Merge remote-tracking branch 'upstream/develop' into feature/small-fixes

Makefile

@@ -1,7 +1,7 @@
 SHELL := /bin/bash
 
 ifndef SPACY_EXTRAS
-override SPACY_EXTRAS = spacy-lookups-data==0.4.0.dev0 jieba pkuseg==0.0.25 sudachipy sudachidict_core
+override SPACY_EXTRAS = spacy-lookups-data==1.0.0rc0 jieba pkuseg==0.0.25 pickle5 sudachipy sudachidict_core
 endif
 
 ifndef PYVER

pyproject.toml

@@ -6,7 +6,7 @@ requires = [
     "cymem>=2.0.2,<2.1.0",
     "preshed>=3.0.2,<3.1.0",
     "murmurhash>=0.28.0,<1.1.0",
-    "thinc>=8.0.0a41,<8.0.0a50",
+    "thinc>=8.0.0a43,<8.0.0a50",
     "blis>=0.4.0,<0.5.0",
     "pytokenizations",
     "pathy"

requirements.txt

@@ -1,12 +1,12 @@
 # Our libraries
 cymem>=2.0.2,<2.1.0
 preshed>=3.0.2,<3.1.0
-thinc>=8.0.0a41,<8.0.0a50
+thinc>=8.0.0a43,<8.0.0a50
 blis>=0.4.0,<0.5.0
 ml_datasets==0.2.0a0
 murmurhash>=0.28.0,<1.1.0
 wasabi>=0.8.0,<1.1.0
-srsly>=2.1.0,<3.0.0
+srsly>=2.3.0,<3.0.0
 catalogue>=2.0.1,<2.1.0
 typer>=0.3.0,<0.4.0
 pathy

setup.cfg

@@ -34,16 +34,16 @@ setup_requires =
     cymem>=2.0.2,<2.1.0
     preshed>=3.0.2,<3.1.0
     murmurhash>=0.28.0,<1.1.0
-    thinc>=8.0.0a41,<8.0.0a50
+    thinc>=8.0.0a43,<8.0.0a50
 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>=8.0.0a41,<8.0.0a50
+    thinc>=8.0.0a43,<8.0.0a50
     blis>=0.4.0,<0.5.0
     wasabi>=0.8.0,<1.1.0
-    srsly>=2.1.0,<3.0.0
+    srsly>=2.3.0,<3.0.0
     catalogue>=2.0.1,<2.1.0
     typer>=0.3.0,<0.4.0
     pathy
@@ -65,7 +65,7 @@ console_scripts =
 
 [options.extras_require]
 lookups =
-    spacy_lookups_data==0.4.0.dev0
+    spacy_lookups_data==1.0.0rc0
 cuda =
     cupy>=5.0.0b4,<9.0.0
 cuda80 =
@@ -84,7 +84,7 @@ cuda102 =
     cupy-cuda102>=5.0.0b4,<9.0.0
 # Language tokenizers with external dependencies
 ja =
-    sudachipy>=0.4.5
+    sudachipy>=0.4.9
     sudachidict_core>=20200330
 ko =
     natto-py==0.9.0
@@ -98,7 +98,7 @@ universal = false
 formats = gztar
 
 [flake8]
-ignore = E203, E266, E501, E731, W503
+ignore = E203, E266, E501, E731, W503, E741
 max-line-length = 80
 select = B,C,E,F,W,T4,B9
 exclude =

spacy/about.py

@@ -1,6 +1,6 @@
 # fmt: off
 __title__ = "spacy-nightly"
-__version__ = "3.0.0a26"
+__version__ = "3.0.0a29"
 __release__ = True
 __download_url__ = "https://github.com/explosion/spacy-models/releases/download"
 __compatibility__ = "https://raw.githubusercontent.com/explosion/spacy-models/master/compatibility.json"

spacy/cli/__init__.py

@@ -15,7 +15,7 @@ from .debug_config import debug_config  # noqa: F401
 from .debug_model import debug_model  # noqa: F401
 from .evaluate import evaluate  # noqa: F401
 from .convert import convert  # noqa: F401
-from .init_model import init_model  # noqa: F401
+from .init_pipeline import init_pipeline_cli  # noqa: F401
 from .init_config import init_config, fill_config  # noqa: F401
 from .validate import validate  # noqa: F401
 from .project.clone import project_clone  # noqa: F401

spacy/cli/_util.py

@@ -10,12 +10,13 @@ from click import NoSuchOption
 from click.parser import split_arg_string
 from typer.main import get_command
 from contextlib import contextmanager
-from thinc.api import Config, ConfigValidationError
+from thinc.api import Config, ConfigValidationError, require_gpu
 from configparser import InterpolationError
 import os
 
 from ..schemas import ProjectConfigSchema, validate
 from ..util import import_file, run_command, make_tempdir, registry, logger
+from ..util import ENV_VARS
 
 if TYPE_CHECKING:
     from pathy import Pathy  # noqa: F401
@@ -39,7 +40,6 @@ commands to check and validate your config files, training and evaluation data,
 and custom model implementations.
 """
 INIT_HELP = """Commands for initializing configs and pipeline packages."""
-OVERRIDES_ENV_VAR = "SPACY_CONFIG_OVERRIDES"
 
 # Wrappers for Typer's annotations. Initially created to set defaults and to
 # keep the names short, but not needed at the moment.
@@ -65,7 +65,7 @@ def setup_cli() -> None:
 
 
 def parse_config_overrides(
-    args: List[str], env_var: Optional[str] = OVERRIDES_ENV_VAR
+    args: List[str], env_var: Optional[str] = ENV_VARS.CONFIG_OVERRIDES
 ) -> Dict[str, Any]:
     """Generate a dictionary of config overrides based on the extra arguments
     provided on the CLI, e.g. --training.batch_size to override
@@ -275,18 +275,6 @@ def import_code(code_path: Optional[Union[Path, str]]) -> None:
             msg.fail(f"Couldn't load Python code: {code_path}", e, exits=1)
 
 
-def get_sourced_components(config: Union[Dict[str, Any], Config]) -> List[str]:
-    """RETURNS (List[str]): All sourced components in the original config,
-        e.g. {"source": "en_core_web_sm"}. If the config contains a key
-        "factory", we assume it refers to a component factory.
-    """
-    return [
-        name
-        for name, cfg in config.get("components", {}).items()
-        if "factory" not in cfg and "source" in cfg
-    ]
-
-
 def upload_file(src: Path, dest: Union[str, "Pathy"]) -> None:
     """Upload a file.
 
@@ -458,3 +446,12 @@ def string_to_list(value: str, intify: bool = False) -> Union[List[str], List[in
             p = int(p)
         result.append(p)
     return result
+
+
+def setup_gpu(use_gpu: int) -> None:
+    """Configure the GPU and log info."""
+    if use_gpu >= 0:
+        msg.info(f"Using GPU: {use_gpu}")
+        require_gpu(use_gpu)
+    else:
+        msg.info("Using CPU")

spacy/cli/convert.py

@@ -9,7 +9,8 @@ import sys
 from ._util import app, Arg, Opt
 from ..training import docs_to_json
 from ..tokens import DocBin
-from ..training.converters import iob_to_docs, conll_ner_to_docs, json_to_docs, conllu_to_docs
+from ..training.converters import iob_to_docs, conll_ner_to_docs, json_to_docs
+from ..training.converters import conllu_to_docs
 
 
 # Converters are matched by file extension except for ner/iob, which are

spacy/cli/debug_config.py

@@ -1,12 +1,14 @@
 from typing import Optional, Dict, Any, Union, List
 from pathlib import Path
 from wasabi import msg, table
-from thinc.api import Config, ConfigValidationError
+from thinc.api import Config
 from thinc.config import VARIABLE_RE
 import typer
 
 from ._util import Arg, Opt, show_validation_error, parse_config_overrides
 from ._util import import_code, debug_cli
+from ..schemas import ConfigSchemaTraining
+from ..util import registry
 from .. import util
 
 
@@ -52,10 +54,10 @@ def debug_config(
     with show_validation_error(config_path):
         config = util.load_config(config_path, overrides=overrides)
         nlp = util.load_model_from_config(config)
-        # Use the resolved config here in case user has one function returning
+        config = nlp.config.interpolate()
-        # a dict of corpora etc.
+        T = registry.resolve(config["training"], schema=ConfigSchemaTraining)
-        resolved = util.resolve_training_config(nlp.config)
+        dot_names = [T["train_corpus"], T["dev_corpus"]]
-        check_section_refs(resolved, ["training.dev_corpus", "training.train_corpus"])
+        util.resolve_dot_names(config, dot_names)
     msg.good("Config is valid")
     if show_vars:
         variables = get_variables(config)
@@ -97,23 +99,3 @@ def get_variables(config: Config) -> Dict[str, Any]:
         value = util.dot_to_object(config, path)
         result[variable] = repr(value)
     return result
-
-
-def check_section_refs(config: Config, fields: List[str]) -> None:
-    """Validate fields in the config that refer to other sections or values
-    (e.g. in the corpora) and make sure that those references exist.
-    """
-    errors = []
-    for field in fields:
-        # If the field doesn't exist in the config, we ignore it
-        try:
-            value = util.dot_to_object(config, field)
-        except KeyError:
-            continue
-        try:
-            util.dot_to_object(config, value)
-        except KeyError:
-            msg = f"not a valid section reference: {value}"
-            errors.append({"loc": field.split("."), "msg": msg})
-    if errors:
-        raise ConfigValidationError(config=config, errors=errors)

spacy/cli/debug_data.py

@@ -7,10 +7,13 @@ from wasabi import Printer, MESSAGES, msg
 import typer
 
 from ._util import app, Arg, Opt, show_validation_error, parse_config_overrides
-from ._util import import_code, debug_cli, get_sourced_components
+from ._util import import_code, debug_cli
-from ..training import Corpus, Example
+from ..training import Example
+from ..training.initialize import get_sourced_components
+from ..schemas import ConfigSchemaTraining
 from ..pipeline._parser_internals import nonproj
 from ..language import Language
+from ..util import registry, resolve_dot_names
 from .. import util
 
 
@@ -24,7 +27,7 @@ BLANK_MODEL_THRESHOLD = 2000
 
 
 @debug_cli.command(
-    "data", context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+    "data", context_settings={"allow_extra_args": True, "ignore_unknown_options": True}
 )
 @app.command(
     "debug-data",
@@ -34,8 +37,6 @@ BLANK_MODEL_THRESHOLD = 2000
 def debug_data_cli(
     # fmt: off
     ctx: typer.Context,  # This is only used to read additional arguments
-    train_path: Path = Arg(..., help="Location of JSON-formatted training data", exists=True),
-    dev_path: Path = Arg(..., help="Location of JSON-formatted development data", exists=True),
     config_path: Path = Arg(..., help="Path to config file", exists=True),
     code_path: Optional[Path] = Opt(None, "--code-path", "-c", help="Path to Python file with additional code (registered functions) to be imported"),
     ignore_warnings: bool = Opt(False, "--ignore-warnings", "-IW", help="Ignore warnings, only show stats and errors"),
@@ -59,8 +60,6 @@ def debug_data_cli(
     overrides = parse_config_overrides(ctx.args)
     import_code(code_path)
     debug_data(
-        train_path,
-        dev_path,
         config_path,
         config_overrides=overrides,
         ignore_warnings=ignore_warnings,
@@ -71,8 +70,6 @@ def debug_data_cli(
 
 
 def debug_data(
-    train_path: Path,
-    dev_path: Path,
     config_path: Path,
     *,
     config_overrides: Dict[str, Any] = {},
@@ -85,57 +82,29 @@ def debug_data(
         no_print=silent, pretty=not no_format, ignore_warnings=ignore_warnings
     )
     # Make sure all files and paths exists if they are needed
-    if not train_path.exists():
-        msg.fail("Training data not found", train_path, exits=1)
-    if not dev_path.exists():
-        msg.fail("Development data not found", dev_path, exits=1)
-    if not config_path.exists():
-        msg.fail("Config file not found", config_path, exists=1)
     with show_validation_error(config_path):
         cfg = util.load_config(config_path, overrides=config_overrides)
         nlp = util.load_model_from_config(cfg)
-        C = util.resolve_training_config(nlp.config)
+        config = nlp.config.interpolate()
+        T = registry.resolve(config["training"], schema=ConfigSchemaTraining)
     # Use original config here, not resolved version
     sourced_components = get_sourced_components(cfg)
-    frozen_components = C["training"]["frozen_components"]
+    frozen_components = T["frozen_components"]
     resume_components = [p for p in sourced_components if p not in frozen_components]
     pipeline = nlp.pipe_names
     factory_names = [nlp.get_pipe_meta(pipe).factory for pipe in nlp.pipe_names]
-    tag_map_path = util.ensure_path(C["training"]["tag_map"])
-    tag_map = {}
-    if tag_map_path is not None:
-        tag_map = srsly.read_json(tag_map_path)
-    morph_rules_path = util.ensure_path(C["training"]["morph_rules"])
-    morph_rules = {}
-    if morph_rules_path is not None:
-        morph_rules = srsly.read_json(morph_rules_path)
-    # Replace tag map with provided mapping
-    nlp.vocab.morphology.load_tag_map(tag_map)
-    # Load morph rules
-    nlp.vocab.morphology.load_morph_exceptions(morph_rules)
-
     msg.divider("Data file validation")
 
     # Create the gold corpus to be able to better analyze data
-    loading_train_error_message = ""
+    dot_names = [T["train_corpus"], T["dev_corpus"]]
-    loading_dev_error_message = ""
+    train_corpus, dev_corpus = resolve_dot_names(config, dot_names)
-    with msg.loading("Loading corpus..."):
+    train_dataset = list(train_corpus(nlp))
-        try:
+    dev_dataset = list(dev_corpus(nlp))
-            train_dataset = list(Corpus(train_path)(nlp))
-        except ValueError as e:
-            loading_train_error_message = f"Training data cannot be loaded: {e}"
-        try:
-            dev_dataset = list(Corpus(dev_path)(nlp))
-        except ValueError as e:
-            loading_dev_error_message = f"Development data cannot be loaded: {e}"
-    if loading_train_error_message or loading_dev_error_message:
-        if loading_train_error_message:
-            msg.fail(loading_train_error_message)
-        if loading_dev_error_message:
-            msg.fail(loading_dev_error_message)
-        sys.exit(1)
     msg.good("Corpus is loadable")
 
+    nlp.initialize(lambda: train_dataset)
+    msg.good("Pipeline can be initialized with data")
+
     # Create all gold data here to avoid iterating over the train_dataset constantly
     gold_train_data = _compile_gold(train_dataset, factory_names, nlp, make_proj=True)
     gold_train_unpreprocessed_data = _compile_gold(
@@ -145,10 +114,10 @@ def debug_data(
 
     train_texts = gold_train_data["texts"]
     dev_texts = gold_dev_data["texts"]
-    frozen_components = C["training"]["frozen_components"]
+    frozen_components = T["frozen_components"]
 
     msg.divider("Training stats")
-    msg.text(f"Language: {C['nlp']['lang']}")
+    msg.text(f"Language: {nlp.lang}")
     msg.text(f"Training pipeline: {', '.join(pipeline)}")
     if resume_components:
         msg.text(f"Components from other pipelines: {', '.join(resume_components)}")
@@ -355,17 +324,12 @@ def debug_data(
     if "tagger" in factory_names:
         msg.divider("Part-of-speech Tagging")
         labels = [label for label in gold_train_data["tags"]]
-        tag_map = nlp.vocab.morphology.tag_map
+        # TODO: does this need to be updated?
-        msg.info(f"{len(labels)} label(s) in data ({len(tag_map)} label(s) in tag map)")
+        msg.info(f"{len(labels)} label(s) in data")
         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(f"All labels present in tag map for language '{nlp.lang}'")
-        for label in non_tagmap:
-            msg.fail(f"Label '{label}' not found in tag map for language '{nlp.lang}'")
 
     if "parser" in factory_names:
         has_low_data_warning = False

spacy/cli/debug_model.py

@@ -2,18 +2,23 @@ from typing import Dict, Any, Optional, Iterable
 from pathlib import Path
 
 from spacy.training import Example
-from spacy.util import dot_to_object
+from spacy.util import resolve_dot_names
 from wasabi import msg
-from thinc.api import require_gpu, fix_random_seed, set_dropout_rate, Adam
+from thinc.api import fix_random_seed, set_dropout_rate, Adam
 from thinc.api import Model, data_validation, set_gpu_allocator
 import typer
 
 from ._util import Arg, Opt, debug_cli, show_validation_error
-from ._util import parse_config_overrides, string_to_list
+from ._util import parse_config_overrides, string_to_list, setup_gpu
+from ..schemas import ConfigSchemaTraining
+from ..util import registry
 from .. import util
 
 
-@debug_cli.command("model")
+@debug_cli.command(
+    "model",
+    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+)
 def debug_model_cli(
     # fmt: off
     ctx: typer.Context,  # This is only used to read additional arguments
@@ -37,11 +42,7 @@ def debug_model_cli(
 
     DOCS: https://nightly.spacy.io/api/cli#debug-model
     """
-    if use_gpu >= 0:
+    setup_gpu(use_gpu)
-        msg.info("Using GPU")
-        require_gpu(use_gpu)
-    else:
-        msg.info("Using CPU")
     layers = string_to_list(layers, intify=True)
     print_settings = {
         "dimensions": dimensions,
@@ -59,14 +60,15 @@ def debug_model_cli(
         raw_config = util.load_config(
             config_path, overrides=config_overrides, interpolate=False
         )
-    config = raw_config.iterpolate()
+    config = raw_config.interpolate()
     allocator = config["training"]["gpu_allocator"]
     if use_gpu >= 0 and allocator:
         set_gpu_allocator(allocator)
     with show_validation_error(config_path):
         nlp = util.load_model_from_config(raw_config)
-        C = util.resolve_training_config(nlp.config)
+        config = nlp.config.interpolate()
-    seed = C["training"]["seed"]
+        T = registry.resolve(config["training"], schema=ConfigSchemaTraining)
+    seed = T["seed"]
     if seed is not None:
         msg.info(f"Fixing random seed: {seed}")
         fix_random_seed(seed)
@@ -77,11 +79,16 @@ def debug_model_cli(
             exits=1,
         )
     model = pipe.model
-    debug_model(C, nlp, model, print_settings=print_settings)
+    debug_model(config, T, nlp, model, print_settings=print_settings)
 
 
 def debug_model(
-    config, nlp, model: Model, *, print_settings: Optional[Dict[str, Any]] = None
+    config,
+    resolved_train_config,
+    nlp,
+    model: Model,
+    *,
+    print_settings: Optional[Dict[str, Any]] = None,
 ):
     if not isinstance(model, Model):
         msg.fail(
@@ -102,13 +109,16 @@ def debug_model(
     # The output vector might differ from the official type of the output layer
     with data_validation(False):
         try:
-            train_corpus = dot_to_object(config, config["training"]["train_corpus"])
+            dot_names = [resolved_train_config["train_corpus"]]
-            nlp.begin_training(lambda: train_corpus(nlp))
+            with show_validation_error():
+                (train_corpus,) = resolve_dot_names(config, dot_names)
+                nlp.initialize(lambda: train_corpus(nlp))
             msg.info("Initialized the model with the training corpus.")
         except ValueError:
             try:
                 _set_output_dim(nO=7, model=model)
-                nlp.begin_training(lambda: [Example.from_dict(x, {}) for x in X])
+                with show_validation_error():
+                    nlp.initialize(lambda: [Example.from_dict(x, {}) for x in X])
                 msg.info("Initialized the model with dummy data.")
             except Exception:
                 msg.fail(

spacy/cli/evaluate.py

@@ -3,11 +3,11 @@ from wasabi import Printer
 from pathlib import Path
 import re
 import srsly
-from thinc.api import require_gpu, fix_random_seed
+from thinc.api import fix_random_seed
 
 from ..training import Corpus
 from ..tokens import Doc
-from ._util import app, Arg, Opt
+from ._util import app, Arg, Opt, setup_gpu, import_code
 from ..scorer import Scorer
 from .. import util
 from .. import displacy
@@ -19,6 +19,7 @@ def evaluate_cli(
     model: str = Arg(..., help="Model name or path"),
     data_path: Path = Arg(..., help="Location of binary evaluation data in .spacy format", exists=True),
     output: Optional[Path] = Opt(None, "--output", "-o", help="Output JSON file for metrics", dir_okay=False),
+    code_path: Optional[Path] = Opt(None, "--code", "-c", help="Path to Python file with additional code (registered functions) to be imported"),
     use_gpu: int = Opt(-1, "--gpu-id", "-g", help="GPU ID or -1 for CPU"),
     gold_preproc: bool = Opt(False, "--gold-preproc", "-G", help="Use gold preprocessing"),
     displacy_path: Optional[Path] = Opt(None, "--displacy-path", "-dp", help="Directory to output rendered parses as HTML", exists=True, file_okay=False),
@@ -37,6 +38,7 @@ def evaluate_cli(
 
