependencyRenderer, EntityRenderer
 from ..tokens import Doc, Span
-from ..compat import b_to_str
 from ..errors import Errors, Warnings, user_warning
 from ..util import is_in_jupyter
 
@@ -93,20 +89,20 @@ def serve(
 
     render(docs, style=style, page=page, minify=minify, options=options, manual=manual)
     httpd = simple_server.make_server(host, port, app)
-    print("\nUsing the '{}' visualizer".format(style))
+    print(f"\nUsing the '{style}' visualizer")
-    print("Serving on http://{}:{} ...\n".format(host, port))
+    print(f"Serving on http://{host}:{port} ...\n")
     try:
         httpd.serve_forever()
     except KeyboardInterrupt:
-        print("Shutting down server on port {}.".format(port))
+        print(f"Shutting down server on port {port}.")
     finally:
         httpd.server_close()
 
 
 def app(environ, start_response):
     # Headers and status need to be bytes in Python 2, see #1227
-    headers = [(b_to_str(b"Content-type"), b_to_str(b"text/html; charset=utf-8"))]
+    headers = [("Content-type", "text/html; charset=utf-8")]
-    start_response(b_to_str(b"200 OK"), headers)
+    start_response("200 OK", headers)
     res = _html["parsed"].encode(encoding="utf-8")
     return [res]
 

spacy/displacy/render.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 import uuid
 
 from .templates import TPL_DEP_SVG, TPL_DEP_WORDS, TPL_DEP_ARCS, TPL_ENTS
@@ -55,7 +52,7 @@ class DependencyRenderer(object):
                 settings = p.get("settings", {})
                 self.direction = settings.get("direction", DEFAULT_DIR)
                 self.lang = settings.get("lang", DEFAULT_LANG)
-            render_id = "{}-{}".format(id_prefix, i)
+            render_id = f"{id_prefix}-{i}"
             svg = self.render_svg(render_id, p["words"], p["arcs"])
             rendered.append(svg)
         if page:

spacy/displacy/templates.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 # Setting explicit height and max-width: none on the SVG is required for
 # Jupyter to render it properly in a cell
 

spacy/errors.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 import os
 import warnings
 import inspect
@@ -12,7 +9,7 @@ def add_codes(err_cls):
     class ErrorsWithCodes(object):
         def __getattribute__(self, code):
             msg = getattr(err_cls, code)
-            return "[{code}] {msg}".format(code=code, msg=msg)
+            return f"[{code}] {msg}"
 
     return ErrorsWithCodes()
 
@@ -97,8 +94,6 @@ class Warnings(object):
             "you can ignore this warning by setting SPACY_WARNING_IGNORE=W022. "
             "If this is surprising, make sure you have the spacy-lookups-data "
             "package installed.")
-    W023 = ("Multiprocessing of Language.pipe is not supported in Python 2. "
-            "'n_process' will be set to 1.")
     W024 = ("Entity '{entity}' - Alias '{alias}' combination already exists in "
             "the Knowledge Base.")
     W025 = ("'{name}' requires '{attr}' to be assigned, but none of the "
@@ -107,7 +102,9 @@ class Warnings(object):
     W027 = ("Found a large training file of {size} bytes. Note that it may "
             "be more efficient to split your training data into multiple "
             "smaller JSON files instead.")
-
+    W028 = ("Skipping unsupported morphological feature(s): {feature}. "
+            "Provide features as a dict {{\"Field1\": \"Value1,Value2\"}} or "
+            "string \"Field1=Value1,Value2|Field2=Value3\".")
 