     DOCS: https://nightly.spacy.io/api/cli#evaluate
     """
+    import_code(code_path)
     evaluate(
         model,
         data_path,
@@ -61,8 +63,7 @@ def evaluate(
 ) -> Scorer:
     msg = Printer(no_print=silent, pretty=not silent)
     fix_random_seed()
-    if use_gpu >= 0:
+    setup_gpu(use_gpu)
-        require_gpu(use_gpu)
     data_path = util.ensure_path(data_path)
     output_path = util.ensure_path(output)
     displacy_path = util.ensure_path(displacy_path)

spacy/cli/init_model.py

@@ -1,360 +0,0 @@
-from typing import Optional, List, Dict, Any, Union, IO
-import math
-from tqdm import tqdm
-import numpy
-from ast import literal_eval
-from pathlib import Path
-from preshed.counter import PreshCounter
-import tarfile
-import gzip
-import zipfile
-import srsly
-import warnings
-from wasabi import msg, Printer
-import typer
-
-from ._util import app, init_cli, Arg, Opt
-from ..vectors import Vectors
-from ..errors import Errors, Warnings
-from ..language import Language
-from ..util import ensure_path, get_lang_class, load_model, OOV_RANK
-
-try:
-    import ftfy
-except ImportError:
-    ftfy = None
-
-
-DEFAULT_OOV_PROB = -20
-
-
-@init_cli.command("vocab")
-@app.command(
-    "init-model",
-    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
-    hidden=True,  # hide this from main CLI help but still allow it to work with warning
-)
-def init_model_cli(
-    # fmt: off
-    ctx: typer.Context,  # This is only used to read additional arguments
-    lang: str = Arg(..., help="Pipeline language"),
-    output_dir: Path = Arg(..., help="Pipeline output directory"),
-    freqs_loc: Optional[Path] = Arg(None, help="Location of words frequencies file", exists=True),
-    clusters_loc: Optional[Path] = Opt(None, "--clusters-loc", "-c", help="Optional location of brown clusters data", exists=True),
-    jsonl_loc: Optional[Path] = Opt(None, "--jsonl-loc", "-j", help="Location of JSONL-formatted attributes file", exists=True),
-    vectors_loc: Optional[Path] = Opt(None, "--vectors-loc", "-v", help="Optional vectors file in Word2Vec format", exists=True),
-    prune_vectors: int = Opt(-1, "--prune-vectors", "-V", help="Optional number of vectors to prune to"),
-    truncate_vectors: int = Opt(0, "--truncate-vectors", "-t", help="Optional number of vectors to truncate to when reading in vectors file"),
-    vectors_name: Optional[str] = Opt(None, "--vectors-name", "-vn", help="Optional name for the word vectors, e.g. en_core_web_lg.vectors"),
-    model_name: Optional[str] = Opt(None, "--meta-name", "-mn", help="Optional name of the package for the pipeline meta"),
-    base_model: Optional[str] = Opt(None, "--base", "-b", help="Name of or path to base pipeline to start with (mostly relevant for pipelines with custom tokenizers)")
-    # fmt: on
-):
-    """
-    Create a new blank pipeline directory with vocab and vectors from raw data.
-    If vectors are provided in Word2Vec format, they can be either a .txt or
-    zipped as a .zip or .tar.gz.
-
-    DOCS: https://nightly.spacy.io/api/cli#init-vocab
-    """
-    if ctx.command.name == "init-model":
-        msg.warn(
-            "The init-model command is now called 'init vocab'. You can run "
-            "'python -m spacy init --help' for an overview of the other "
-            "available initialization commands."
-        )
-    init_model(
-        lang,
-        output_dir,
-        freqs_loc=freqs_loc,
-        clusters_loc=clusters_loc,
-        jsonl_loc=jsonl_loc,
-        vectors_loc=vectors_loc,
-        prune_vectors=prune_vectors,
-        truncate_vectors=truncate_vectors,
-        vectors_name=vectors_name,
-        model_name=model_name,
-        base_model=base_model,
-        silent=False,
-    )
-
-
-def init_model(
-    lang: str,
-    output_dir: Path,
-    freqs_loc: Optional[Path] = None,
-    clusters_loc: Optional[Path] = None,
-    jsonl_loc: Optional[Path] = None,
-    vectors_loc: Optional[Path] = None,
-    prune_vectors: int = -1,
-    truncate_vectors: int = 0,
-    vectors_name: Optional[str] = None,
-    model_name: Optional[str] = None,
-    base_model: Optional[str] = None,
-    silent: bool = True,
-) -> Language:
-    msg = Printer(no_print=silent, pretty=not silent)
-    if jsonl_loc is not None:
-        if freqs_loc is not None or clusters_loc is not None:
-            settings = ["-j"]
-            if freqs_loc:
-                settings.append("-f")
-            if clusters_loc:
-                settings.append("-c")
-            msg.warn(
-                "Incompatible arguments",
-                "The -f and -c arguments are deprecated, and not compatible "
-                "with the -j argument, which should specify the same "
-                "information. Either merge the frequencies and clusters data "
-                "into the JSONL-formatted file (recommended), or use only the "
-                "-f and -c files, without the other lexical attributes.",
-            )
-        jsonl_loc = ensure_path(jsonl_loc)
-        lex_attrs = srsly.read_jsonl(jsonl_loc)
-    else:
-        clusters_loc = ensure_path(clusters_loc)
-        freqs_loc = ensure_path(freqs_loc)
-        if freqs_loc is not None and not freqs_loc.exists():
-            msg.fail("Can't find words frequencies file", freqs_loc, exits=1)
-        lex_attrs = read_attrs_from_deprecated(msg, freqs_loc, clusters_loc)
-
-    with msg.loading("Creating blank pipeline..."):
-        nlp = create_model(lang, lex_attrs, name=model_name, base_model=base_model)
-
-    msg.good("Successfully created blank pipeline")
-    if vectors_loc is not None:
-        add_vectors(
-            msg, nlp, vectors_loc, truncate_vectors, prune_vectors, vectors_name
-        )
-    vec_added = len(nlp.vocab.vectors)
-    lex_added = len(nlp.vocab)
-    msg.good(
-        "Sucessfully compiled vocab", f"{lex_added} entries, {vec_added} vectors",
-    )
-    if not output_dir.exists():
-        output_dir.mkdir()
-    nlp.to_disk(output_dir)
-    return nlp
-
-
-def open_file(loc: Union[str, Path]) -> IO:
-    """Handle .gz, .tar.gz or unzipped files"""
-    loc = ensure_path(loc)
-    if tarfile.is_tarfile(str(loc)):
-        return tarfile.open(str(loc), "r:gz")
-    elif loc.parts[-1].endswith("gz"):
-        return (line.decode("utf8") for line in gzip.open(str(loc), "r"))
-    elif loc.parts[-1].endswith("zip"):
-        zip_file = zipfile.ZipFile(str(loc))
-        names = zip_file.namelist()
-        file_ = zip_file.open(names[0])
-        return (line.decode("utf8") for line in file_)
-    else:
-        return loc.open("r", encoding="utf8")
-
-
-def read_attrs_from_deprecated(
-    msg: Printer, freqs_loc: Optional[Path], clusters_loc: Optional[Path]
-) -> List[Dict[str, Any]]:
-    if freqs_loc is not None:
-        with msg.loading("Counting frequencies..."):
-            probs, _ = read_freqs(freqs_loc)
-        msg.good("Counted frequencies")
-    else:
-        probs, _ = ({}, DEFAULT_OOV_PROB)  # noqa: F841
-    if clusters_loc:
-        with msg.loading("Reading clusters..."):
-            clusters = read_clusters(clusters_loc)
-        msg.good("Read clusters")
-    else:
-        clusters = {}
-    lex_attrs = []
-    sorted_probs = sorted(probs.items(), key=lambda item: item[1], reverse=True)
-    if len(sorted_probs):
-        for i, (word, prob) in tqdm(enumerate(sorted_probs)):
-            attrs = {"orth": word, "id": i, "prob": prob}
-            # Decode as a little-endian string, so that we can do & 15 to get
-            # the first 4 bits. See _parse_features.pyx
-            if word in clusters:
-                attrs["cluster"] = int(clusters[word][::-1], 2)
-            else:
-                attrs["cluster"] = 0
-            lex_attrs.append(attrs)
-    return lex_attrs
-
-
-def create_model(
-    lang: str,
-    lex_attrs: List[Dict[str, Any]],
-    name: Optional[str] = None,
-    base_model: Optional[Union[str, Path]] = None,
-) -> Language:
-    if base_model:
-        nlp = load_model(base_model)
-        # keep the tokenizer but remove any existing pipeline components due to
-        # potentially conflicting vectors
-        for pipe in nlp.pipe_names:
-            nlp.remove_pipe(pipe)
-    else:
-        lang_class = get_lang_class(lang)
-        nlp = lang_class()
-    for lexeme in nlp.vocab:
-        lexeme.rank = OOV_RANK
-    for attrs in lex_attrs:
-        if "settings" in attrs:
-            continue
-        lexeme = nlp.vocab[attrs["orth"]]
-        lexeme.set_attrs(**attrs)
-    if len(nlp.vocab):
-        oov_prob = min(lex.prob for lex in nlp.vocab) - 1
-    else:
-        oov_prob = DEFAULT_OOV_PROB
-    nlp.vocab.cfg.update({"oov_prob": oov_prob})
-    if name:
-        nlp.meta["name"] = name
-    return nlp
-
-
-def add_vectors(
-    msg: Printer,
-    nlp: Language,
-    vectors_loc: Optional[Path],
-    truncate_vectors: int,
-    prune_vectors: int,
-    name: Optional[str] = None,
-) -> None:
-    vectors_loc = ensure_path(vectors_loc)
-    if vectors_loc and vectors_loc.parts[-1].endswith(".npz"):
-        nlp.vocab.vectors = Vectors(data=numpy.load(vectors_loc.open("rb")))
-        for lex in nlp.vocab:
-            if lex.rank and lex.rank != OOV_RANK:
-                nlp.vocab.vectors.add(lex.orth, row=lex.rank)
-    else:
-        if vectors_loc:
-            with msg.loading(f"Reading vectors from {vectors_loc}"):
-                vectors_data, vector_keys = read_vectors(
-                    msg, vectors_loc, truncate_vectors
-                )
-            msg.good(f"Loaded vectors from {vectors_loc}")
-        else:
-            vectors_data, vector_keys = (None, None)
-        if vector_keys is not None:
-            for word in vector_keys:
-                if word not in nlp.vocab:
-                    nlp.vocab[word]
-        if vectors_data is not None:
-            nlp.vocab.vectors = Vectors(data=vectors_data, keys=vector_keys)
-    if name is None:
-        # TODO: Is this correct? Does this matter?
-        nlp.vocab.vectors.name = f"{nlp.meta['lang']}_{nlp.meta['name']}.vectors"
-    else:
-        nlp.vocab.vectors.name = name
-    nlp.meta["vectors"]["name"] = nlp.vocab.vectors.name
-    if prune_vectors >= 1:
-        nlp.vocab.prune_vectors(prune_vectors)
-
-
-def read_vectors(msg: Printer, vectors_loc: Path, truncate_vectors: int):
-    f = open_file(vectors_loc)
-    f = ensure_shape(f)
-    shape = tuple(int(size) for size in next(f).split())
-    if truncate_vectors >= 1:
-        shape = (truncate_vectors, shape[1])
-    vectors_data = numpy.zeros(shape=shape, dtype="f")
-    vectors_keys = []
-    for i, line in enumerate(tqdm(f)):
-        line = line.rstrip()
-        pieces = line.rsplit(" ", vectors_data.shape[1])
-        word = pieces.pop(0)
-        if len(pieces) != vectors_data.shape[1]:
-            msg.fail(Errors.E094.format(line_num=i, loc=vectors_loc), exits=1)
-        vectors_data[i] = numpy.asarray(pieces, dtype="f")
-        vectors_keys.append(word)
-        if i == truncate_vectors - 1:
-            break
-    return vectors_data, vectors_keys
-
-
-def ensure_shape(lines):
-    """Ensure that the first line of the data is the vectors shape.
-
-    If it's not, we read in the data and output the shape as the first result,
-    so that the reader doesn't have to deal with the problem.
-    """
-    first_line = next(lines)
-    try:
-        shape = tuple(int(size) for size in first_line.split())
-    except ValueError:
-        shape = None
-    if shape is not None:
-        # All good, give the data
-        yield first_line
-        yield from lines
-    else:
-        # Figure out the shape, make it the first value, and then give the
-        # rest of the data.
-        width = len(first_line.split()) - 1
-        captured = [first_line] + list(lines)
-        length = len(captured)
-        yield f"{length} {width}"
-        yield from captured
-
-
-def read_freqs(
-    freqs_loc: Path, max_length: int = 100, min_doc_freq: int = 5, min_freq: int = 50
-):
-    counts = PreshCounter()
-    total = 0
-    with freqs_loc.open() as f:
-        for i, line in enumerate(f):
-            freq, doc_freq, key = line.rstrip().split("\t", 2)
-            freq = int(freq)
-            counts.inc(i + 1, freq)
-            total += freq
-    counts.smooth()
-    log_total = math.log(total)
-    probs = {}
-    with freqs_loc.open() as f:
-        for line in tqdm(f):
-            freq, doc_freq, key = line.rstrip().split("\t", 2)
-            doc_freq = int(doc_freq)
-            freq = int(freq)
-            if doc_freq >= min_doc_freq and freq >= min_freq and len(key) < max_length:
-                try:
-                    word = literal_eval(key)
-                except SyntaxError:
-                    # Take odd strings literally.
-                    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
-    return probs, oov_prob
-
-
-def read_clusters(clusters_loc: Path) -> dict:
-    clusters = {}
-    if ftfy is None:
-        warnings.warn(Warnings.W004)
-    with clusters_loc.open() as f:
-        for line in tqdm(f):
-            try:
-                cluster, word, freq = line.split()
-                if ftfy is not None:
-                    word = ftfy.fix_text(word)
-            except ValueError:
-                continue
-            # If the clusterer has only seen the word a few times, its
-            # cluster is unreliable.
-            if int(freq) >= 3:
-                clusters[word] = cluster
-            else:
-                clusters[word] = "0"
-    # Expand clusters with re-casing
-    for word, cluster in list(clusters.items()):
-        if word.lower() not in clusters:
-            clusters[word.lower()] = cluster
-        if word.title() not in clusters:
-            clusters[word.title()] = cluster
-        if word.upper() not in clusters:
-            clusters[word.upper()] = cluster
-    return clusters

spacy/cli/init_pipeline.py

@@ -0,0 +1,117 @@
+from typing import Optional
+import logging
+from pathlib import Path
+from wasabi import msg
+import typer
+import srsly
+
+from .. import util
+from ..training.initialize import init_nlp, convert_vectors
+from ..language import Language
+from ._util import init_cli, Arg, Opt, parse_config_overrides, show_validation_error
+from ._util import import_code, setup_gpu
+
+
+@init_cli.command("vectors")
+def init_vectors_cli(
+    # fmt: off
+    lang: str = Arg(..., help="The language of the nlp object to create"),
+    vectors_loc: Path = Arg(..., help="Vectors file in Word2Vec format", exists=True),
+    output_dir: Path = Arg(..., help="Pipeline output directory"),
+    prune: int = Opt(-1, "--prune", "-p", help="Optional number of vectors to prune to"),
+    truncate: int = Opt(0, "--truncate", "-t", help="Optional number of vectors to truncate to when reading in vectors file"),
+    name: Optional[str] = Opt(None, "--name", "-n", help="Optional name for the word vectors, e.g. en_core_web_lg.vectors"),
+    verbose: bool = Opt(False, "--verbose", "-V", "-VV", help="Display more information for debugging purposes"),
+    jsonl_loc: Optional[Path] = Opt(None, "--lexemes-jsonl", "-j", help="Location of JSONL-formatted attributes file", hidden=True),
+    # fmt: on
+):
+    """Convert word vectors for use with spaCy. Will export an nlp object that
+    you can use in the [initialize] block of your config to initialize
+    a model with vectors.
+    """
+    util.logger.setLevel(logging.DEBUG if verbose else logging.INFO)
+    msg.info(f"Creating blank nlp object for language '{lang}'")
+    nlp = util.get_lang_class(lang)()
+    if jsonl_loc is not None:
+        update_lexemes(nlp, jsonl_loc)
+    convert_vectors(nlp, vectors_loc, truncate=truncate, prune=prune, name=name)
+    msg.good(f"Successfully converted {len(nlp.vocab.vectors)} vectors")
+    nlp.to_disk(output_dir)
+    msg.good(
+        "Saved nlp object with vectors to output directory. You can now use the "
+        "path to it in your config as the 'vectors' setting in [initialize.vocab].",
+        output_dir.resolve(),
+    )
+
+
+def update_lexemes(nlp: Language, jsonl_loc: Path) -> None:
+    # Mostly used for backwards-compatibility and may be removed in the future
+    lex_attrs = srsly.read_jsonl(jsonl_loc)
+    for attrs in lex_attrs:
+        if "settings" in attrs:
+            continue
+        lexeme = nlp.vocab[attrs["orth"]]
+        lexeme.set_attrs(**attrs)
+
+
+@init_cli.command(
+    "nlp",
+    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+    hidden=True,
+)
+def init_pipeline_cli(
+    # fmt: off
+    ctx: typer.Context,  # This is only used to read additional arguments
+    config_path: Path = Arg(..., help="Path to config file", exists=True),
+    output_path: Path = Arg(..., help="Output directory for the prepared data"),
+    code_path: Optional[Path] = Opt(None, "--code", "-c", help="Path to Python file with additional code (registered functions) to be imported"),
+    verbose: bool = Opt(False, "--verbose", "-V", "-VV", help="Display more information for debugging purposes"),
+    use_gpu: int = Opt(-1, "--gpu-id", "-g", help="GPU ID or -1 for CPU")
+    # fmt: on
+):
+    util.logger.setLevel(logging.DEBUG if verbose else logging.INFO)
+    overrides = parse_config_overrides(ctx.args)
+    import_code(code_path)
+    setup_gpu(use_gpu)
+    with show_validation_error(config_path):
+        config = util.load_config(config_path, overrides=overrides)
+    with show_validation_error(hint_fill=False):
+        nlp = init_nlp(config, use_gpu=use_gpu)
+    nlp.to_disk(output_path)
+    msg.good(f"Saved initialized pipeline to {output_path}")
+
+
+@init_cli.command(
+    "labels",
+    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+)
+def init_labels_cli(
+    # fmt: off
+    ctx: typer.Context,  # This is only used to read additional arguments
+    config_path: Path = Arg(..., help="Path to config file", exists=True),
+    output_path: Path = Arg(..., help="Output directory for the labels"),
+    code_path: Optional[Path] = Opt(None, "--code", "-c", help="Path to Python file with additional code (registered functions) to be imported"),
+    verbose: bool = Opt(False, "--verbose", "-V", "-VV", help="Display more information for debugging purposes"),
+    use_gpu: int = Opt(-1, "--gpu-id", "-g", help="GPU ID or -1 for CPU")
+    # fmt: on
+):
+    """Generate JSON files for the labels in the data. This helps speed up the
+    training process, since spaCy won't have to preprocess the data to
+    extract the labels."""
+    util.logger.setLevel(logging.DEBUG if verbose else logging.INFO)
+    if not output_path.exists():
+        output_path.mkdir()
+    overrides = parse_config_overrides(ctx.args)
+    import_code(code_path)
+    setup_gpu(use_gpu)
+    with show_validation_error(config_path):
+        config = util.load_config(config_path, overrides=overrides)
+    with show_validation_error(hint_fill=False):
+        nlp = init_nlp(config, use_gpu=use_gpu)
+    for name, component in nlp.pipeline:
+        if getattr(component, "label_data", None) is not None:
+            output_file = output_path / f"{name}.json"
+            srsly.write_json(output_file, component.label_data)
+            msg.good(f"Saving {name} labels to {output_file}")
+        else:
+            msg.info(f"No labels found for {name}")

spacy/cli/pretrain.py

@@ -1,25 +1,13 @@
 from typing import Optional
-import numpy
-import time
-import re
-from collections import Counter
 from pathlib import Path
-from thinc.api import require_gpu, set_gpu_allocator
-from thinc.api import set_dropout_rate, to_categorical, fix_random_seed
-from thinc.api import Config, CosineDistance, L2Distance
 from wasabi import msg
-import srsly
-from functools import partial
 import typer
+import re
 
 from ._util import app, Arg, Opt, parse_config_overrides, show_validation_error
-from ._util import import_code
+from ._util import import_code, setup_gpu
-from ..ml.models.multi_task import build_cloze_multi_task_model
+from ..training.pretrain import pretrain
-from ..ml.models.multi_task import build_cloze_characters_multi_task_model
+from ..util import load_config
-from ..tokens import Doc
-from ..attrs import ID
-from .. import util
-from ..util import dot_to_object
 
 
 @app.command(
@@ -61,15 +49,11 @@ def pretrain_cli(
     config_overrides = parse_config_overrides(ctx.args)
     import_code(code_path)
     verify_cli_args(config_path, output_dir, resume_path, epoch_resume)
-    if use_gpu >= 0:
+    setup_gpu(use_gpu)
-        msg.info("Using GPU")
-        require_gpu(use_gpu)
-    else:
-        msg.info("Using CPU")
     msg.info(f"Loading config from: {config_path}")
 
     with show_validation_error(config_path):
-        raw_config = util.load_config(
+        raw_config = load_config(
             config_path, overrides=config_overrides, interpolate=False
         )
     config = raw_config.interpolate()
@@ -89,250 +73,11 @@ def pretrain_cli(
         resume_path=resume_path,
         epoch_resume=epoch_resume,
         use_gpu=use_gpu,
+        silent=False,
     )
-
-
-def pretrain(
-    config: Config,
-    output_dir: Path,
-    resume_path: Optional[Path] = None,
-    epoch_resume: Optional[int] = None,
-    use_gpu: int = -1,
-):
-    if config["training"]["seed"] is not None:
-        fix_random_seed(config["training"]["seed"])
-    allocator = config["training"]["gpu_allocator"]
-    if use_gpu >= 0 and allocator:
-        set_gpu_allocator(allocator)
-    nlp = util.load_model_from_config(config)
-    C = util.resolve_training_config(nlp.config)
-    P_cfg = C["pretraining"]
-    corpus = dot_to_object(C, P_cfg["corpus"])
-    batcher = P_cfg["batcher"]
-    model = create_pretraining_model(nlp, C["pretraining"])
-    optimizer = C["pretraining"]["optimizer"]
-    # Load in pretrained weights to resume from
-    if resume_path is not None:
-        _resume_model(model, resume_path, epoch_resume)
-    else:
-        # Without '--resume-path' the '--epoch-resume' argument is ignored
-        epoch_resume = 0
-
-    tracker = ProgressTracker(frequency=10000)
-    msg.divider(f"Pre-training tok2vec layer - starting at epoch {epoch_resume}")
-    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 / f"model{epoch}{is_temp_str}.bin").open("wb") as file_:
-                file_.write(model.get_ref("tok2vec").to_bytes())
-            log = {
-                "nr_word": tracker.nr_word,
-                "loss": tracker.loss,
-                "epoch_loss": tracker.epoch_loss,
-                "epoch": epoch,
-            }
-            with (output_dir / "log.jsonl").open("a") as file_:
-                file_.write(srsly.json_dumps(log) + "\n")
-
-    objective = create_objective(P_cfg["objective"])
-    # TODO: I think we probably want this to look more like the
-    # 'create_train_batches' function?
-    for epoch in range(epoch_resume, P_cfg["max_epochs"]):
-        for batch_id, batch in enumerate(batcher(corpus(nlp))):
-            docs = ensure_docs(batch)
-            loss = make_update(model, docs, optimizer, objective)
-            progress = tracker.update(epoch, loss, docs)
-            if progress:
-                msg.row(progress, **row_settings)
-            if P_cfg["n_save_every"] and (batch_id % P_cfg["n_save_every"] == 0):
-                _save_model(epoch, is_temp=True)
-        _save_model(epoch)
-        tracker.epoch_loss = 0.0
     msg.good("Successfully finished pretrain")
 
 
-def ensure_docs(examples_or_docs):
-    docs = []
-    for eg_or_doc in examples_or_docs:
-        if isinstance(eg_or_doc, Doc):
-            docs.append(eg_or_doc)
-        else:
-            docs.append(eg_or_doc.reference)
-    return docs
-
-
-def _resume_model(model, resume_path, epoch_resume):
-    msg.info(f"Resume training tok2vec from: {resume_path}")
-    with resume_path.open("rb") as file_:
-        weights_data = file_.read()
-        model.get_ref("tok2vec").from_bytes(weights_data)
-    # Parse the epoch number from the given weight file
-    model_name = re.search(r"model\d+\.bin", str(resume_path))
-    if model_name:
-        # Default weight file name so read epoch_start from it by cutting off 'model' and '.bin'
-        epoch_resume = int(model_name.group(0)[5:][:-4]) + 1
-        msg.info(f"Resuming from epoch: {epoch_resume}")
-    else:
-        msg.info(f"Resuming from epoch: {epoch_resume}")
-
-
-def make_update(model, docs, optimizer, objective_func):
-    """Perform an update over a single batch of documents.
-
-    docs (iterable): A batch of `Doc` objects.
-    optimizer (callable): An optimizer.
-    RETURNS loss: A float for the loss.
-    """
-    predictions, backprop = model.begin_update(docs)
-    loss, gradients = objective_func(model.ops, docs, predictions)
-    backprop(gradients)
-    model.finish_update(optimizer)
-    # Don't want to return a cupy object here
-    # The gradients are modified in-place by the BERT MLM,
-    # so we get an accurate loss
-    return float(loss)
-
-
-def create_objective(config):
-    """Create the objective for pretraining.
-
-    We'd like to replace this with a registry function but it's tricky because
-    we're also making a model choice based on this. For now we hard-code support
-    for two types (characters, vectors). For characters you can specify
-    n_characters, for vectors you can specify the loss.
-
-    Bleh.
-    """
-    objective_type = config["type"]
-    if objective_type == "characters":
-        return partial(get_characters_loss, nr_char=config["n_characters"])
-    elif objective_type == "vectors":
-        if config["loss"] == "cosine":
-            return partial(
-                get_vectors_loss,
-                distance=CosineDistance(normalize=True, ignore_zeros=True),
-            )
-        elif config["loss"] == "L2":
-            return partial(
-                get_vectors_loss, distance=L2Distance(normalize=True, ignore_zeros=True)
-            )
-        else:
-            raise ValueError("Unexpected loss type", config["loss"])
-    else:
-        raise ValueError("Unexpected objective_type", objective_type)
-
-
-def get_vectors_loss(ops, docs, prediction, distance):
-    """Compute a loss based on a distance between the documents' vectors and
-    the prediction.
-    """
-    # The simplest way to implement this would be to vstack the
-    # token.vector values, but that's a bit inefficient, especially on GPU.
-    # Instead we fetch the index into the vectors table for each of our tokens,
-    # 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]
-    d_target, loss = distance(prediction, target)
-    return loss, d_target
-
-
-def get_characters_loss(ops, docs, prediction, nr_char):
-    """Compute a loss based on a number of characters predicted from the docs."""
-    target_ids = numpy.vstack([doc.to_utf8_array(nr_char=nr_char) for doc in docs])
-    target_ids = target_ids.reshape((-1,))
-    target = ops.asarray(to_categorical(target_ids, n_classes=256), dtype="f")
-    target = target.reshape((-1, 256 * nr_char))
-    diff = prediction - target
-    loss = (diff ** 2).sum()
-    d_target = diff / float(prediction.shape[0])
-    return loss, d_target
-
-
-def create_pretraining_model(nlp, pretrain_config):
-    """Define a network for the pretraining. We simply add an output layer onto
-    the tok2vec input model. The tok2vec input model needs to be a model that
-    takes a batch of Doc objects (as a list), and returns a list of arrays.
-    Each array in the output needs to have one row per token in the doc.
-    The actual tok2vec layer is stored as a reference, and only this bit will be
-    serialized to file and read back in when calling the 'train' command.
-    """
-    component = nlp.get_pipe(pretrain_config["component"])
-    if pretrain_config.get("layer"):
-        tok2vec = component.model.get_ref(pretrain_config["layer"])
-    else:
-        tok2vec = component.model
-
-    # TODO
-    maxout_pieces = 3
-    hidden_size = 300
-    if pretrain_config["objective"]["type"] == "vectors":
-        model = build_cloze_multi_task_model(
-            nlp.vocab, tok2vec, hidden_size=hidden_size, maxout_pieces=maxout_pieces
-        )
-    elif pretrain_config["objective"]["type"] == "characters":
-        model = build_cloze_characters_multi_task_model(
-            nlp.vocab,
-            tok2vec,
-            hidden_size=hidden_size,
-            maxout_pieces=maxout_pieces,
-            nr_char=pretrain_config["objective"]["n_characters"],
-        )
-    model.initialize(X=[nlp.make_doc("Give it a doc to infer shapes")])
-    set_dropout_rate(model, pretrain_config["dropout"])
-    return model
-
-
-class ProgressTracker:
-    def __init__(self, frequency=1000000):
-        self.loss = 0.0
-        self.prev_loss = 0.0
-        self.nr_word = 0
-        self.words_per_epoch = Counter()
-        self.frequency = frequency
-        self.last_time = time.time()
-        self.last_update = 0
-        self.epoch_loss = 0.0
-
-    def update(self, epoch, loss, docs):
-        self.loss += loss
-        self.epoch_loss += loss
-        words_in_batch = sum(len(doc) for doc in docs)
-        self.words_per_epoch[epoch] += words_in_batch
-        self.nr_word += words_in_batch
-        words_since_update = self.nr_word - self.last_update
-        if words_since_update >= self.frequency:
-            wps = words_since_update / (time.time() - self.last_time)
-            self.last_update = self.nr_word
-            self.last_time = time.time()
-            loss_per_word = self.loss - self.prev_loss
-            status = (
-                epoch,
-                self.nr_word,
-                _smart_round(self.loss, width=10),
-                _smart_round(loss_per_word, width=6),
-                int(wps),
-            )
-            self.prev_loss = float(self.loss)
-            return status
-        else:
-            return None
-
-
-def _smart_round(figure, width=10, max_decimal=4):
-    """Round large numbers as integers, smaller numbers as decimals."""
-    n_digits = len(str(int(figure)))
-    n_decimal = width - (n_digits + 1)
-    if n_decimal <= 1:
-        return str(int(figure))
-    else:
-        n_decimal = min(n_decimal, max_decimal)
-        format_str = "%." + str(n_decimal) + "f"
-        return format_str % figure
-
-
 def verify_cli_args(config_path, output_dir, resume_path, epoch_resume):
     if not config_path or not config_path.exists():
         msg.fail("Config file not found", config_path, exits=1)

spacy/cli/project/dvc.py

@@ -134,7 +134,7 @@ def update_dvc_config(
 
 
 def run_dvc_commands(
-    commands: Iterable[str] = SimpleFrozenList(), flags: Dict[str, bool] = {},
+    commands: Iterable[str] = SimpleFrozenList(), flags: Dict[str, bool] = {}
 ) -> None:
     """Run a sequence of DVC commands in a subprocess, in order.
 

spacy/cli/templates/quickstart_training.jinja

@@ -4,8 +4,8 @@ can help generate the best possible configuration, given a user's requirements.
 {%- set use_transformer = (transformer_data and hardware != "cpu") -%}
 {%- set transformer = transformer_data[optimize] if use_transformer else {} -%}
 [paths]
-train = ""
+train = null
-dev = ""
+dev = null
 
 [system]
 {% if use_transformer -%}
@@ -37,6 +37,22 @@ tokenizer_config = {"use_fast": true}
 window = 128
 stride = 96
 
+{% if "morphologizer" in components %}
+[components.morphologizer]
+factory = "morphologizer"
+
+[components.morphologizer.model]
+@architectures = "spacy.Tagger.v1"
+nO = null
+
+[components.morphologizer.model.tok2vec]
+@architectures = "spacy-transformers.TransformerListener.v1"
+grad_factor = 1.0
+
+[components.morphologizer.model.tok2vec.pooling]
+@layers = "reduce_mean.v1"
+{%- endif %}
+
 {% if "tagger" in components %}
 [components.tagger]
 factory = "tagger"
@@ -166,6 +182,19 @@ depth = {{ 4 if optimize == "efficiency" else 8 }}
 window_size = 1
 maxout_pieces = 3
 
+{% if "morphologizer" in components %}
+[components.morphologizer]
+factory = "morphologizer"
+
+[components.morphologizer.model]
+@architectures = "spacy.Tagger.v1"
+nO = null
+
+[components.morphologizer.model.tok2vec]
+@architectures = "spacy.Tok2VecListener.v1"
+width = ${components.tok2vec.model.encode.width}
+{%- endif %}
+
 {% if "tagger" in components %}
 [components.tagger]
 factory = "tagger"
@@ -257,7 +286,7 @@ no_output_layer = false
 {% endif %}
 
 {% for pipe in components %}
-{% if pipe not in ["tagger", "parser", "ner", "textcat", "entity_linker"] %}
+{% if pipe not in ["tagger", "morphologizer", "parser", "ner", "textcat", "entity_linker"] %}
 {# Other components defined by the user: we just assume they're factories #}
 [components.{{ pipe }}]
 factory = "{{ pipe }}"
@@ -270,7 +299,6 @@ factory = "{{ pipe }}"
 @readers = "spacy.Corpus.v1"
 path = ${paths.train}
 max_length = {{ 500 if hardware == "gpu" else 2000 }}
-augmenter = {"@augmenters": "spacy.orth_variants.v1", "level": 0.1, "lower": 0.5}
 
 [corpora.dev]
 @readers = "spacy.Corpus.v1"
@@ -278,11 +306,6 @@ path = ${paths.dev}
 max_length = 0
 
 [training]
-{% if use_transformer or optimize == "efficiency" or not word_vectors -%}
-vectors = null
-{% else -%}
-vectors = "{{ word_vectors }}"
-{% endif -%}
 {% if use_transformer -%}
 accumulate_gradient = {{ transformer["size_factor"] }}
 {% endif -%}
@@ -318,3 +341,10 @@ start = 100
 stop = 1000
 compound = 1.001
 {% endif %}
+
+[initialize]
+{% if use_transformer or optimize == "efficiency" or not word_vectors -%}
+vectors = null
+{% else -%}
+vectors = "{{ word_vectors }}"
+{% endif -%}

spacy/cli/train.py

@@ -1,23 +1,14 @@
-from typing import Optional, Dict, Any, Tuple, Union, Callable, List
+from typing import Optional
-from timeit import default_timer as timer
-import srsly
-import tqdm
 from pathlib import Path
 from wasabi import msg
-import thinc
-import thinc.schedules
-from thinc.api import Config, Optimizer, require_gpu, fix_random_seed, set_gpu_allocator
-import random
 import typer
 import logging
 
 from ._util import app, Arg, Opt, parse_config_overrides, show_validation_error
-from ._util import import_code, get_sourced_components
+from ._util import import_code, setup_gpu
-from ..language import Language
+from ..training.loop import train
+from ..training.initialize import init_nlp
 from .. import util
-from ..training.example import Example
-from ..errors import Errors
-from ..util import dot_to_object
 
 
 @app.command(
@@ -30,8 +21,7 @@ def train_cli(
     output_path: Optional[Path] = Opt(None, "--output", "--output-path", "-o", help="Output directory to store trained pipeline in"),
     code_path: Optional[Path] = Opt(None, "--code", "-c", help="Path to Python file with additional code (registered functions) to be imported"),
     verbose: bool = Opt(False, "--verbose", "-V", "-VV", help="Display more information for debugging purposes"),
-    use_gpu: int = Opt(-1, "--gpu-id", "-g", help="GPU ID or -1 for CPU"),
+    use_gpu: int = Opt(-1, "--gpu-id", "-g", help="GPU ID or -1 for CPU")
-    resume: bool = Opt(False, "--resume", "-R", help="Resume training"),
     # fmt: on
 ):
     """
@@ -48,393 +38,19 @@ def train_cli(
 