 
 @add_codes
@@ -227,13 +224,8 @@ class Errors(object):
     E047 = ("Can't assign a value to unregistered extension attribute "
             "'{name}'. Did you forget to call the `set_extension` method?")
     E048 = ("Can't import language {lang} from spacy.lang: {err}")
-    E049 = ("Can't find spaCy data directory: '{path}'. Check your "
+    E050 = ("Can't find model '{name}'. It doesn't seem to be a Python "
-            "installation and permissions, or use spacy.util.set_data_path "
+            "package or a valid path to a data directory.")
-            "to customise the location if necessary.")
-    E050 = ("Can't find model '{name}'. It doesn't seem to be a shortcut "
-            "link, a Python package or a valid path to a data directory.")
-    E051 = ("Cant' load '{name}'. If you're using a shortcut link, make sure "
-            "it points to a valid package (not just a data directory).")
     E052 = ("Can't find model directory: {path}")
     E053 = ("Could not read meta.json from {path}")
     E054 = ("No valid '{setting}' setting found in model meta.json.")
@@ -424,8 +416,6 @@ class Errors(object):
     E134 = ("Entity '{entity}' is not defined in the Knowledge Base.")
     E135 = ("If you meant to replace a built-in component, use `create_pipe`: "
             "`nlp.replace_pipe('{name}', nlp.create_pipe('{name}'))`")
-    E136 = ("This additional feature requires the jsonschema library to be "
-            "installed:\npip install jsonschema")
     E137 = ("Expected 'dict' type, but got '{type}' from '{line}'. Make sure "
             "to provide a valid JSON object as input with either the `text` "
             "or `tokens` key. For more info, see the docs:\n"
@@ -541,6 +531,15 @@ class Errors(object):
     E188 = ("Could not match the gold entity links to entities in the doc - "
             "make sure the gold EL data refers to valid results of the "
             "named entity recognizer in the `nlp` pipeline.")
+    # TODO: fix numbering after merging develop into master
+    E996 = ("Could not parse {file}: {msg}")
+    E997 = ("Tokenizer special cases are not allowed to modify the text. "
+            "This would map '{chunk}' to '{orth}' given token attributes "
+            "'{token_attrs}'.")
+    E998 = ("Can only create GoldParse objects from Example objects without a "
+            "Doc if get_gold_parses() is called with a Vocab object.")
+    E999 = ("Encountered an unexpected format for the dictionary holding "
+            "gold annotations: {gold_dict}")
 
 
 @add_codes
@@ -566,10 +565,10 @@ class MatchPatternError(ValueError):
         errors (dict): Validation errors (sequence of strings) mapped to pattern
             ID, i.e. the index of the added pattern.
         """
-        msg = "Invalid token patterns for matcher rule '{}'\n".format(key)
+        msg = f"Invalid token patterns for matcher rule '{key}'\n"
         for pattern_idx, error_msgs in errors.items():
-            pattern_errors = "\n".join(["- {}".format(e) for e in error_msgs])
+            pattern_errors = "\n".join([f"- {e}" for e in error_msgs])
-            msg += "\nPattern {}:\n{}\n".format(pattern_idx, pattern_errors)
+            msg += f"\nPattern {pattern_idx}:\n{pattern_errors}\n"
         ValueError.__init__(self, msg)
 
 

spacy/glossary.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 def explain(term):
     """Get a description for a given POS tag, dependency label or entity type.
 

spacy/gold.pxd

@@ -1,6 +1,6 @@
 from cymem.cymem cimport Pool
 
-from .structs cimport TokenC
+from .tokens import Doc
 from .typedefs cimport attr_t
 from .syntax.transition_system cimport Transition
 
@@ -19,23 +19,49 @@ cdef class GoldParse:
     cdef Pool mem
 
     cdef GoldParseC c
+    cdef readonly TokenAnnotation orig
 
     cdef int length
     cdef public int loss
     cdef public list words
     cdef public list tags
-    cdef public list morphology
+    cdef public list pos
+    cdef public list morphs
+    cdef public list lemmas
+    cdef public list sent_starts
     cdef public list heads
     cdef public list labels
     cdef public dict orths
     cdef public list ner
-    cdef public list ents
     cdef public dict brackets
-    cdef public object cats
+    cdef public dict cats
     cdef public dict links
 
     cdef readonly list cand_to_gold
     cdef readonly list gold_to_cand
-    cdef readonly list orig_annot
 
 
+cdef class TokenAnnotation:
+    cdef public list ids
+    cdef public list words
+    cdef public list tags
+    cdef public list pos
+    cdef public list morphs
+    cdef public list lemmas
+    cdef public list heads
+    cdef public list deps
+    cdef public list entities
+    cdef public list sent_starts
+    cdef public list brackets
+
+
+cdef class DocAnnotation:
+    cdef public object cats
+    cdef public object links
+
+
+cdef class Example:
+    cdef public object doc
+    cdef public TokenAnnotation token_annotation
+    cdef public DocAnnotation doc_annotation
+    cdef public object goldparse

spacy/kb.pxd

@@ -6,7 +6,7 @@ from libcpp.vector cimport vector
 from libc.stdint cimport int32_t, int64_t
 from libc.stdio cimport FILE
 
-from spacy.vocab cimport Vocab
+from .vocab cimport Vocab
 from .typedefs cimport hash_t
 
 from .structs cimport KBEntryC, AliasC
@@ -169,4 +169,3 @@ cdef class Reader:
     cdef int read_alias(self, int64_t* entry_index, float* prob) except -1
 
     cdef int _read(self, void* value, size_t size) except -1
-

spacy/kb.pyx

@@ -1,22 +1,17 @@
 # cython: infer_types=True
 # cython: profile=True
-# coding: utf8
-from spacy.errors import Errors, Warnings, user_warning
-
 from pathlib import Path
 from cymem.cymem cimport Pool
 from preshed.maps cimport PreshMap
-
 from cpython.exc cimport PyErr_SetFromErrno
-
 from libc.stdio cimport fopen, fclose, fread, fwrite, feof, fseek
 from libc.stdint cimport int32_t, int64_t
-
-from .typedefs cimport hash_t
-
 from os import path
 from libcpp.vector cimport vector
 
+from .typedefs cimport hash_t
+from .errors import Errors, Warnings, user_warning
+
 
 cdef class Candidate:
     """A `Candidate` object refers to a textual mention (`alias`) that may or may not be resolved
@@ -447,7 +442,7 @@ cdef class KnowledgeBase:
 cdef class Writer:
     def __init__(self, object loc):
         if path.exists(loc):
-            assert not path.isdir(loc), "%s is directory." % loc
+            assert not path.isdir(loc), f"{loc} is directory"
         if isinstance(loc, Path):
             loc = bytes(loc)
         cdef bytes bytes_loc = loc.encode('utf8') if type(loc) == unicode else loc
@@ -584,5 +579,3 @@ cdef class Reader:
     cdef int _read(self, void* value, size_t size) except -1:
         status = fread(value, size, 1, self._fp)
         return status
-
-

spacy/lang/af/__init__.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from .stop_words import STOP_WORDS
 from ...language import Language
 from ...attrs import LANG

spacy/lang/af/stop_words.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 # Source: https://github.com/stopwords-iso/stopwords-af
 
 STOP_WORDS = set(

spacy/lang/ar/__init__.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from .stop_words import STOP_WORDS
 from .lex_attrs import LEX_ATTRS
 from .punctuation import TOKENIZER_SUFFIXES

spacy/lang/ar/examples.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 """
 Example sentences to test spaCy and its language models.
 

spacy/lang/ar/lex_attrs.py

@@ -1,5 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
 from ...attrs import LIKE_NUM
 
 _num_words = set(

spacy/lang/ar/punctuation.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ..char_classes import LIST_PUNCT, LIST_ELLIPSES, LIST_QUOTES, CURRENCY
 from ..char_classes import UNITS, ALPHA_UPPER
 

spacy/lang/ar/stop_words.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 STOP_WORDS = set(
     """
 من

spacy/lang/ar/tokenizer_exceptions.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ...symbols import ORTH, LEMMA
 
 

spacy/lang/bg/__init__.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from .stop_words import STOP_WORDS
 from ...language import Language
 from ...attrs import LANG

spacy/lang/bg/examples.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 """
 Example sentences to test spaCy and its language models.
 

spacy/lang/bg/stop_words.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 # Source: https://github.com/Alir3z4/stop-words
 
 STOP_WORDS = set(

spacy/lang/bn/__init__.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
 from .punctuation import TOKENIZER_PREFIXES, TOKENIZER_SUFFIXES, TOKENIZER_INFIXES
 from .tag_map import TAG_MAP

spacy/lang/bn/examples.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 """
 Example sentences to test spaCy and its language models.
 

spacy/lang/bn/morph_rules.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ...symbols import LEMMA, PRON_LEMMA
 
 

spacy/lang/bn/punctuation.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ..char_classes import LIST_PUNCT, LIST_ELLIPSES, LIST_QUOTES, LIST_ICONS
 from ..char_classes import ALPHA_LOWER, ALPHA, HYPHENS, CONCAT_QUOTES, UNITS
 

spacy/lang/bn/stop_words.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 STOP_WORDS = set(
     """
 অতএব অথচ অথবা অনুযায়ী অনেক অনেকে অনেকেই অন্তত  অবধি অবশ্য অর্থাৎ অন্য অনুযায়ী অর্ধভাগে

spacy/lang/bn/tag_map.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ...symbols import POS, PUNCT, ADJ, CONJ, SCONJ, NUM, DET, ADV, ADP, X, VERB
 from ...symbols import CCONJ, NOUN, PROPN, PART, INTJ, SPACE, PRON, AUX, SYM
 
@@ -14,8 +11,8 @@ TAG_MAP = {
     '""': {POS: PUNCT, "PunctType": "quot", "PunctSide": "fin"},
     "''": {POS: PUNCT, "PunctType": "quot", "PunctSide": "fin"},
     ":": {POS: PUNCT},
-    "৳": {POS: SYM, "Other": {"SymType": "currency"}},
+    "৳": {POS: SYM, "SymType": "currency"},
-    "#": {POS: SYM, "Other": {"SymType": "numbersign"}},
+    "#": {POS: SYM, "SymType": "numbersign"},
     "AFX": {POS: ADJ, "Hyph": "yes"},
     "CC": {POS: CONJ, "ConjType": "coor"},
     "CD": {POS: NUM, "NumType": "card"},

spacy/lang/bn/tokenizer_exceptions.py

@@ -1,6 +1,3 @@
-# coding=utf-8
-from __future__ import unicode_literals
-
 from ...symbols import ORTH, LEMMA
 