     DOCS: https://nightly.spacy.io/api/cli#train
     """
-    util.logger.setLevel(logging.DEBUG if verbose else logging.ERROR)
+    util.logger.setLevel(logging.DEBUG if verbose else logging.INFO)
     verify_cli_args(config_path, output_path)
     overrides = parse_config_overrides(ctx.args)
     import_code(code_path)
-    train(
+    setup_gpu(use_gpu)
-        config_path,
-        output_path=output_path,
-        config_overrides=overrides,
-        use_gpu=use_gpu,
-        resume_training=resume,
-    )
-
-
-def train(
-    config_path: Path,
-    output_path: Optional[Path] = None,
-    config_overrides: Dict[str, Any] = {},
-    use_gpu: int = -1,
-    resume_training: bool = False,
-) -> None:
-    if use_gpu >= 0:
-        msg.info(f"Using GPU: {use_gpu}")
-        require_gpu(use_gpu)
-    else:
-        msg.info("Using CPU")
-    msg.info(f"Loading config and nlp from: {config_path}")
     with show_validation_error(config_path):
-        # Keep an un-interpolated config so we can preserve variables in
+        config = util.load_config(config_path, overrides=overrides, interpolate=False)
-        # the final nlp object we train and serialize
+    msg.divider("Initializing pipeline")
-        raw_config = util.load_config(
+    with show_validation_error(config_path, hint_fill=False):
-            config_path, overrides=config_overrides, interpolate=False
+        nlp = init_nlp(config, use_gpu=use_gpu)
-        )
+    msg.good("Initialized pipeline")
-    config = raw_config.interpolate()
+    msg.divider("Training pipeline")
-    if config["training"]["seed"] is not None:
+    train(nlp, output_path, use_gpu=use_gpu, silent=False)
-        fix_random_seed(config["training"]["seed"])
-    allocator = config["training"]["gpu_allocator"]
-    if use_gpu >= 0 and allocator:
-        set_gpu_allocator(allocator)
-    # Use original config here before it's resolved to functions
-    sourced_components = get_sourced_components(config)
-    with show_validation_error(config_path):
-        nlp = util.load_model_from_config(raw_config)
-        # Resolve all training-relevant sections using the filled nlp config
-        C = util.resolve_training_config(nlp.config)
-    util.load_vocab_data_into_model(nlp, lookups=C["training"]["lookups"])
-    if C["training"]["vectors"] is not None:
-        add_vectors(nlp, C["training"]["vectors"])
-    raw_text, tag_map, morph_rules, weights_data = load_from_paths(C)
-    T_cfg = C["training"]
-    optimizer = T_cfg["optimizer"]
-    train_corpus = dot_to_object(C, T_cfg["train_corpus"])
-    dev_corpus = dot_to_object(C, T_cfg["dev_corpus"])
-    batcher = T_cfg["batcher"]
-    train_logger = T_cfg["logger"]
-    before_to_disk = create_before_to_disk_callback(T_cfg["before_to_disk"])
-    # Components that shouldn't be updated during training
-    frozen_components = T_cfg["frozen_components"]
-    # Sourced components that require resume_training
-    resume_components = [p for p in sourced_components if p not in frozen_components]
-    msg.info(f"Pipeline: {nlp.pipe_names}")
-    if resume_components:
-        with nlp.select_pipes(enable=resume_components):
-            msg.info(f"Resuming training for: {resume_components}")
-            nlp.resume_training(sgd=optimizer)
-    with nlp.select_pipes(disable=[*frozen_components, *resume_components]):
-        nlp.begin_training(lambda: train_corpus(nlp), sgd=optimizer)
-    # Verify the config after calling 'begin_training' to ensure labels are properly initialized
-    verify_config(nlp)
-
-    if tag_map:
-        # Replace tag map with provided mapping
-        nlp.vocab.morphology.load_tag_map(tag_map)
-    if morph_rules:
-        # Load morph rules
-        nlp.vocab.morphology.load_morph_exceptions(morph_rules)
-
-    # Load pretrained tok2vec weights - cf. CLI command 'pretrain'
-    if weights_data is not None:
-        tok2vec_component = C["pretraining"]["component"]
-        if tok2vec_component is None:
-            msg.fail(
-                f"To use pretrained tok2vec weights, [pretraining.component] "
-                f"needs to specify the component that should load them.",
-                exits=1,
-            )
-        layer = nlp.get_pipe(tok2vec_component).model
-        tok2vec_layer = C["pretraining"]["layer"]
-        if tok2vec_layer:
-            layer = layer.get_ref(tok2vec_layer)
-        layer.from_bytes(weights_data)
-        msg.info(f"Loaded pretrained weights into component '{tok2vec_component}'")
-
-    # Create iterator, which yields out info after each optimization step.
-    msg.info("Start training")
-    score_weights = T_cfg["score_weights"]
-    training_step_iterator = train_while_improving(
-        nlp,
-        optimizer,
-        create_train_batches(train_corpus(nlp), batcher, T_cfg["max_epochs"]),
-        create_evaluation_callback(nlp, dev_corpus, score_weights),
-        dropout=T_cfg["dropout"],
-        accumulate_gradient=T_cfg["accumulate_gradient"],
-        patience=T_cfg["patience"],
-        max_steps=T_cfg["max_steps"],
-        eval_frequency=T_cfg["eval_frequency"],
-        raw_text=None,
-        exclude=frozen_components,
-    )
-    msg.info(f"Training. Initial learn rate: {optimizer.learn_rate}")
-    with nlp.select_pipes(disable=frozen_components):
-        print_row, finalize_logger = train_logger(nlp)
-
-    try:
-        progress = tqdm.tqdm(total=T_cfg["eval_frequency"], leave=False)
-        progress.set_description(f"Epoch 1")
-        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:
-                    with nlp.select_pipes(disable=frozen_components):
-                        update_meta(T_cfg, nlp, info)
-                    with nlp.use_params(optimizer.averages):
-                        nlp = before_to_disk(nlp)
-                        nlp.to_disk(output_path / "model-best")
-                progress = tqdm.tqdm(total=T_cfg["eval_frequency"], leave=False)
-                progress.set_description(f"Epoch {info['epoch']}")
-    except Exception as e:
-        finalize_logger()
-        if output_path is not None:
-            # We don't want to swallow the traceback if we don't have a
-            # specific error.
-            msg.warn(
-                f"Aborting and saving the final best model. "
-                f"Encountered exception: {str(e)}"
-            )
-            nlp = before_to_disk(nlp)
-            nlp.to_disk(output_path / "model-final")
-        raise e
-    finally:
-        finalize_logger()
-        if output_path is not None:
-            final_model_path = output_path / "model-final"
-            if optimizer.averages:
-                with nlp.use_params(optimizer.averages):
-                    nlp.to_disk(final_model_path)
-            else:
-                nlp.to_disk(final_model_path)
-            msg.good(f"Saved pipeline to output directory {final_model_path}")
-
-
-def add_vectors(nlp: Language, vectors: str) -> None:
-    title = f"Config validation error for vectors {vectors}"
-    desc = (
-        "This typically means that there's a problem in the config.cfg included "
-        "with the packaged vectors. Make sure that the vectors package you're "
-        "loading is compatible with the current version of spaCy."
-    )
-    with show_validation_error(
-        title=title, desc=desc, hint_fill=False, show_config=False
-    ):
-        util.load_vectors_into_model(nlp, vectors)
-
-
-def create_train_batches(iterator, batcher, max_epochs: int):
-    epoch = 0
-    examples = list(iterator)
-    if not examples:
-        # Raise error if no data
-        raise ValueError(Errors.E986)
-    while max_epochs < 1 or epoch != max_epochs:
-        random.shuffle(examples)
-        for batch in batcher(examples):
-            yield epoch, batch
-        epoch += 1
-
-
-def create_evaluation_callback(
-    nlp: Language, dev_corpus: Callable, weights: Dict[str, float]
-) -> Callable[[], Tuple[float, Dict[str, float]]]:
-    weights = {key: value for key, value in weights.items() if value is not None}
-
-    def evaluate() -> Tuple[float, Dict[str, float]]:
-        dev_examples = list(dev_corpus(nlp))
-        scores = nlp.evaluate(dev_examples)
-        # Calculate a weighted sum based on score_weights for the main score.
-        # We can only consider scores that are ints/floats, not dicts like
-        # entity scores per type etc.
-        for key, value in scores.items():
-            if key in weights and not isinstance(value, (int, float)):
-                raise ValueError(Errors.E915.format(name=key, score_type=type(value)))
-        try:
-            weighted_score = sum(
-                scores.get(s, 0.0) * weights.get(s, 0.0) for s in weights
-            )
-        except KeyError as e:
-            keys = list(scores.keys())
-            err = Errors.E983.format(dict="score_weights", key=str(e), keys=keys)
-            raise KeyError(err) from None
-        return weighted_score, scores
-
-    return evaluate
-
-
-def create_before_to_disk_callback(
-    callback: Optional[Callable[[Language], Language]]
-) -> Callable[[Language], Language]:
-    def before_to_disk(nlp: Language) -> Language:
-        if not callback:
-            return nlp
-        modified_nlp = callback(nlp)
-        if not isinstance(modified_nlp, Language):
-            err = Errors.E914.format(name="before_to_disk", value=type(modified_nlp))
-            raise ValueError(err)
-        return modified_nlp
-
-    return before_to_disk
-
-
-def train_while_improving(
-    nlp: Language,
-    optimizer: Optimizer,
-    train_data,
-    evaluate,
-    *,
-    dropout: float,
-    eval_frequency: int,
-    accumulate_gradient: int,
-    patience: int,
-    max_steps: int,
-    raw_text: List[Dict[str, str]],
-    exclude: List[str],
-):
-    """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.
-
-    Positional arguments:
-        nlp: The spaCy pipeline to evaluate.
-        optimizer: The optimizer callable.
-        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 list of Example objects.
-    * 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 from the last evaluation.
-        other_scores: : The other scores from the last evaluation.
-        losses: 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 = {}
-    if raw_text:
-        random.shuffle(raw_text)
-        raw_examples = [
-            Example.from_dict(nlp.make_doc(rt["text"]), {}) for rt in raw_text
-        ]
-        raw_batches = util.minibatch(raw_examples, size=8)
-
-    words_seen = 0
-    start_time = timer()
-    for step, (epoch, batch) in enumerate(train_data):
-        dropout = next(dropouts)
-        for subbatch in subdivide_batch(batch, accumulate_gradient):
-
-            nlp.update(
-                subbatch, drop=dropout, losses=losses, sgd=False, exclude=exclude
-            )
-            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, exclude=exclude)
-        # TODO: refactor this so we don't have to run it separately in here
-        for name, proc in nlp.pipeline:
-            if (
-                name not in exclude
-                and hasattr(proc, "model")
-                and proc.model not in (True, False, None)
-            ):
-                proc.model.finish_update(optimizer)
-        optimizer.step_schedules()
-        if not (step % eval_frequency):
-            if optimizer.averages:
-                with nlp.use_params(optimizer.averages):
-                    score, other_scores = evaluate()
-            else:
-                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
-        words_seen += sum(len(eg) for eg in batch)
-        info = {
-            "epoch": epoch,
-            "step": step,
-            "score": score,
-            "other_scores": other_scores,
-            "losses": losses,
-            "checkpoints": results,
-            "seconds": int(timer() - start_time),
-            "words": words_seen,
-        }
-        yield batch, info, is_best_checkpoint
-        if is_best_checkpoint is not None:
-            losses = {}
-        # Stop if no improvement in `patience` updates (if specified)
-        best_score, best_step = max(results)
-        if patience and (step - best_step) >= patience:
-            break
-        # Stop if we've exhausted our max steps (if specified)
-        if max_steps and step >= max_steps:
-            break
-
-
-def subdivide_batch(batch, accumulate_gradient):
-    batch = list(batch)
-    batch.sort(key=lambda eg: len(eg.predicted))
-    sub_len = len(batch) // accumulate_gradient
-    start = 0
-    for i in range(accumulate_gradient):
-        subbatch = batch[start : start + sub_len]
-        if subbatch:
-            yield subbatch
-        start += len(subbatch)
-    subbatch = batch[start:]
-    if subbatch:
-        yield subbatch
-
-
-def update_meta(
-    training: Union[Dict[str, Any], Config], nlp: Language, info: Dict[str, Any]
-) -> None:
-    nlp.meta["performance"] = {}
-    for metric in training["score_weights"]:
-        if metric is not None:
-            nlp.meta["performance"][metric] = info["other_scores"].get(metric, 0.0)
-    for pipe_name in nlp.pipe_names:
-        nlp.meta["performance"][f"{pipe_name}_loss"] = info["losses"][pipe_name]
-
-
-def load_from_paths(
-    config: Config,
-) -> Tuple[List[Dict[str, str]], Dict[str, dict], bytes]:
-    # TODO: separate checks from loading
-    raw_text = util.ensure_path(config["training"]["raw_text"])
-    if raw_text is not None:
-        if not raw_text.exists():
-            msg.fail("Can't find raw text", raw_text, exits=1)
-        raw_text = list(srsly.read_jsonl(config["training"]["raw_text"]))
-    tag_map = {}
-    morph_rules = {}
-    weights_data = None
-    init_tok2vec = util.ensure_path(config["training"]["init_tok2vec"])
-    if init_tok2vec is not None:
-        if not init_tok2vec.exists():
-            msg.fail("Can't find pretrained tok2vec", init_tok2vec, exits=1)
-        with init_tok2vec.open("rb") as file_:
-            weights_data = file_.read()
-    return raw_text, tag_map, morph_rules, weights_data
 
 
 def verify_cli_args(config_path: Path, output_path: Optional[Path] = None) -> None:
@@ -445,30 +61,3 @@ def verify_cli_args(config_path: Path, output_path: Optional[Path] = None) -> No
         if not output_path.exists():
             output_path.mkdir()
             msg.good(f"Created output directory: {output_path}")
-
-
-def verify_config(nlp: Language) -> None:
-    """Perform additional checks based on the config, loaded nlp object and training data."""
-    # TODO: maybe we should validate based on the actual components, the list
-    # in config["nlp"]["pipeline"] instead?
-    for pipe_config in nlp.config["components"].values():
-        # We can't assume that the component name == the factory
-        factory = pipe_config["factory"]
-        if factory == "textcat":
-            verify_textcat_config(nlp, pipe_config)
-
-
-def verify_textcat_config(nlp: Language, pipe_config: Dict[str, Any]) -> None:
-    # if 'positive_label' is provided: double check whether it's in the data and
-    # the task is binary
-    if pipe_config.get("positive_label"):
-        textcat_labels = nlp.get_pipe("textcat").labels
-        pos_label = pipe_config.get("positive_label")
-        if pos_label not in textcat_labels:
-            raise ValueError(
-                Errors.E920.format(pos_label=pos_label, labels=textcat_labels)
-            )
-        if len(list(textcat_labels)) != 2:
-            raise ValueError(
-                Errors.E919.format(pos_label=pos_label, labels=textcat_labels)
-            )

spacy/default_config.cfg

@@ -1,7 +1,7 @@
 [paths]
-train = ""
+train = null
-dev = ""
+dev = null
-raw = null
+vectors = null
 init_tok2vec = null
 
 [system]
@@ -10,8 +10,13 @@ gpu_allocator = null
 
 [nlp]
 lang = null
+# List of pipeline component names, in order. The names should correspond to
+# components defined in the [components block]
 pipeline = []
+# Components that are loaded but disabled by default
 disabled = []
+# Optional callbacks to modify the nlp object before it's initialized, after
+# it's created and after the pipeline has been set up
 before_creation = null
 after_creation = null
 after_pipeline_creation = null
@@ -19,6 +24,7 @@ after_pipeline_creation = null
 [nlp.tokenizer]
 @tokenizers = "spacy.Tokenizer.v1"
 
+# The pipeline components and their models
 [components]
 
 # Readers for corpora like dev and train.
@@ -37,9 +43,8 @@ max_length = 0
 limit = 0
 # Apply some simply data augmentation, where we replace tokens with variations.
 # This is especially useful for punctuation and case replacement, to help
-# generalize beyond corpora that don't have smart-quotes, or only have smart
+# generalize beyond corpora that don't/only have smart quotes etc.
-# quotes, etc.
+augmenter = null
-augmenter = {"@augmenters": "spacy.orth_variants.v1", "level": 0.1, "lower": 0.5}
 
 [corpora.dev]
 @readers = "spacy.Corpus.v1"
@@ -52,6 +57,8 @@ gold_preproc = false
 max_length = 0
 # Limitation on number of training examples
 limit = 0
+# Optional callback for data augmentation
+augmenter = null
 
 # Training hyper-parameters and additional features.
 [training]
@@ -59,11 +66,6 @@ seed = ${system.seed}
 gpu_allocator = ${system.gpu_allocator}
 dropout = 0.1
 accumulate_gradient = 1
-# Extra resources for transfer-learning or pseudo-rehearsal
-init_tok2vec = ${paths.init_tok2vec}
-raw_text = ${paths.raw}
-vectors = null
-lookups = null
 # Controls early-stopping. 0 or -1 mean unlimited.
 patience = 1600
 max_epochs = 0
@@ -104,3 +106,19 @@ grad_clip = 1.0
 use_averages = false
 eps = 1e-8
 learn_rate = 0.001
+
+# These settings are used when nlp.initialize() is called (typically before
+# training or pretraining). Components and the tokenizer can each define their
+# own arguments via their initialize methods that are populated by the config.
+# This lets them gather data resources, build label sets etc.
+[initialize]
+vectors = ${paths.vectors}
+# Extra resources for transfer-learning or pseudo-rehearsal
+init_tok2vec = ${paths.init_tok2vec}
+# Data and lookups for vocabulary
+vocab_data = null
+lookups = null
+# Arguments passed to the tokenizer's initialize method
+tokenizer = {}
+# Arguments for initialize methods of the components (keyed by component)
+components = {}

spacy/default_config_pretraining.cfg

@@ -1,3 +1,6 @@
+[paths]
+raw_text = null
+
 [pretraining]
 max_epochs = 1000
 dropout = 0.2
@@ -31,8 +34,8 @@ learn_rate = 0.001
 [corpora]
 
 [corpora.pretrain]
-@readers = "spacy.JsonlReader.v1"
+@readers = "spacy.JsonlCorpus.v1"
-path = ${paths.raw}
+path = ${paths.raw_text}
 min_length = 5
 max_length = 500
 limit = 0

spacy/errors.py

@@ -85,6 +85,7 @@ class Warnings:
             "attribute or operator.")
 
     # TODO: fix numbering after merging develop into master
+    W089 = ("The nlp.begin_training method has been renamed to nlp.initialize.")
     W090 = ("Could not locate any {format} files in path '{path}'.")
     W091 = ("Could not clean/remove the temp directory at {dir}: {msg}.")
     W092 = ("Ignoring annotations for sentence starts, as dependency heads are set.")
@@ -306,7 +307,7 @@ class Errors:
             "settings: {opts}")
     E107 = ("Value of doc._.{attr} is not JSON-serializable: {value}")
     E109 = ("Component '{name}' could not be run. Did you forget to "
-            "call begin_training()?")
+            "call initialize()?")
     E110 = ("Invalid displaCy render wrapper. Expected callable, got: {obj}")
     E111 = ("Pickling a token is not supported, because tokens are only views "
             "of the parent Doc and can't exist on their own. A pickled token "
@@ -376,7 +377,7 @@ class Errors:
             "provided {found}.")
     E143 = ("Labels for component '{name}' not initialized. This can be fixed "
             "by calling add_label, or by providing a representative batch of "
-            "examples to the component's begin_training method.")
+            "examples to the component's initialize method.")
     E145 = ("Error reading `{param}` from input file.")
     E146 = ("Could not access `{path}`.")
     E147 = ("Unexpected error in the {method} functionality of the "
@@ -418,7 +419,7 @@ class Errors:
     E164 = ("x is neither increasing nor decreasing: {}.")
     E165 = ("Only one class present in y_true. ROC AUC score is not defined in "
             "that case.")
-    E166 = ("Can only merge DocBins with the same pre-defined attributes.\n"
+    E166 = ("Can only merge DocBins with the same value for '{param}'.\n"
             "Current DocBin: {current}\nOther DocBin: {other}")
     E169 = ("Can't find module: {module}")
     E170 = ("Cannot apply transition {name}: invalid for the current state.")
@@ -476,6 +477,10 @@ class Errors:
     E201 = ("Span index out of range.")
 
     # TODO: fix numbering after merging develop into master
+    E912 = ("Failed to initialize lemmatizer. Missing lemmatizer table(s) found "
+            "for mode '{mode}'. Required tables: {tables}. Found: {found}.")
+    E913 = ("Corpus path can't be None. Maybe you forgot to define it in your "
+            "config.cfg or override it on the CLI?")
     E914 = ("Executing {name} callback failed. Expected the function to "
             "return the nlp object but got: {value}. Maybe you forgot to return "
             "the modified object in your function?")
@@ -517,7 +522,7 @@ class Errors:
             "but the provided argument {loc} points to a file.")
     E929 = ("A 'KnowledgeBase' could not be read from {loc} - the path does "
             "not seem to exist.")
-    E930 = ("Received invalid get_examples callback in {name}.begin_training. "
+    E930 = ("Received invalid get_examples callback in {name}.initialize. "
             "Expected function that returns an iterable of Example objects but "
             "got: {obj}")
     E931 = ("Encountered Pipe subclass without Pipe.{method} method in component "
@@ -553,7 +558,10 @@ class Errors:
     E953 = ("Mismatched IDs received by the Tok2Vec listener: {id1} vs. {id2}")
     E954 = ("The Tok2Vec listener did not receive any valid input from an upstream "
             "component.")
-    E955 = ("Can't find table(s) '{table}' for language '{lang}' in spacy-lookups-data.")
+    E955 = ("Can't find table(s) {table} for language '{lang}' in "
+            "spacy-lookups-data. Make sure you have the package installed or "
+            "provide your own lookup tables if no default lookups are available "
+            "for your language.")
     E956 = ("Can't find component '{name}' in [components] block in the config. "
             "Available components: {opts}")
     E957 = ("Writing directly to Language.factories isn't needed anymore in "
@@ -670,18 +678,17 @@ class Errors:
             "'{token_attrs}'.")
     E999 = ("Unable to merge the `Doc` objects because they do not all share "
             "the same `Vocab`.")
-    E1000 = ("No pkuseg model available. Provide a pkuseg model when "
+    E1000 = ("The Chinese word segmenter is pkuseg but no pkuseg model was "
-             "initializing the pipeline:\n"
+             "loaded. Provide the name of a pretrained model or the path to "
-             'cfg = {"tokenizer": {"segmenter": "pkuseg", "pkuseg_model": name_or_path}}\n'
+             "a model and initialize the pipeline:\n\n"
-             'nlp = Chinese(config=cfg)')
+             'nlp.tokenizer.initialize(pkuseg_model="default")')
     E1001 = ("Target token outside of matched span for match with tokens "
              "'{span}' and offset '{index}' matched by patterns '{patterns}'.")
     E1002 = ("Span index out of range.")
     E1003 = ("Unsupported lemmatizer mode '{mode}'.")
     E1004 = ("Missing lemmatizer table(s) found for lemmatizer mode '{mode}'. "
-             "Required tables '{tables}', found '{found}'. If you are not "
+             "Required tables: {tables}. Found: {found}. Maybe you forgot to "
-             "providing custom lookups, make sure you have the package "
+             "call nlp.initialize() to load in the data?")
-             "spacy-lookups-data installed.")
     E1005 = ("Unable to set attribute '{attr}' in tokenizer exception for "
              "'{chunk}'. Tokenizer exceptions are only allowed to specify "
              "`ORTH` and `NORM`.")
@@ -698,6 +705,9 @@ class Errors:
              "options: {modes}")
     E1012 = ("Entity spans and blocked/missing/outside spans should be "
              "provided to doc.set_ents as lists of `Span` objects.")
+    E1013 = ("Invalid morph: the MorphAnalysis must have the same vocab as the "
+             "token itself. To set the morph from this MorphAnalysis, set from "
+             "the string value with: `token.set_morph(str(other_morph))`.")
 
 
 @add_codes

spacy/lang/bn/__init__.py

@@ -4,7 +4,6 @@ from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
 from .punctuation import TOKENIZER_PREFIXES, TOKENIZER_SUFFIXES, TOKENIZER_INFIXES
 from .stop_words import STOP_WORDS
 from ...language import Language
-from ...lookups import Lookups
 from ...pipeline import Lemmatizer
 
 
@@ -24,18 +23,11 @@ class Bengali(Language):
 @Bengali.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return Lemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = Lemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return Lemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Bengali"]

spacy/lang/el/__init__.py

@@ -7,7 +7,6 @@ from .lex_attrs import LEX_ATTRS
 from .syntax_iterators import SYNTAX_ITERATORS
 from .punctuation import TOKENIZER_PREFIXES, TOKENIZER_SUFFIXES, TOKENIZER_INFIXES
 from .lemmatizer import GreekLemmatizer
-from ...lookups import Lookups
 from ...language import Language
 
 
@@ -29,18 +28,11 @@ class Greek(Language):
 @Greek.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return GreekLemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = GreekLemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return GreekLemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Greek"]

spacy/lang/en/__init__.py

@@ -1,5 +1,4 @@
 from typing import Optional
-
 from thinc.api import Model
 
 from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
@@ -9,7 +8,6 @@ from .syntax_iterators import SYNTAX_ITERATORS
 from .punctuation import TOKENIZER_INFIXES
 from .lemmatizer import EnglishLemmatizer
 from ...language import Language
-from ...lookups import Lookups
 
 
 class EnglishDefaults(Language.Defaults):
@@ -28,18 +26,11 @@ class English(Language):
 @English.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return EnglishLemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = EnglishLemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return EnglishLemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["English"]

spacy/lang/en/lemmatizer.py

@@ -3,8 +3,7 @@ from ...tokens import Token
 
 
 class EnglishLemmatizer(Lemmatizer):
-    """English lemmatizer. Only overrides is_base_form.
+    """English lemmatizer. Only overrides is_base_form."""
-    """
 
     def is_base_form(self, token: Token) -> bool:
         """

spacy/lang/es/syntax_iterators.py

@@ -58,7 +58,7 @@ def noun_bounds(
                 doc, token, np_left_deps, np_right_deps, stop_deps
             )
             filter_func = lambda t: is_verb_token(t) or t.dep in stop_deps
-            if list(filter(filter_func, doc[left_bound.i : right.i],)):
+            if list(filter(filter_func, doc[left_bound.i : right.i])):
                 break
             else:
                 right_bound = right

spacy/lang/fa/__init__.py

@@ -6,7 +6,6 @@ from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
 from .punctuation import TOKENIZER_SUFFIXES
 from .syntax_iterators import SYNTAX_ITERATORS
 from ...language import Language
-from ...lookups import Lookups
 from ...pipeline import Lemmatizer
 
 
@@ -27,18 +26,11 @@ class Persian(Language):
 @Persian.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return Lemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = Lemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return Lemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Persian"]

spacy/lang/fr/__init__.py

@@ -9,7 +9,6 @@ from .stop_words import STOP_WORDS
 from .lex_attrs import LEX_ATTRS
 from .syntax_iterators import SYNTAX_ITERATORS
 from .lemmatizer import FrenchLemmatizer
-from ...lookups import Lookups
 from ...language import Language
 
 
@@ -32,18 +31,11 @@ class French(Language):
 @French.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return FrenchLemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = FrenchLemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return FrenchLemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["French"]

spacy/lang/ja/__init__.py

@@ -2,7 +2,6 @@ from typing import Optional, Union, Dict, Any
 from pathlib import Path
 import srsly
 from collections import namedtuple
-from thinc.api import Config
 
 from .stop_words import STOP_WORDS
 from .syntax_iterators import SYNTAX_ITERATORS
@@ -16,7 +15,7 @@ from ...scorer import Scorer
 from ...symbols import POS
 from ...tokens import Doc
 from ...training import validate_examples
-from ...util import DummyTokenizer, registry
+from ...util import DummyTokenizer, registry, load_config_from_str
 from ... import util
 
 
@@ -166,7 +165,7 @@ class JapaneseTokenizer(DummyTokenizer):
 
 
 class JapaneseDefaults(Language.Defaults):
-    config = Config().from_str(DEFAULT_CONFIG)
+    config = load_config_from_str(DEFAULT_CONFIG)
     stop_words = STOP_WORDS
     syntax_iterators = SYNTAX_ITERATORS
     writing_system = {"direction": "ltr", "has_case": False, "has_letters": False}

spacy/lang/ko/__init__.py

@@ -1,5 +1,4 @@
 from typing import Optional, Any, Dict
-from thinc.api import Config
 
 from .stop_words import STOP_WORDS
 from .tag_map import TAG_MAP
@@ -10,7 +9,7 @@ from ...compat import copy_reg
 from ...scorer import Scorer
 from ...symbols import POS
 from ...training import validate_examples
-from ...util import DummyTokenizer, registry
+from ...util import DummyTokenizer, registry, load_config_from_str
 
 
 DEFAULT_CONFIG = """
@@ -70,7 +69,7 @@ class KoreanTokenizer(DummyTokenizer):
 
 
 class KoreanDefaults(Language.Defaults):
-    config = Config().from_str(DEFAULT_CONFIG)
+    config = load_config_from_str(DEFAULT_CONFIG)
     lex_attr_getters = LEX_ATTRS
     stop_words = STOP_WORDS
     writing_system = {"direction": "ltr", "has_case": False, "has_letters": False}

spacy/lang/nb/__init__.py

@@ -6,7 +6,6 @@ from .punctuation import TOKENIZER_SUFFIXES
 from .stop_words import STOP_WORDS
 from .syntax_iterators import SYNTAX_ITERATORS
 from ...language import Language
-from ...lookups import Lookups
 from ...pipeline import Lemmatizer
 
 
@@ -27,18 +26,11 @@ class Norwegian(Language):
 @Norwegian.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return Lemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = Lemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return Lemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Norwegian"]

spacy/lang/nl/__init__.py

@@ -1,5 +1,4 @@
 from typing import Optional
-
 from thinc.api import Model
 
 from .stop_words import STOP_WORDS
@@ -8,7 +7,6 @@ from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
 from .punctuation import TOKENIZER_PREFIXES, TOKENIZER_INFIXES
 from .punctuation import TOKENIZER_SUFFIXES
 from .lemmatizer import DutchLemmatizer
-from ...lookups import Lookups
 from ...language import Language
 
 
@@ -29,18 +27,11 @@ class Dutch(Language):
 @Dutch.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return DutchLemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = DutchLemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return DutchLemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Dutch"]

spacy/lang/pl/__init__.py

@@ -34,18 +34,11 @@ class Polish(Language):
 @Polish.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "pos_lookup", "lookups": None},
+    default_config={"model": None, "mode": "pos_lookup"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return PolishLemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = PolishLemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return PolishLemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Polish"]

spacy/lang/ru/__init__.py

@@ -1,5 +1,4 @@
 from typing import Optional
-
 from thinc.api import Model
 
 from .stop_words import STOP_WORDS
@@ -7,7 +6,6 @@ from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
 from .lex_attrs import LEX_ATTRS
 from .lemmatizer import RussianLemmatizer
 from ...language import Language
-from ...lookups import Lookups
 
 
 class RussianDefaults(Language.Defaults):
@@ -24,17 +22,11 @@ class Russian(Language):
 @Russian.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "pymorphy2", "lookups": None},
+    default_config={"model": None, "mode": "pymorphy2"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return RussianLemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    return RussianLemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Russian"]

spacy/lang/sv/__init__.py

@@ -5,7 +5,6 @@ from .stop_words import STOP_WORDS
 from .lex_attrs import LEX_ATTRS
 from .syntax_iterators import SYNTAX_ITERATORS
 from ...language import Language
-from ...lookups import Lookups
 from ...pipeline import Lemmatizer
 
 
@@ -30,18 +29,11 @@ class Swedish(Language):
 @Swedish.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "rule", "lookups": None},
+    default_config={"model": None, "mode": "rule"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return Lemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    lookups = Lemmatizer.load_lookups(nlp.lang, mode, lookups)
-    return Lemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Swedish"]

spacy/lang/th/__init__.py

@@ -1,10 +1,8 @@
-from thinc.api import Config
-
 from .stop_words import STOP_WORDS
 from .lex_attrs import LEX_ATTRS
 from ...language import Language
 from ...tokens import Doc
-from ...util import DummyTokenizer, registry
+from ...util import DummyTokenizer, registry, load_config_from_str
 
 
 DEFAULT_CONFIG = """
@@ -42,7 +40,7 @@ class ThaiTokenizer(DummyTokenizer):
 
 
 class ThaiDefaults(Language.Defaults):
-    config = Config().from_str(DEFAULT_CONFIG)
+    config = load_config_from_str(DEFAULT_CONFIG)
     lex_attr_getters = LEX_ATTRS
     stop_words = STOP_WORDS
 

spacy/lang/uk/__init__.py

@@ -7,7 +7,6 @@ from .stop_words import STOP_WORDS
 from .lex_attrs import LEX_ATTRS
 from .lemmatizer import UkrainianLemmatizer
 from ...language import Language
-from ...lookups import Lookups
 
 
 class UkrainianDefaults(Language.Defaults):
@@ -24,17 +23,11 @@ class Ukrainian(Language):
 @Ukrainian.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={"model": None, "mode": "pymorphy2", "lookups": None},
+    default_config={"model": None, "mode": "pymorphy2"},
     default_score_weights={"lemma_acc": 1.0},
 )
-def make_lemmatizer(
+def make_lemmatizer(nlp: Language, model: Optional[Model], name: str, mode: str):
-    nlp: Language,
+    return UkrainianLemmatizer(nlp.vocab, model, name, mode=mode)
-    model: Optional[Model],
-    name: str,
-    mode: str,
-    lookups: Optional[Lookups],
-):
-    return UkrainianLemmatizer(nlp.vocab, model, name, mode=mode, lookups=lookups)
 
 
 __all__ = ["Ukrainian"]

spacy/lang/vi/__init__.py

@@ -1,10 +1,8 @@
-from thinc.api import Config
+from .stop_words import STOP_WORDS
-
+from .lex_attrs import LEX_ATTRS
 from ...language import Language
 from ...tokens import Doc
-from .stop_words import STOP_WORDS
+from ...util import DummyTokenizer, registry, load_config_from_str
-from ...util import DummyTokenizer, registry
-from .lex_attrs import LEX_ATTRS
 
 
 DEFAULT_CONFIG = """
@@ -17,7 +15,7 @@ use_pyvi = true
 
 
 @registry.tokenizers("spacy.vi.VietnameseTokenizer")
-def create_vietnamese_tokenizer(use_pyvi: bool = True,):
+def create_vietnamese_tokenizer(use_pyvi: bool = True):
     def vietnamese_tokenizer_factory(nlp):
         return VietnameseTokenizer(nlp, use_pyvi=use_pyvi)
 
@@ -55,7 +53,7 @@ class VietnameseTokenizer(DummyTokenizer):
 
 
 class VietnameseDefaults(Language.Defaults):
-    config = Config().from_str(DEFAULT_CONFIG)
+    config = load_config_from_str(DEFAULT_CONFIG)
     lex_attr_getters = LEX_ATTRS
     stop_words = STOP_WORDS
 

spacy/lang/zh/__init__.py

@@ -1,23 +1,25 @@
-from typing import Optional, List, Dict, Any
+from typing import Optional, List, Dict, Any, Callable, Iterable
 from enum import Enum
 import tempfile
 import srsly
 import warnings
 from pathlib import Path
-from thinc.api import Config
 
 from ...errors import Warnings, Errors
 from ...language import Language
 from ...scorer import Scorer
 from ...tokens import Doc
-from ...training import validate_examples
+from ...training import validate_examples, Example
-from ...util import DummyTokenizer, registry
+from ...util import DummyTokenizer, registry, load_config_from_str
 from .lex_attrs import LEX_ATTRS
 from .stop_words import STOP_WORDS
 from ... import util
 
 
-_PKUSEG_INSTALL_MSG = "install it with `pip install pkuseg==0.0.25` or from https://github.com/lancopku/pkuseg-python"
+# fmt: off
+_PKUSEG_INSTALL_MSG = "install pkuseg and pickle5 with `pip install pkuseg==0.0.25 pickle5`"
+_PKUSEG_PICKLE_WARNING = "Failed to force pkuseg model to use pickle protocol 4. If you're saving this model with python 3.8, it may not work with python 3.6-3.7."
+# fmt: on
 
 DEFAULT_CONFIG = """
 [nlp]
@@ -25,6 +27,10 @@ DEFAULT_CONFIG = """
 [nlp.tokenizer]
 @tokenizers = "spacy.zh.ChineseTokenizer"
 segmenter = "char"
+
+[initialize]
+
+[initialize.tokenizer]
 pkuseg_model = null
 pkuseg_user_dict = "default"
 """
@@ -41,41 +47,23 @@ class Segmenter(str, Enum):
 
 
 @registry.tokenizers("spacy.zh.ChineseTokenizer")
-def create_chinese_tokenizer(
+def create_chinese_tokenizer(segmenter: Segmenter = Segmenter.char,):
-    segmenter: Segmenter = Segmenter.char,
-    pkuseg_model: Optional[str] = None,
-    pkuseg_user_dict: Optional[str] = "default",
-):
     def chinese_tokenizer_factory(nlp):
-        return ChineseTokenizer(
+        return ChineseTokenizer(nlp, segmenter=segmenter)
-            nlp,
-            segmenter=segmenter,
-            pkuseg_model=pkuseg_model,
-            pkuseg_user_dict=pkuseg_user_dict,
-        )
 
     return chinese_tokenizer_factory
 
 
 class ChineseTokenizer(DummyTokenizer):
     def __init__(
-        self,
+        self, nlp: Language, segmenter: Segmenter = Segmenter.char,
-        nlp: Language,
-        segmenter: Segmenter = Segmenter.char,
-        pkuseg_model: Optional[str] = None,
-        pkuseg_user_dict: Optional[str] = None,
     ):
         self.vocab = nlp.vocab
-        if isinstance(segmenter, Segmenter):  # we might have the Enum here
+        if isinstance(segmenter, Segmenter):
             segmenter = segmenter.value
         self.segmenter = segmenter
-        self.pkuseg_model = pkuseg_model
-        self.pkuseg_user_dict = pkuseg_user_dict
         self.pkuseg_seg = None
         self.jieba_seg = None
-        self.configure_segmenter(segmenter)
-
-    def configure_segmenter(self, segmenter: str):
         if segmenter not in Segmenter.values():
             warn_msg = Warnings.W103.format(
                 lang="Chinese",
@@ -85,11 +73,20 @@ class ChineseTokenizer(DummyTokenizer):
             )
             warnings.warn(warn_msg)
             self.segmenter = Segmenter.char
-        self.jieba_seg = try_jieba_import(self.segmenter)
+        if segmenter == Segmenter.jieba:
+            self.jieba_seg = try_jieba_import()
+
+    def initialize(
+        self,
+        get_examples: Optional[Callable[[], Iterable[Example]]] = None,
+        *,
+        nlp: Optional[Language] = None,
+        pkuseg_model: Optional[str] = None,
+        pkuseg_user_dict: str = "default",
+    ):
+        if self.segmenter == Segmenter.pkuseg:
             self.pkuseg_seg = try_pkuseg_import(
-            self.segmenter,
+                pkuseg_model=pkuseg_model, pkuseg_user_dict=pkuseg_user_dict,
-            pkuseg_model=self.pkuseg_model,
-            pkuseg_user_dict=self.pkuseg_user_dict,
             )
 
     def __call__(self, text: str) -> Doc:
@@ -145,14 +142,10 @@ class ChineseTokenizer(DummyTokenizer):
     def _get_config(self) -> Dict[str, Any]:
         return {
             "segmenter": self.segmenter,
-            "pkuseg_model": self.pkuseg_model,
-            "pkuseg_user_dict": self.pkuseg_user_dict,
         }
 
     def _set_config(self, config: Dict[str, Any] = {}) -> None:
         self.segmenter = config.get("segmenter", Segmenter.char)
-        self.pkuseg_model = config.get("pkuseg_model", None)
-        self.pkuseg_user_dict = config.get("pkuseg_user_dict", "default")
 
     def to_bytes(self, **kwargs):
         pkuseg_features_b = b""
@@ -163,6 +156,22 @@ class ChineseTokenizer(DummyTokenizer):
                 self.pkuseg_seg.feature_extractor.save(tempdir)
                 self.pkuseg_seg.model.save(tempdir)
                 tempdir = Path(tempdir)
+                # pkuseg saves features.pkl with pickle.HIGHEST_PROTOCOL, which
+                # means that it will be saved with pickle protocol 5 with
+                # python 3.8, which can't be reloaded with python 3.6-3.7.
+                # To try to make the model compatible with python 3.6+, reload
+                # the data with pickle5 and convert it back to protocol 4.
+                try:
+                    import pickle5
+
+                    with open(tempdir / "features.pkl", "rb") as fileh:
+                        features = pickle5.load(fileh)
+                    with open(tempdir / "features.pkl", "wb") as fileh:
+                        pickle5.dump(features, fileh, protocol=4)
+                except ImportError as e:
+                    raise e
+                except Exception:
+                    warnings.warn(_PKUSEG_PICKLE_WARNING)
                 with open(tempdir / "features.pkl", "rb") as fileh:
                     pkuseg_features_b = fileh.read()
                 with open(tempdir / "weights.npz", "rb") as fileh:
@@ -235,6 +244,18 @@ class ChineseTokenizer(DummyTokenizer):
                     path.mkdir(parents=True)
                 self.pkuseg_seg.model.save(path)
                 self.pkuseg_seg.feature_extractor.save(path)
+                # try to convert features.pkl to pickle protocol 4
+                try:
+                    import pickle5
+
+                    with open(path / "features.pkl", "rb") as fileh:
+                        features = pickle5.load(fileh)
+                    with open(path / "features.pkl", "wb") as fileh:
+                        pickle5.dump(features, fileh, protocol=4)
+                except ImportError as e:
+                    raise e
+                except Exception:
+                    warnings.warn(_PKUSEG_PICKLE_WARNING)
 
         def save_pkuseg_processors(path):
             if self.pkuseg_seg:
@@ -291,7 +312,7 @@ class ChineseTokenizer(DummyTokenizer):
 
 
 class ChineseDefaults(Language.Defaults):
-    config = Config().from_str(DEFAULT_CONFIG)
+    config = load_config_from_str(DEFAULT_CONFIG)
     lex_attr_getters = LEX_ATTRS
     stop_words = STOP_WORDS
     writing_system = {"direction": "ltr", "has_case": False, "has_letters": False}
@@ -302,17 +323,15 @@ class Chinese(Language):
     Defaults = ChineseDefaults
 