 

spacy/lang/ca/__init__.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
 from .stop_words import STOP_WORDS
 from .lex_attrs import LEX_ATTRS

spacy/lang/ca/examples.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 """
 Example sentences to test spaCy and its language models.
 

spacy/lang/ca/lex_attrs.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ...attrs import LIKE_NUM
 
 

spacy/lang/ca/punctuation.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ..punctuation import TOKENIZER_INFIXES
 from ..char_classes import ALPHA
 

spacy/lang/ca/stop_words.py

@@ -1,7 +1,3 @@
-# encoding: utf8
-from __future__ import unicode_literals
-
-
 STOP_WORDS = set(
     """
 a abans ací ah així això al aleshores algun alguna algunes alguns alhora allà allí allò

spacy/lang/ca/tag_map.py

@@ -1,28 +0,0 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-from ..symbols import POS, ADV, NOUN, ADP, PRON, SCONJ, PROPN, DET, SYM, INTJ
-from ..symbols import PUNCT, NUM, AUX, X, CONJ, ADJ, VERB, PART, SPACE, CCONJ
-
-
-TAG_MAP = {
-    "ADV": {POS: ADV},
-    "NOUN": {POS: NOUN},
-    "ADP": {POS: ADP},
-    "PRON": {POS: PRON},
-    "SCONJ": {POS: SCONJ},
-    "PROPN": {POS: PROPN},
-    "DET": {POS: DET},
-    "SYM": {POS: SYM},
-    "INTJ": {POS: INTJ},
-    "PUNCT": {POS: PUNCT},
-    "NUM": {POS: NUM},
-    "AUX": {POS: AUX},
-    "X": {POS: X},
-    "CONJ": {POS: CONJ},
-    "CCONJ": {POS: CCONJ},
-    "ADJ": {POS: ADJ},
-    "VERB": {POS: VERB},
-    "PART": {POS: PART},
-    "SP": {POS: SPACE},
-}

spacy/lang/ca/tokenizer_exceptions.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from ...symbols import ORTH, LEMMA
 
 
@@ -33,9 +30,9 @@ _exc["12m."] = [{ORTH: "12"}, {ORTH: "m.", LEMMA: "p.m."}]
 
 for h in range(1, 12 + 1):
     for period in ["a.m.", "am"]:
-        _exc["%d%s" % (h, period)] = [{ORTH: "%d" % h}, {ORTH: period, LEMMA: "a.m."}]
+        _exc[f"{h}{period}"] = [{ORTH: f"{h}"}, {ORTH: period, LEMMA: "a.m."}]
     for period in ["p.m.", "pm"]:
-        _exc["%d%s" % (h, period)] = [{ORTH: "%d" % h}, {ORTH: period, LEMMA: "p.m."}]
+        _exc[f"{h}{period}"] = [{ORTH: f"{h}"}, {ORTH: period, LEMMA: "p.m."}]
 
 
 TOKENIZER_EXCEPTIONS = _exc

spacy/lang/char_classes.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 split_chars = lambda char: list(char.strip().split(" "))
 merge_chars = lambda char: char.strip().replace(" ", "|")
 group_chars = lambda char: char.strip().replace(" ", "")

spacy/lang/cs/__init__.py

@@ -1,6 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from .stop_words import STOP_WORDS
 from ...language import Language
 from ...attrs import LANG

spacy/lang/cs/stop_words.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 # Source: https://github.com/Alir3z4/stop-words
 
 STOP_WORDS = set(

spacy/lang/da/__init__.py

@@ -1,13 +1,9 @@
-# coding: utf8
-from __future__ import unicode_literals
-
 from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
 from .norm_exceptions import NORM_EXCEPTIONS
 from .punctuation import TOKENIZER_INFIXES, TOKENIZER_SUFFIXES
 from .stop_words import STOP_WORDS
 from .lex_attrs import LEX_ATTRS
 from .morph_rules import MORPH_RULES
-from ..tag_map import TAG_MAP
 
 from ..tokenizer_exceptions import BASE_EXCEPTIONS
 from ..norm_exceptions import BASE_NORMS
@@ -27,7 +23,6 @@ class DanishDefaults(Language.Defaults):
     morph_rules = MORPH_RULES
     infixes = TOKENIZER_INFIXES
     suffixes = TOKENIZER_SUFFIXES
-    tag_map = TAG_MAP
     stop_words = STOP_WORDS
 
 

spacy/lang/da/examples.py

@@ -1,7 +1,3 @@
-# coding: utf8
-from __future__ import unicode_literals
-
-
 """
 Example sentences to test spaCy and its language models.