 
-def try_jieba_import(segmenter: str) -> None:
+def try_jieba_import() -> None:
     try:
         import jieba
 
-        if segmenter == Segmenter.jieba:
         # segment a short text to have jieba initialize its cache in advance
         list(jieba.cut("作为", cut_all=False))
 
         return jieba
     except ImportError:
-        if segmenter == Segmenter.jieba:
         msg = (
             "Jieba not installed. To use jieba, install it with `pip "
             " install jieba` or from https://github.com/fxsjy/jieba"
@@ -320,27 +339,15 @@ def try_jieba_import(segmenter: str) -> None:
         raise ImportError(msg) from None
 
 
-def try_pkuseg_import(segmenter: str, pkuseg_model: str, pkuseg_user_dict: str) -> None:
+def try_pkuseg_import(pkuseg_model: str, pkuseg_user_dict: str) -> None:
     try:
         import pkuseg
 
-        if pkuseg_model:
         return pkuseg.pkuseg(pkuseg_model, pkuseg_user_dict)
-        elif segmenter == Segmenter.pkuseg:
-            msg = (
-                "The Chinese word segmenter is 'pkuseg' but no pkuseg model "
-                "was specified. Please provide the name of a pretrained model "
-                "or the path to a model with:\n"
-                'cfg = {"nlp": {"tokenizer": {"segmenter": "pkuseg", "pkuseg_model": name_or_path }}\n'
-                "nlp = Chinese.from_config(cfg)"
-            )
-            raise ValueError(msg)
     except ImportError:
-        if segmenter == Segmenter.pkuseg:
         msg = "pkuseg not installed. To use pkuseg, " + _PKUSEG_INSTALL_MSG
         raise ImportError(msg) from None
     except FileNotFoundError:
-        if segmenter == Segmenter.pkuseg:
         msg = "Unable to load pkuseg model from: " + pkuseg_model
         raise FileNotFoundError(msg) from None
 

spacy/language.py

@@ -8,7 +8,7 @@ from contextlib import contextmanager
 from copy import deepcopy
 from pathlib import Path
 import warnings
-from thinc.api import Model, get_current_ops, Config, require_gpu, Optimizer
+from thinc.api import Model, get_current_ops, Config, Optimizer
 import srsly
 import multiprocessing as mp
 from itertools import chain, cycle
@@ -18,8 +18,9 @@ from .tokens.underscore import Underscore
 from .vocab import Vocab, create_vocab
 from .pipe_analysis import validate_attrs, analyze_pipes, print_pipe_analysis
 from .training import Example, validate_examples
+from .training.initialize import init_vocab, init_tok2vec
 from .scorer import Scorer
-from .util import create_default_optimizer, registry, SimpleFrozenList
+from .util import registry, SimpleFrozenList
 from .util import SimpleFrozenDict, combine_score_weights, CONFIG_SECTION_ORDER
 from .lang.tokenizer_exceptions import URL_MATCH, BASE_EXCEPTIONS
 from .lang.punctuation import TOKENIZER_PREFIXES, TOKENIZER_SUFFIXES
@@ -27,7 +28,8 @@ from .lang.punctuation import TOKENIZER_INFIXES
 from .tokens import Doc
 from .tokenizer import Tokenizer
 from .errors import Errors, Warnings
-from .schemas import ConfigSchema, ConfigSchemaNlp
+from .schemas import ConfigSchema, ConfigSchemaNlp, ConfigSchemaInit
+from .schemas import ConfigSchemaPretrain, validate_init_settings
 from .git_info import GIT_VERSION
 from . import util
 from . import about
@@ -1066,7 +1068,7 @@ class Language:
         validate_examples(examples, "Language.update")
         if sgd is None:
             if self._optimizer is None:
-                self._optimizer = create_default_optimizer()
+                self._optimizer = self.create_optimizer()
             sgd = self._optimizer
         if component_cfg is None:
             component_cfg = {}
@@ -1124,7 +1126,7 @@ class Language:
         validate_examples(examples, "Language.rehearse")
         if sgd is None:
             if self._optimizer is None:
-                self._optimizer = create_default_optimizer()
+                self._optimizer = self.create_optimizer()
             sgd = self._optimizer
         pipes = list(self.pipeline)
         random.shuffle(pipes)
@@ -1154,61 +1156,73 @@ class Language:
         get_examples: Optional[Callable[[], Iterable[Example]]] = None,
         *,
         sgd: Optional[Optimizer] = None,
-        device: int = -1,
+    ) -> Optimizer:
+        warnings.warn(Warnings.W089, DeprecationWarning)
+        return self.initialize(get_examples, sgd=sgd)
+
+    def initialize(
+        self,
+        get_examples: Optional[Callable[[], Iterable[Example]]] = None,
+        *,
+        sgd: Optional[Optimizer] = None,
     ) -> Optimizer:
         """Initialize the pipe for training, using data examples if available.
 
         get_examples (Callable[[], Iterable[Example]]): Optional function that
             returns gold-standard Example objects.
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
+        sgd (Optional[Optimizer]): An optimizer to use for updates. If not
-            create_optimizer if it doesn't exist.
+            provided, will be created using the .create_optimizer() method.
         RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/language#begin_training
+        DOCS: https://nightly.spacy.io/api/language#initialize
         """
         if get_examples is None:
             util.logger.debug(
-                "No 'get_examples' callback provided to 'Language.begin_training', creating dummy examples"
+                "No 'get_examples' callback provided to 'Language.initialize', creating dummy examples"
             )
             doc = Doc(self.vocab, words=["x", "y", "z"])
             get_examples = lambda: [Example.from_dict(doc, {})]
-        # Populate vocab
         if not hasattr(get_examples, "__call__"):
             err = Errors.E930.format(name="Language", obj=type(get_examples))
             raise ValueError(err)
-        valid_examples = False
+        # Make sure the config is interpolated so we can resolve subsections
-        for example in get_examples():
+        config = self.config.interpolate()
-            if not isinstance(example, Example):
+        # These are the settings provided in the [initialize] block in the config
-                err = Errors.E978.format(
+        I = registry.resolve(config["initialize"], schema=ConfigSchemaInit)
-                    name="Language.begin_training", types=type(example)
+        init_vocab(
+            self, data=I["vocab_data"], lookups=I["lookups"], vectors=I["vectors"]
         )
-                raise ValueError(err)
+        pretrain_cfg = config.get("pretraining")
-            else:
+        if pretrain_cfg:
-                valid_examples = True
+            P = registry.resolve(pretrain_cfg, schema=ConfigSchemaPretrain)
-            for word in [t.text for t in example.reference]:
+            init_tok2vec(self, P, I)
-                _ = self.vocab[word]  # noqa: F841
-        if not valid_examples:
-            err = Errors.E930.format(name="Language", obj="empty list")
-            raise ValueError(err)
-        if device >= 0:  # TODO: do we need this here?
-            require_gpu(device)
         if self.vocab.vectors.data.shape[1] >= 1:
             ops = get_current_ops()
             self.vocab.vectors.data = ops.asarray(self.vocab.vectors.data)
-        if sgd is None:
+        if hasattr(self.tokenizer, "initialize"):
-            sgd = create_default_optimizer()
+            tok_settings = validate_init_settings(
-        self._optimizer = sgd
+                self.tokenizer.initialize,
-        for name, proc in self.pipeline:
+                I["tokenizer"],
-            if hasattr(proc, "begin_training"):
+                section="tokenizer",
-                proc.begin_training(
+                name="tokenizer",
-                    get_examples, pipeline=self.pipeline, sgd=self._optimizer
             )
+            self.tokenizer.initialize(get_examples, nlp=self, **tok_settings)
+        for name, proc in self.pipeline:
+            if hasattr(proc, "initialize"):
+                p_settings = I["components"].get(name, {})
+                p_settings = validate_init_settings(
+                    proc.initialize, p_settings, section="components", name=name
+                )
+                proc.initialize(get_examples, nlp=self, **p_settings)
         self._link_components()
+        self._optimizer = sgd
+        if sgd is not None:
+            self._optimizer = sgd
+        elif self._optimizer is None:
+            self._optimizer = self.create_optimizer()
         return self._optimizer
 
-    def resume_training(
+    def resume_training(self, *, sgd: Optional[Optimizer] = None) -> Optimizer:
-        self, *, sgd: Optional[Optimizer] = None, device: int = -1
-    ) -> Optimizer:
         """Continue training a pretrained model.
 
         Create and return an optimizer, and initialize "rehearsal" for any pipeline
@@ -1217,22 +1231,20 @@ class Language:
         rehearsal, collect samples of text you want the models to retain performance
         on, and call nlp.rehearse() with a batch of Example objects.
 
-        sgd (Optional[Optimizer]): An optimizer.
         RETURNS (Optimizer): The optimizer.
 
         DOCS: https://nightly.spacy.io/api/language#resume_training
         """
-        if device >= 0:  # TODO: do we need this here?
-            require_gpu(device)
         ops = get_current_ops()
         if self.vocab.vectors.data.shape[1] >= 1:
             self.vocab.vectors.data = ops.asarray(self.vocab.vectors.data)
-        if sgd is None:
-            sgd = create_default_optimizer()
-        self._optimizer = sgd
         for name, proc in self.pipeline:
             if hasattr(proc, "_rehearsal_model"):
                 proc._rehearsal_model = deepcopy(proc.model)
+        if sgd is not None:
+            self._optimizer = sgd
+        elif self._optimizer is None:
+            self._optimizer = self.create_optimizer()
         return self._optimizer
 
     def evaluate(
@@ -1294,6 +1306,11 @@ class Language:
         results["speed"] = n_words / (end_time - start_time)
         return results
 
+    def create_optimizer(self):
+        """Create an optimizer, usually using the [training.optimizer] config."""
+        subconfig = {"optimizer": self.config["training"]["optimizer"]}
+        return registry.resolve(subconfig)["optimizer"]
+
     @contextmanager
     def use_params(self, params: Optional[dict]):
         """Replace weights of models in the pipeline with those provided in the
@@ -1502,7 +1519,7 @@ class Language:
             ).merge(config)
         if "nlp" not in config:
             raise ValueError(Errors.E985.format(config=config))
-        config_lang = config["nlp"]["lang"]
+        config_lang = config["nlp"].get("lang")
         if config_lang is not None and config_lang != cls.lang:
             raise ValueError(
                 Errors.E958.format(

spacy/ml/featureextractor.py

@@ -0,0 +1,28 @@
+from typing import List, Union, Callable, Tuple
+from thinc.types import Ints2d
+from thinc.api import Model, registry
+
+from ..tokens import Doc
+
+
+@registry.layers("spacy.FeatureExtractor.v1")
+def FeatureExtractor(columns: List[Union[int, str]]) -> Model[List[Doc], List[Ints2d]]:
+    return Model("extract_features", forward, attrs={"columns": columns})
+
+
+def forward(
+    model: Model[List[Doc], List[Ints2d]], docs, is_train: bool
+) -> Tuple[List[Ints2d], Callable]:
+    columns = model.attrs["columns"]
+    features: List[Ints2d] = []
+    for doc in docs:
+        if hasattr(doc, "to_array"):
+            attrs = doc.to_array(columns)
+        else:
+            attrs = doc.doc.to_array(columns)[doc.start : doc.end]
+        if attrs.ndim == 1:
+            attrs = attrs.reshape((attrs.shape[0], 1))
+        features.append(model.ops.asarray2i(attrs, dtype="uint64"))
+
+    backprop: Callable[[List[Ints2d]], List] = lambda d_features: []
+    return features, backprop

spacy/ml/models/textcat.py

@@ -3,12 +3,13 @@ from thinc.api import Model, reduce_mean, Linear, list2ragged, Logistic
 from thinc.api import chain, concatenate, clone, Dropout, ParametricAttention
 from thinc.api import SparseLinear, Softmax, softmax_activation, Maxout, reduce_sum
 from thinc.api import HashEmbed, with_array, with_cpu, uniqued
-from thinc.api import Relu, residual, expand_window, FeatureExtractor
+from thinc.api import Relu, residual, expand_window
 
 from ...attrs import ID, ORTH, PREFIX, SUFFIX, SHAPE, LOWER
 from ...util import registry
 from ..extract_ngrams import extract_ngrams
 from ..staticvectors import StaticVectors
+from ..featureextractor import FeatureExtractor
 
 
 @registry.architectures.register("spacy.TextCatCNN.v1")

spacy/ml/models/tok2vec.py

@@ -1,16 +1,16 @@
-from typing import Optional, List
+from typing import Optional, List, Union
-from thinc.api import chain, clone, concatenate, with_array, with_padded
-from thinc.api import Model, noop, list2ragged, ragged2list
-from thinc.api import FeatureExtractor, HashEmbed
-from thinc.api import expand_window, residual, Maxout, Mish, PyTorchLSTM
 from thinc.types import Floats2d
+from thinc.api import chain, clone, concatenate, with_array, with_padded
+from thinc.api import Model, noop, list2ragged, ragged2list, HashEmbed
+from thinc.api import expand_window, residual, Maxout, Mish, PyTorchLSTM
 
 from ...tokens import Doc
 from ...util import registry
 from ...ml import _character_embed
 from ..staticvectors import StaticVectors
+from ..featureextractor import FeatureExtractor
 from ...pipeline.tok2vec import Tok2VecListener
-from ...attrs import ORTH, NORM, PREFIX, SUFFIX, SHAPE
+from ...attrs import ORTH, LOWER, PREFIX, SUFFIX, SHAPE, intify_attr
 
 
 @registry.architectures.register("spacy.Tok2VecListener.v1")
@@ -98,7 +98,7 @@ def MultiHashEmbed(
     attributes using hash embedding, concatenates the results, and passes it
     through a feed-forward subnetwork to build a mixed representations.
 
-    The features used are the NORM, PREFIX, SUFFIX and SHAPE, which can have
+    The features used are the LOWER, PREFIX, SUFFIX and SHAPE, which can have
     varying definitions depending on the Vocab of the Doc object passed in.
     Vectors from pretrained static vectors can also be incorporated into the
     concatenated representation.
@@ -115,7 +115,7 @@ def MultiHashEmbed(
     also_use_static_vectors (bool): Whether to also use static word vectors.
         Requires a vectors table to be loaded in the Doc objects' vocab.
     """
-    cols = [NORM, PREFIX, SUFFIX, SHAPE, ORTH]
+    cols = [LOWER, PREFIX, SUFFIX, SHAPE, ORTH]
     seed = 7
 
     def make_hash_embed(feature):
@@ -123,7 +123,7 @@ def MultiHashEmbed(
         seed += 1
         return HashEmbed(
             width,
-            rows if feature == NORM else rows // 2,
+            rows if feature == LOWER else rows // 2,
             column=cols.index(feature),
             seed=seed,
             dropout=0.0,
@@ -131,13 +131,13 @@ def MultiHashEmbed(
 
     if also_embed_subwords:
         embeddings = [
-            make_hash_embed(NORM),
+            make_hash_embed(LOWER),
             make_hash_embed(PREFIX),
             make_hash_embed(SUFFIX),
             make_hash_embed(SHAPE),
         ]
     else:
-        embeddings = [make_hash_embed(NORM)]
+        embeddings = [make_hash_embed(LOWER)]
     concat_size = width * (len(embeddings) + also_use_static_vectors)
     if also_use_static_vectors:
         model = chain(
@@ -165,7 +165,8 @@ def MultiHashEmbed(
 
 @registry.architectures.register("spacy.CharacterEmbed.v1")
 def CharacterEmbed(
-    width: int, rows: int, nM: int, nC: int, also_use_static_vectors: bool
+    width: int, rows: int, nM: int, nC: int, also_use_static_vectors: bool,
+    feature: Union[int, str]="LOWER"
 ) -> Model[List[Doc], List[Floats2d]]:
     """Construct an embedded representation based on character embeddings, using
     a feed-forward network. A fixed number of UTF-8 byte characters are used for
@@ -179,12 +180,13 @@ def CharacterEmbed(
     of being in an arbitrary position depending on the word length.
 
     The characters are embedded in a embedding table with a given number of rows,
-    and the vectors concatenated. A hash-embedded vector of the NORM of the word is
+    and the vectors concatenated. A hash-embedded vector of the LOWER of the word is
     also concatenated on, and the result is then passed through a feed-forward
     network to construct a single vector to represent the information.
 
-    width (int): The width of the output vector and the NORM hash embedding.
+    feature (int or str): An attribute to embed, to concatenate with the characters.
-    rows (int): The number of rows in the NORM hash embedding table.
+    width (int): The width of the output vector and the feature embedding.
+    rows (int): The number of rows in the LOWER hash embedding table.
     nM (int): The dimensionality of the character embeddings. Recommended values
         are between 16 and 64.
     nC (int): The number of UTF-8 bytes to embed per word. Recommended values
@@ -193,12 +195,15 @@ def CharacterEmbed(
     also_use_static_vectors (bool): Whether to also use static word vectors.
         Requires a vectors table to be loaded in the Doc objects' vocab.
     """
+    feature = intify_attr(feature)
+    if feature is None:
+        raise ValueError("Invalid feature: Must be a token attribute.")
     if also_use_static_vectors:
         model = chain(
             concatenate(
                 chain(_character_embed.CharacterEmbed(nM=nM, nC=nC), list2ragged()),
                 chain(
-                    FeatureExtractor([NORM]),
+                    FeatureExtractor([feature]),
                     list2ragged(),
                     with_array(HashEmbed(nO=width, nV=rows, column=0, seed=5)),
                 ),
@@ -214,7 +219,7 @@ def CharacterEmbed(
             concatenate(
                 chain(_character_embed.CharacterEmbed(nM=nM, nC=nC), list2ragged()),
                 chain(
-                    FeatureExtractor([NORM]),
+                    FeatureExtractor([feature]),
                     list2ragged(),
                     with_array(HashEmbed(nO=width, nV=rows, column=0, seed=5)),
                 ),

spacy/pipe_analysis.py

@@ -78,7 +78,7 @@ def get_attr_info(nlp: "Language", attr: str) -> Dict[str, List[str]]:
 
 
 def analyze_pipes(
-    nlp: "Language", *, keys: List[str] = DEFAULT_KEYS,
+    nlp: "Language", *, keys: List[str] = DEFAULT_KEYS
 ) -> Dict[str, Union[List[str], Dict[str, List[str]]]]:
     """Print a formatted summary for the current nlp object's pipeline. Shows
     a table with the pipeline components and why they assign and require, as

spacy/pipeline/attributeruler.py

@@ -82,8 +82,7 @@ class AttributeRuler(Pipe):
         matches = self.matcher(doc, allow_missing=True)
         # Sort by the attribute ID, so that later rules have precendence
         matches = [
-            (int(self.vocab.strings[m_id]), m_id, s, e)
+            (int(self.vocab.strings[m_id]), m_id, s, e) for m_id, s, e in matches
-            for m_id, s, e in matches
         ]
         matches.sort()
         for attr_id, match_id, start, end in matches:
@@ -93,7 +92,7 @@ class AttributeRuler(Pipe):
             try:
                 # The index can be negative, which makes it annoying to do
                 # the boundscheck. Let Span do it instead.
-                token = span[index]
+                token = span[index]  # noqa: F841
             except IndexError:
                 # The original exception is just our conditional logic, so we
                 # raise from.

spacy/pipeline/dep_parser.pyx

@@ -126,13 +126,13 @@ cdef class DependencyParser(Parser):
     def add_multitask_objective(self, mt_component):
         self._multitasks.append(mt_component)
 
-    def init_multitask_objectives(self, get_examples, pipeline, sgd=None, **cfg):
+    def init_multitask_objectives(self, get_examples, nlp=None, **cfg):
         # TODO: transfer self.model.get_ref("tok2vec") to the multitask's model ?
         for labeller in self._multitasks:
             labeller.model.set_dim("nO", len(self.labels))
             if labeller.model.has_ref("output_layer"):
                 labeller.model.get_ref("output_layer").set_dim("nO", len(self.labels))
-            labeller.begin_training(get_examples, pipeline=pipeline, sgd=sgd)
+            labeller.initialize(get_examples, nlp=nlp)
 
     @property
     def labels(self):

spacy/pipeline/entity_linker.py

@@ -1,5 +1,5 @@
 from itertools import islice
-from typing import Optional, Iterable, Callable, Dict, Iterator, Union, List, Tuple
+from typing import Optional, Iterable, Callable, Dict, Iterator, Union, List
 from pathlib import Path
 import srsly
 import random
@@ -140,26 +140,20 @@ class EntityLinker(Pipe):
         if len(self.kb) == 0:
             raise ValueError(Errors.E139.format(name=self.name))
 
-    def begin_training(
+    def initialize(
         self,
         get_examples: Callable[[], Iterable[Example]],
         *,
-        pipeline: Optional[List[Tuple[str, Callable[[Doc], Doc]]]] = None,
+        nlp: Optional[Language] = None,
-        sgd: Optional[Optimizer] = None,
+    ):
-    ) -> Optimizer:
         """Initialize the pipe for training, using a representative set
         of data examples.
 
         get_examples (Callable[[], Iterable[Example]]): Function that
             returns a representative sample of gold-standard Example objects.
-        pipeline (List[Tuple[str, Callable]]): Optional list of pipeline
+        nlp (Language): The current nlp object the component is part of.
-            components that this component is part of. Corresponds to
-            nlp.pipeline.
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
-            create_optimizer if it doesn't exist.
-        RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/entitylinker#begin_training
+        DOCS: https://nightly.spacy.io/api/entitylinker#initialize
         """
         self._ensure_examples(get_examples)
         self._require_kb()
@@ -174,9 +168,6 @@ class EntityLinker(Pipe):
         self.model.initialize(
             X=doc_sample, Y=self.model.ops.asarray(vector_sample, dtype="float32")
         )
-        if sgd is None:
-            sgd = self.create_optimizer()
-        return sgd
 
     def update(
         self,

spacy/pipeline/lemmatizer.py

@@ -1,26 +1,25 @@
-from typing import Optional, List, Dict, Any
+from typing import Optional, List, Dict, Any, Callable, Iterable, Iterator, Union
+from typing import Tuple
 from thinc.api import Model
+from pathlib import Path
 
 from .pipe import Pipe
 from ..errors import Errors
 from ..language import Language
+from ..training import Example
 from ..lookups import Lookups, load_lookups
 from ..scorer import Scorer
 from ..tokens import Doc, Token
 from ..vocab import Vocab
 from ..training import validate_examples
+from ..util import logger, SimpleFrozenList
 from .. import util
 
 
 @Language.factory(
     "lemmatizer",
     assigns=["token.lemma"],
-    default_config={
+    default_config={"model": None, "mode": "lookup", "overwrite": False},
-        "model": None,
-        "mode": "lookup",
-        "lookups": None,
-        "overwrite": False,
-    },
     default_score_weights={"lemma_acc": 1.0},
 )
 def make_lemmatizer(
@@ -28,13 +27,9 @@ def make_lemmatizer(
     model: Optional[Model],
     name: str,
     mode: str,
-    lookups: Optional[Lookups],
     overwrite: bool = False,
 ):
-    lookups = Lemmatizer.load_lookups(nlp.lang, mode, lookups)
+    return Lemmatizer(nlp.vocab, model, name, mode=mode, overwrite=overwrite)
-    return Lemmatizer(
-        nlp.vocab, model, name, mode=mode, lookups=lookups, overwrite=overwrite
-    )
 
 
 class Lemmatizer(Pipe):
@@ -46,59 +41,19 @@ class Lemmatizer(Pipe):
     """
 
     @classmethod
-    def get_lookups_config(cls, mode: str) -> Dict:
+    def get_lookups_config(cls, mode: str) -> Tuple[List[str], List[str]]:
         """Returns the lookups configuration settings for a given mode for use
         in Lemmatizer.load_lookups.
 
         mode (str): The lemmatizer mode.
-        RETURNS (dict): The lookups configuration settings for this mode.
+        RETURNS (Tuple[List[str], List[str]]): The required and optional
-
+            lookup tables for this mode.
-        DOCS: https://nightly.spacy.io/api/lemmatizer#get_lookups_config
         """
         if mode == "lookup":
-            return {
+            return (["lemma_lookup"], [])
-                "required_tables": ["lemma_lookup"],
-            }
         elif mode == "rule":
-            return {
+            return (["lemma_rules"], ["lemma_exc", "lemma_index"])
-                "required_tables": ["lemma_rules"],
+        return ([], [])
-                "optional_tables": ["lemma_exc", "lemma_index"],
-            }
-        return {}
-
-    @classmethod
-    def load_lookups(cls, lang: str, mode: str, lookups: Optional[Lookups],) -> Lookups:
-        """Load and validate lookups tables. If the provided lookups is None,
-        load the default lookups tables according to the language and mode
-        settings. Confirm that all required tables for the language and mode
-        are present.
-
-        lang (str): The language code.
-        mode (str): The lemmatizer mode.
-        lookups (Lookups): The provided lookups, may be None if the default
-            lookups should be loaded.
-        RETURNS (Lookups): The Lookups object.
-
-        DOCS: https://nightly.spacy.io/api/lemmatizer#get_lookups_config
-        """
-        config = cls.get_lookups_config(mode)
-        required_tables = config.get("required_tables", [])
-        optional_tables = config.get("optional_tables", [])
-        if lookups is None:
-            lookups = load_lookups(lang=lang, tables=required_tables)
-            optional_lookups = load_lookups(
-                lang=lang, tables=optional_tables, strict=False
-            )
-            for table in optional_lookups.tables:
-                lookups.set_table(table, optional_lookups.get_table(table))
-        for table in required_tables:
-            if table not in lookups:
-                raise ValueError(
-                    Errors.E1004.format(
-                        mode=mode, tables=required_tables, found=lookups.tables
-                    )
-                )
-        return lookups
 
     def __init__(
         self,
@@ -107,7 +62,6 @@ class Lemmatizer(Pipe):
         name: str = "lemmatizer",
         *,
         mode: str = "lookup",
-        lookups: Optional[Lookups] = None,
         overwrite: bool = False,
     ) -> None:
         """Initialize a Lemmatizer.
@@ -116,9 +70,6 @@ class Lemmatizer(Pipe):
         model (Model): A model (not yet implemented).
         name (str): The component name. Defaults to "lemmatizer".
         mode (str): The lemmatizer mode: "lookup", "rule". Defaults to "lookup".
-        lookups (Lookups): The lookups object containing the (optional) tables
-            such as "lemma_rules", "lemma_index", "lemma_exc" and
-            "lemma_lookup". Defaults to None
         overwrite (bool): Whether to overwrite existing lemmas. Defaults to
             `False`.
 
@@ -128,8 +79,9 @@ class Lemmatizer(Pipe):
         self.model = model
         self.name = name
         self._mode = mode
-        self.lookups = lookups if lookups is not None else Lookups()
+        self.lookups = Lookups()
         self.overwrite = overwrite
+        self._validated = False
         if self.mode == "lookup":
             self.lemmatize = self.lookup_lemmatize
         elif self.mode == "rule":
@@ -153,12 +105,56 @@ class Lemmatizer(Pipe):
 
         DOCS: https://nightly.spacy.io/api/lemmatizer#call
         """
+        if not self._validated:
+            self._validate_tables(Errors.E1004)
         for token in doc:
             if self.overwrite or token.lemma == 0:
                 token.lemma_ = self.lemmatize(token)[0]
         return doc
 
-    def pipe(self, stream, *, batch_size=128):
+    def initialize(
+        self,
+        get_examples: Optional[Callable[[], Iterable[Example]]] = None,
+        *,
+        nlp: Optional[Language] = None,
+        lookups: Optional[Lookups] = None,
+    ):
+        """Initialize the lemmatizer and load in data.
+
+        get_examples (Callable[[], Iterable[Example]]): Function that
+            returns a representative sample of gold-standard Example objects.
+        nlp (Language): The current nlp object the component is part of.
+        lookups (Lookups): The lookups object containing the (optional) tables
+            such as "lemma_rules", "lemma_index", "lemma_exc" and
+            "lemma_lookup". Defaults to None.
+        """
+        required_tables, optional_tables = self.get_lookups_config(self.mode)
+        if lookups is None:
+            logger.debug("Lemmatizer: loading tables from spacy-lookups-data")
+            lookups = load_lookups(lang=self.vocab.lang, tables=required_tables)
+            optional_lookups = load_lookups(
+                lang=self.vocab.lang, tables=optional_tables, strict=False
+            )
+            for table in optional_lookups.tables:
+                lookups.set_table(table, optional_lookups.get_table(table))
+        self.lookups = lookups
+        self._validate_tables(Errors.E1004)
+
+    def _validate_tables(self, error_message: str = Errors.E912) -> None:
+        """Check that the lookups are correct for the current mode."""
+        required_tables, optional_tables = self.get_lookups_config(self.mode)
+        for table in required_tables:
+            if table not in self.lookups:
+                raise ValueError(
+                    error_message.format(
+                        mode=self.mode,
+                        tables=required_tables,
+                        found=self.lookups.tables,
+                    )
+                )
+        self._validated = True
+
+    def pipe(self, stream: Iterable[Doc], *, batch_size: int = 128) -> Iterator[Doc]:
         """Apply the pipe to a stream of documents. This usually happens under
         the hood when the nlp object is called on a text and all components are
         applied to the Doc.
@@ -263,7 +259,7 @@ class Lemmatizer(Pipe):
         """
         return False
 
-    def score(self, examples, **kwargs) -> Dict[str, Any]:
+    def score(self, examples: Iterable[Example], **kwargs) -> Dict[str, Any]:
         """Score a batch of examples.
 
         examples (Iterable[Example]): The examples to score.
@@ -274,58 +270,66 @@ class Lemmatizer(Pipe):
         validate_examples(examples, "Lemmatizer.score")
         return Scorer.score_token_attr(examples, "lemma", **kwargs)
 
-    def to_disk(self, path, *, exclude=tuple()):
+    def to_disk(
-        """Save the current state to a directory.
+        self, path: Union[str, Path], *, exclude: Iterable[str] = SimpleFrozenList()
+    ):
+        """Serialize the pipe to disk.
 
-        path (unicode or Path): A path to a directory, which will be created if
+        path (str / Path): Path to a directory.
-            it doesn't exist.
+        exclude (Iterable[str]): String names of serialization fields to exclude.
-        exclude (list): String names of serialization fields to exclude.
 
-        DOCS: https://nightly.spacy.io/api/vocab#to_disk
+        DOCS: https://nightly.spacy.io/api/lemmatizer#to_disk
         """
         serialize = {}
         serialize["vocab"] = lambda p: self.vocab.to_disk(p)
         serialize["lookups"] = lambda p: self.lookups.to_disk(p)
         util.to_disk(path, serialize, exclude)
 
-    def from_disk(self, path, *, exclude=tuple()):
+    def from_disk(
-        """Loads state from a directory. Modifies the object in place and
+        self, path: Union[str, Path], *, exclude: Iterable[str] = SimpleFrozenList()
-        returns it.
+    ) -> "Lemmatizer":
+        """Load the pipe from disk. Modifies the object in place and returns it.
 
-        path (unicode or Path): A path to a directory.
+        path (str / Path): Path to a directory.
-        exclude (list): String names of serialization fields to exclude.
+        exclude (Iterable[str]): String names of serialization fields to exclude.
-        RETURNS (Vocab): The modified `Vocab` object.
+        RETURNS (Lemmatizer): The modified Lemmatizer object.
 
-        DOCS: https://nightly.spacy.io/api/vocab#to_disk
+        DOCS: https://nightly.spacy.io/api/lemmatizer#from_disk
         """
         deserialize = {}
         deserialize["vocab"] = lambda p: self.vocab.from_disk(p)
         deserialize["lookups"] = lambda p: self.lookups.from_disk(p)
         util.from_disk(path, deserialize, exclude)
+        self._validate_tables()
+        return self
 
-    def to_bytes(self, *, exclude=tuple()) -> bytes:
+    def to_bytes(self, *, exclude: Iterable[str] = SimpleFrozenList()) -> bytes:
-        """Serialize the current state to a binary string.
+        """Serialize the pipe to a bytestring.
 
-        exclude (list): String names of serialization fields to exclude.
+        exclude (Iterable[str]): String names of serialization fields to exclude.
-        RETURNS (bytes): The serialized form of the `Vocab` object.
+        RETURNS (bytes): The serialized object.
 
-        DOCS: https://nightly.spacy.io/api/vocab#to_bytes
+        DOCS: https://nightly.spacy.io/api/lemmatizer#to_bytes
         """
         serialize = {}
         serialize["vocab"] = self.vocab.to_bytes
         serialize["lookups"] = self.lookups.to_bytes
         return util.to_bytes(serialize, exclude)
 
-    def from_bytes(self, bytes_data: bytes, *, exclude=tuple()):
+    def from_bytes(
-        """Load state from a binary string.
+        self, bytes_data: bytes, *, exclude: Iterable[str] = SimpleFrozenList()
+    ) -> "Lemmatizer":
+        """Load the pipe from a bytestring.
 
-        bytes_data (bytes): The data to load from.
+        bytes_data (bytes): The serialized pipe.
-        exclude (list): String names of serialization fields to exclude.
+        exclude (Iterable[str]): String names of serialization fields to exclude.
-        RETURNS (Vocab): The `Vocab` object.
+        RETURNS (Lemmatizer): The loaded Lemmatizer.
 
-        DOCS: https://nightly.spacy.io/api/vocab#from_bytes
+        DOCS: https://nightly.spacy.io/api/lemmatizer#from_bytes
         """
         deserialize = {}
         deserialize["vocab"] = lambda b: self.vocab.from_bytes(b)
         deserialize["lookups"] = lambda b: self.lookups.from_bytes(b)
         util.from_bytes(bytes_data, deserialize, exclude)
+        self._validate_tables()
+        return self

spacy/pipeline/morphologizer.pyx

@@ -1,5 +1,5 @@
 # cython: infer_types=True, profile=True, binding=True
-from typing import Optional
+from typing import Optional, Union, Dict
 import srsly
 from thinc.api import SequenceCategoricalCrossentropy, Model, Config
 from itertools import islice
@@ -101,6 +101,11 @@ class Morphologizer(Tagger):
         """RETURNS (Tuple[str]): The labels currently added to the component."""
         return tuple(self.cfg["labels_morph"].keys())
 
+    @property
+    def label_data(self) -> Dict[str, Dict[str, Union[str, float, int, None]]]:
+        """A dictionary with all labels data."""
+        return {"morph": self.cfg["labels_morph"], "pos": self.cfg["labels_pos"]}
+
     def add_label(self, label):
         """Add a new label to the pipe.
 
@@ -129,27 +134,22 @@ class Morphologizer(Tagger):
             self.cfg["labels_pos"][norm_label] = POS_IDS[pos]
         return 1
 
-    def begin_training(self, get_examples, *, pipeline=None, sgd=None):
+    def initialize(self, get_examples, *, nlp=None):
         """Initialize the pipe for training, using a representative set
         of data examples.
 
         get_examples (Callable[[], Iterable[Example]]): Function that
             returns a representative sample of gold-standard Example objects.
-        pipeline (List[Tuple[str, Callable]]): Optional list of pipeline
+        nlp (Language): The current nlp object the component is part of.
-            components that this component is part of. Corresponds to
-            nlp.pipeline.
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
-            create_optimizer if it doesn't exist.
-        RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/morphologizer#begin_training
+        DOCS: https://nightly.spacy.io/api/morphologizer#initialize
         """
         self._ensure_examples(get_examples)
         # First, fetch all labels from the data
         for example in get_examples():
             for i, token in enumerate(example.reference):
                 pos = token.pos_
-                morph = token.morph_
+                morph = str(token.morph)
                 # create and add the combined morph+POS label
                 morph_dict = Morphology.feats_to_dict(morph)
                 if pos:
@@ -167,7 +167,7 @@ class Morphologizer(Tagger):
             gold_array = []
             for i, token in enumerate(example.reference):
                 pos = token.pos_
-                morph = token.morph_
+                morph = str(token.morph)
                 morph_dict = Morphology.feats_to_dict(morph)
                 if pos:
                     morph_dict[self.POS_FEAT] = pos
@@ -178,9 +178,6 @@ class Morphologizer(Tagger):
         assert len(doc_sample) > 0, Errors.E923.format(name=self.name)
         assert len(label_sample) > 0, Errors.E923.format(name=self.name)
         self.model.initialize(X=doc_sample, Y=label_sample)
-        if sgd is None:
-            sgd = self.create_optimizer()
-        return sgd
 
     def set_annotations(self, docs, batch_tag_ids):
         """Modify a batch of documents, using pre-computed scores.

spacy/pipeline/multitask.pyx

@@ -81,7 +81,7 @@ class MultitaskObjective(Tagger):
     def set_annotations(self, docs, dep_ids):
         pass
 
-    def begin_training(self, get_examples, pipeline=None, sgd=None):
+    def initialize(self, get_examples, nlp=None):
         if not hasattr(get_examples, "__call__"):
             err = Errors.E930.format(name="MultitaskObjective", obj=type(get_examples))
             raise ValueError(err)
@@ -91,9 +91,6 @@ class MultitaskObjective(Tagger):
                 if label is not None and label not in self.labels:
                     self.labels[label] = len(self.labels)
         self.model.initialize()   # TODO: fix initialization by defining X and Y
-        if sgd is None:
-            sgd = self.create_optimizer()
-        return sgd
 
     def predict(self, docs):
         tokvecs = self.model.get_ref("tok2vec")(docs)
@@ -177,13 +174,10 @@ class ClozeMultitask(Pipe):
     def set_annotations(self, docs, dep_ids):
         pass
 
-    def begin_training(self, get_examples, pipeline=None, sgd=None):
+    def initialize(self, get_examples, nlp=None):
         self.model.initialize()  # TODO: fix initialization by defining X and Y
         X = self.model.ops.alloc((5, self.model.get_ref("tok2vec").get_dim("nO")))
-        self.model.output_layer.begin_training(X)
+        self.model.output_layer.initialize(X)
-        if sgd is None:
-            sgd = self.create_optimizer()
-        return sgd
 
     def predict(self, docs):
         tokvecs = self.model.get_ref("tok2vec")(docs)

spacy/pipeline/ner.pyx

@@ -96,14 +96,14 @@ cdef class EntityRecognizer(Parser):
         """Register another component as a multi-task objective. Experimental."""
         self._multitasks.append(mt_component)
 
-    def init_multitask_objectives(self, get_examples, pipeline, sgd=None, **cfg):
+    def init_multitask_objectives(self, get_examples, nlp=None, **cfg):
         """Setup multi-task objective components. Experimental and internal."""
         # TODO: transfer self.model.get_ref("tok2vec") to the multitask's model ?
         for labeller in self._multitasks:
             labeller.model.set_dim("nO", len(self.labels))
             if labeller.model.has_ref("output_layer"):
                 labeller.model.get_ref("output_layer").set_dim("nO", len(self.labels))
-            labeller.begin_training(get_examples, pipeline=pipeline)
+            labeller.initialize(get_examples, nlp=nlp)
 
     @property
     def labels(self):

spacy/pipeline/pipe.pyx

@@ -1,4 +1,5 @@
 # cython: infer_types=True, profile=True
+from typing import Optional, Tuple
 import srsly
 from thinc.api import set_dropout_rate, Model
 
@@ -32,6 +33,17 @@ cdef class Pipe:
         self.name = name
         self.cfg = dict(cfg)
 
+    @property
+    def labels(self) -> Optional[Tuple[str]]:
+        return []
+
+    @property
+    def label_data(self):
+        """Optional JSON-serializable data that would be sufficient to recreate
+        the label set if provided to the `pipe.initialize()` method.
+        """
+        return None
+
     def __call__(self, Doc doc):
         """Apply the pipe to one document. The document is modified in place,
         and returned. This usually happens under the hood when the nlp object
@@ -183,7 +195,7 @@ cdef class Pipe:
         """
         return util.create_default_optimizer()
 
-    def begin_training(self, get_examples, *, pipeline=None, sgd=None):
+    def initialize(self, get_examples, *, nlp=None):
         """Initialize the pipe for training, using data examples if available.
         This method needs to be implemented by each Pipe component,
         ensuring the internal model (if available) is initialized properly
@@ -191,16 +203,11 @@ cdef class Pipe:
 
         get_examples (Callable[[], Iterable[Example]]): Function that
             returns a representative sample of gold-standard Example objects.
-        pipeline (List[Tuple[str, Callable]]): Optional list of pipeline
+        nlp (Language): The current nlp object the component is part of.
-            components that this component is part of. Corresponds to
-            nlp.pipeline.
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
-            create_optimizer if it doesn't exist.
-        RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/pipe#begin_training
+        DOCS: https://nightly.spacy.io/api/pipe#initialize
         """
-        raise NotImplementedError(Errors.E931.format(method="add_label", name=self.name))
+        pass
 
     def _ensure_examples(self, get_examples):
         if get_examples is None or not hasattr(get_examples, "__call__"):

spacy/pipeline/sentencizer.pyx

@@ -58,7 +58,7 @@ class Sentencizer(Pipe):
         else:
             self.punct_chars = set(self.default_punct_chars)
 
-    def begin_training(self, get_examples, pipeline=None, sgd=None):
+    def initialize(self, get_examples, nlp=None):
         pass
 
     def __call__(self, doc):

spacy/pipeline/senter.pyx

@@ -71,6 +71,10 @@ class SentenceRecognizer(Tagger):
         # are 0
         return tuple(["I", "S"])
 
+    @property
+    def label_data(self):
+        return self.labels
+
     def set_annotations(self, docs, batch_tag_ids):
         """Modify a batch of documents, using pre-computed scores.
 
@@ -124,20 +128,15 @@ class SentenceRecognizer(Tagger):
             raise ValueError("nan value when computing loss")
         return float(loss), d_scores
 
-    def begin_training(self, get_examples, *, pipeline=None, sgd=None):
+    def initialize(self, get_examples, *, nlp=None):
         """Initialize the pipe for training, using a representative set
         of data examples.
 
         get_examples (Callable[[], Iterable[Example]]): Function that
             returns a representative sample of gold-standard Example objects.
-        pipeline (List[Tuple[str, Callable]]): Optional list of pipeline
+        nlp (Language): The current nlp object the component is part of.
-            components that this component is part of. Corresponds to
-            nlp.pipeline.
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
-            create_optimizer if it doesn't exist.
-        RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/sentencerecognizer#begin_training
+        DOCS: https://nightly.spacy.io/api/sentencerecognizer#initialize
         """
         self._ensure_examples(get_examples)
         doc_sample = []
@@ -151,9 +150,6 @@ class SentenceRecognizer(Tagger):
         assert len(doc_sample) > 0, Errors.E923.format(name=self.name)
         assert len(label_sample) > 0, Errors.E923.format(name=self.name)
         self.model.initialize(X=doc_sample, Y=label_sample)
-        if sgd is None:
-            sgd = self.create_optimizer()
-        return sgd
 
     def add_label(self, label, values=None):
         raise NotImplementedError

spacy/pipeline/tagger.pyx

@@ -90,6 +90,11 @@ class Tagger(Pipe):
         """
         return tuple(self.cfg["labels"])
 
+    @property
+    def label_data(self):
+        """Data about the labels currently added to the component."""
+        return tuple(self.cfg["labels"])
+
     def __call__(self, doc):
         """Apply the pipe to a Doc.
 
@@ -256,24 +261,24 @@ class Tagger(Pipe):
             raise ValueError("nan value when computing loss")
         return float(loss), d_scores
 
-    def begin_training(self, get_examples, *, pipeline=None, sgd=None):
+    def initialize(self, get_examples, *, nlp=None, labels=None):
         """Initialize the pipe for training, using a representative set
         of data examples.
 
         get_examples (Callable[[], Iterable[Example]]): Function that
             returns a representative sample of gold-standard Example objects..
-        pipeline (List[Tuple[str, Callable]]): Optional list of pipeline
+        nlp (Language): The current nlp object the component is part of.
-            components that this component is part of. Corresponds to
+        labels: The labels to add to the component, typically generated by the
-            nlp.pipeline.
+            `init labels` command. If no labels are provided, the get_examples
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
+            callback is used to extract the labels from the data.
-            create_optimizer if it doesn't exist.
-        RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/tagger#begin_training
+        DOCS: https://nightly.spacy.io/api/tagger#initialize
         """
         self._ensure_examples(get_examples)
-        doc_sample = []
+        if labels is not None:
-        label_sample = []
+            for tag in labels:
+                self.add_label(tag)
+        else:
             tags = set()
             for example in get_examples():
                 for token in example.y:
@@ -281,6 +286,8 @@ class Tagger(Pipe):
                         tags.add(token.tag_)
             for tag in sorted(tags):
                 self.add_label(tag)
+        doc_sample = []
+        label_sample = []
         for example in islice(get_examples(), 10):
             doc_sample.append(example.x)
             gold_tags = example.get_aligned("TAG", as_string=True)
@@ -289,9 +296,6 @@ class Tagger(Pipe):
         assert len(doc_sample) > 0, Errors.E923.format(name=self.name)
         assert len(label_sample) > 0, Errors.E923.format(name=self.name)
         self.model.initialize(X=doc_sample, Y=label_sample)
-        if sgd is None:
-            sgd = self.create_optimizer()
-        return sgd
 
     def add_label(self, label):
         """Add a new label to the pipe.

spacy/pipeline/textcat.py

@@ -154,8 +154,16 @@ class TextCategorizer(Pipe):
 
     @labels.setter
     def labels(self, value: List[str]) -> None:
+        # TODO: This really shouldn't be here. I had a look and I added it when
+        # I added the labels property, but it's pretty nasty to have this, and
+        # will lead to problems.
         self.cfg["labels"] = tuple(value)
 
+    @property
+    def label_data(self) -> List[str]:
+        """RETURNS (List[str]): Information about the component's labels."""
+        return self.labels
+
     def pipe(self, stream: Iterable[Doc], *, batch_size: int = 128) -> Iterator[Doc]:
         """Apply the pipe to a stream of documents. This usually happens under
         the hood when the nlp object is called on a text and all components are
@@ -334,43 +342,40 @@ class TextCategorizer(Pipe):
         self.labels = tuple(list(self.labels) + [label])
         return 1
 
-    def begin_training(
+    def initialize(
         self,
         get_examples: Callable[[], Iterable[Example]],
         *,
-        pipeline: Optional[List[Tuple[str, Callable[[Doc], Doc]]]] = None,
+        nlp: Optional[Language] = None,
-        sgd: Optional[Optimizer] = None,
+        labels: Optional[Dict] = None,
-    ) -> Optimizer:
+    ):
         """Initialize the pipe for training, using a representative set
         of data examples.
 
         get_examples (Callable[[], Iterable[Example]]): Function that
             returns a representative sample of gold-standard Example objects.
-        pipeline (List[Tuple[str, Callable]]): Optional list of pipeline
+        nlp (Language): The current nlp object the component is part of.
-            components that this component is part of. Corresponds to
+        labels: The labels to add to the component, typically generated by the
-            nlp.pipeline.
+            `init labels` command. If no labels are provided, the get_examples
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
+            callback is used to extract the labels from the data.
-            create_optimizer if it doesn't exist.
-        RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/textcategorizer#begin_training
+        DOCS: https://nightly.spacy.io/api/textcategorizer#initialize
         """
         self._ensure_examples(get_examples)
-        subbatch = []  # Select a subbatch of examples to initialize the model
+        if labels is None:
-        for example in islice(get_examples(), 10):
+            for example in get_examples():
-            if len(subbatch) < 2:
-                subbatch.append(example)
                 for cat in example.y.cats:
                     self.add_label(cat)
+        else:
+            for label in labels:
+                self.add_label(label)
+        subbatch = list(islice(get_examples(), 10))
         doc_sample = [eg.reference for eg in subbatch]
         label_sample, _ = self._examples_to_truth(subbatch)
         self._require_labels()
         assert len(doc_sample) > 0, Errors.E923.format(name=self.name)
         assert len(label_sample) > 0, Errors.E923.format(name=self.name)
         self.model.initialize(X=doc_sample, Y=label_sample)
-        if sgd is None:
-            sgd = self.create_optimizer()
-        return sgd
 
     def score(self, examples: Iterable[Example], **kwargs) -> Dict[str, Any]:
         """Score a batch of examples.

spacy/pipeline/tok2vec.py

@@ -1,4 +1,4 @@
-from typing import Iterator, Sequence, Iterable, Optional, Dict, Callable, List, Tuple
+from typing import Iterator, Sequence, Iterable, Optional, Dict, Callable, List
 from thinc.api import Model, set_dropout_rate, Optimizer, Config
 from itertools import islice
 
@@ -203,26 +203,20 @@ class Tok2Vec(Pipe):
     def get_loss(self, examples, scores) -> None:
         pass
 
-    def begin_training(
+    def initialize(
         self,
         get_examples: Callable[[], Iterable[Example]],
         *,
-        pipeline: Optional[List[Tuple[str, Callable[[Doc], Doc]]]] = None,
+        nlp: Optional[Language] = None,
-        sgd: Optional[Optimizer] = None,
     ):
         """Initialize the pipe for training, using a representative set
         of data examples.
 
         get_examples (Callable[[], Iterable[Example]]): Function that
             returns a representative sample of gold-standard Example objects.
-        pipeline (List[Tuple[str, Callable]]): Optional list of pipeline
+        nlp (Language): The current nlp object the component is part of.
-            components that this component is part of. Corresponds to
-            nlp.pipeline.
-        sgd (thinc.api.Optimizer): Optional optimizer. Will be created with
-            create_optimizer if it doesn't exist.
-        RETURNS (thinc.api.Optimizer): The optimizer.
 
-        DOCS: https://nightly.spacy.io/api/tok2vec#begin_training
+        DOCS: https://nightly.spacy.io/api/tok2vec#initialize
         """
         self._ensure_examples(get_examples)
         doc_sample = []

spacy/pipeline/transition_parser.pyx

@@ -1,4 +1,4 @@
-# cython: infer_types=True, cdivision=True, boundscheck=False
+# cython: infer_types=True, cdivision=True, boundscheck=False, binding=True
 from __future__ import print_function
 from cymem.cymem cimport Pool
 cimport numpy as np
@@ -7,6 +7,7 @@ from libcpp.vector cimport vector
 from libc.string cimport memset
 from libc.stdlib cimport calloc, free
 import random
+from typing import Optional
 
 import srsly
 from thinc.api import set_dropout_rate
@@ -95,6 +96,10 @@ cdef class Parser(Pipe):
         class_names = [self.moves.get_class_name(i) for i in range(self.moves.n_moves)]
         return class_names
 
+    @property
+    def label_data(self):
+        return self.moves.labels
+
     @property
     def tok2vec(self):
         """Return the embedding and convolutional layer of the model."""
@@ -354,7 +359,7 @@ cdef class Parser(Pipe):
             # If all weights for an output are 0 in the original model, don't
             # supervise that output. This allows us to add classes.
             loss += (d_scores**2).sum()
-            backprop(d_scores, sgd=sgd)
+            backprop(d_scores)
             # Follow the predicted action
             self.transition_states(states, guesses)
             states = [state for state in states if not state.is_final()]
@@ -405,13 +410,15 @@ cdef class Parser(Pipe):
     def set_output(self, nO):
         self.model.attrs["resize_output"](self.model, nO)
 
-    def begin_training(self, get_examples, pipeline=None, sgd=None, **kwargs):
+    def initialize(self, get_examples, nlp=None, labels=None):
         self._ensure_examples(get_examples)
-        self.cfg.update(kwargs)
         lexeme_norms = self.vocab.lookups.get_table("lexeme_norm", {})
         if len(lexeme_norms) == 0 and self.vocab.lang in util.LEXEME_NORM_LANGS:
             langs = ", ".join(util.LEXEME_NORM_LANGS)
             util.logger.debug(Warnings.W033.format(model="parser or NER", langs=langs))
+        if labels is not None:
+            actions = dict(labels)
+        else:
             actions = self.moves.get_actions(
                 examples=get_examples(),
                 min_freq=self.cfg['min_action_freq'],
@@ -425,11 +432,9 @@ cdef class Parser(Pipe):
         self.moves.initialize_actions(actions)
         # make sure we resize so we have an appropriate upper layer
         self._resize()
-        if sgd is None:
-            sgd = self.create_optimizer()
         doc_sample = []
-        if pipeline is not None:
+        if nlp is not None:
-            for name, component in pipeline:
+            for name, component in nlp.pipeline:
                 if component is self:
                     break
                 if hasattr(component, "pipe"):
@@ -441,9 +446,8 @@ cdef class Parser(Pipe):
                 doc_sample.append(example.predicted)
         assert len(doc_sample) > 0, Errors.E923.format(name=self.name)
         self.model.initialize(doc_sample)
-        if pipeline is not None:
+        if nlp is not None:
-            self.init_multitask_objectives(get_examples, pipeline, sgd=sgd, **self.cfg)
+            self.init_multitask_objectives(get_examples, nlp.pipeline)
-        return sgd
 
     def to_disk(self, path, exclude=tuple()):
         serializers = {

spacy/schemas.py

@@ -1,15 +1,17 @@
 from typing import Dict, List, Union, Optional, Any, Callable, Type, Tuple
 from typing import Iterable, TypeVar, TYPE_CHECKING
 from enum import Enum
-from pydantic import BaseModel, Field, ValidationError, validator
+from pydantic import BaseModel, Field, ValidationError, validator, create_model
 from pydantic import StrictStr, StrictInt, StrictFloat, StrictBool
-from pydantic import root_validator
+from pydantic.main import ModelMetaclass
+from thinc.api import Optimizer, ConfigValidationError
 from thinc.config import Promise
 from collections import defaultdict
-from thinc.api import Optimizer
+import inspect
 
 from .attrs import NAMES
 from .lookups import Lookups
+from .util import is_cython_func
 
 if TYPE_CHECKING:
     # This lets us add type hints for mypy etc. without causing circular imports
@@ -44,6 +46,96 @@ def validate(schema: Type[BaseModel], obj: Dict[str, Any]) -> List[str]:
         return [f"[{loc}] {', '.join(msg)}" for loc, msg in data.items()]
 
 
+# Initialization
+
+
+class ArgSchemaConfig:
+    extra = "forbid"
+    arbitrary_types_allowed = True
+
+
+class ArgSchemaConfigExtra:
+    extra = "forbid"
+    arbitrary_types_allowed = True
+
+
+def get_arg_model(
+    func: Callable,
+    *,
+    exclude: Iterable[str] = tuple(),
+    name: str = "ArgModel",
+    strict: bool = True,
+) -> ModelMetaclass:
+    """Generate a pydantic model for function arguments.
+
+    func (Callable): The function to generate the schema for.
+    exclude (Iterable[str]): Parameter names to ignore.
+    name (str): Name of created model class.
+    strict (bool): Don't allow extra arguments if no variable keyword arguments
+        are allowed on the function.
+    RETURNS (ModelMetaclass): A pydantic model.
+    """
+    sig_args = {}
+    try:
+        sig = inspect.signature(func)
+    except ValueError:
+        # Typically happens if the method is part of a Cython module without
+        # binding=True. Here we just use an empty model that allows everything.
+        return create_model(name, __config__=ArgSchemaConfigExtra)
+    has_variable = False
+    for param in sig.parameters.values():
+        if param.name in exclude:
+            continue
+        if param.kind == param.VAR_KEYWORD:
+            # The function allows variable keyword arguments so we shouldn't
+            # include **kwargs etc. in the schema and switch to non-strict
+            # mode and pass through all other values
+            has_variable = True
+            continue
+        # If no annotation is specified assume it's anything
+        annotation = param.annotation if param.annotation != param.empty else Any
+        # If no default value is specified assume that it's required. Cython
+        # functions/methods will have param.empty for default value None so we
+        # need to treat them differently
+        default_empty = None if is_cython_func(func) else ...
+        default = param.default if param.default != param.empty else default_empty
+        sig_args[param.name] = (annotation, default)
+    is_strict = strict and not has_variable
+    sig_args["__config__"] = ArgSchemaConfig if is_strict else ArgSchemaConfigExtra
+    return create_model(name, **sig_args)
+
+
+def validate_init_settings(
+    func: Callable,
+    settings: Dict[str, Any],
+    *,
+    section: Optional[str] = None,
+    name: str = "",
+    exclude: Iterable[str] = ("get_examples", "nlp"),
+) -> Dict[str, Any]:
+    """Validate initialization settings against the expected arguments in
+    the method signature. Will parse values if possible (e.g. int to string)
+    and return the updated settings dict. Will raise a ConfigValidationError
+    if types don't match or required values are missing.
+
+    func (Callable): The initialize method of a given component etc.
+    settings (Dict[str, Any]): The settings from the repsective [initialize] block.
+    section (str): Initialize section, for error message.
+    name (str): Name of the block in the section.
+    exclude (Iterable[str]): Parameter names to exclude from schema.
+    RETURNS (Dict[str, Any]): The validated settings.
+    """
+    schema = get_arg_model(func, exclude=exclude, name="InitArgModel")
+    try:
+        return schema(**settings).dict()
+    except ValidationError as e:
+        block = "initialize" if not section else f"initialize.{section}"
+        title = f"Error validating initialization settings in [{block}]"
+        raise ConfigValidationError(
+            title=title, errors=e.errors(), config=settings, parent=name
+        ) from None
+
+
 # Matcher token patterns
 
 
@@ -190,7 +282,7 @@ class ModelMetaSchema(BaseModel):
     sources: Optional[Union[List[StrictStr], List[Dict[str, str]]]] = Field(None, title="Training data sources")
     vectors: Dict[str, Any] = Field({}, title="Included word vectors")
     labels: Dict[str, List[str]] = Field({}, title="Component labels, keyed by component name")
-    performance: Dict[str, Union[float, Dict[str, float]]] = Field({}, title="Accuracy and speed numbers")
+    performance: Dict[str, Union[float, Dict[str, Union[float, dict]]]] = Field({}, title="Accuracy and speed numbers")
     spacy_git_version: StrictStr = Field("", title="Commit of spaCy version used")
     # fmt: on
 
@@ -205,8 +297,6 @@ class ModelMetaSchema(BaseModel):
 
 class ConfigSchemaTraining(BaseModel):
     # fmt: off
-    vectors: Optional[StrictStr] = Field(..., title="Path to vectors")
-    lookups: Optional[Lookups] = Field(..., title="Vocab lookups")
     dev_corpus: StrictStr = Field(..., title="Path in the config to the dev data")
     train_corpus: StrictStr = Field(..., title="Path in the config to the training data")
     batcher: Batcher = Field(..., title="Batcher for the training data")
@@ -219,8 +309,6 @@ class ConfigSchemaTraining(BaseModel):
     gpu_allocator: Optional[StrictStr] = Field(..., title="Memory allocator when running on GPU")
     accumulate_gradient: StrictInt = Field(..., title="Whether to divide the batch up into substeps")
     score_weights: Dict[StrictStr, Optional[Union[StrictFloat, StrictInt]]] = Field(..., title="Scores to report and their weights for selecting final model")
-    init_tok2vec: Optional[StrictStr] = Field(..., title="Path to pretrained tok2vec weights")
-    raw_text: Optional[StrictStr] = Field(default=None, title="Raw text")
     optimizer: Optimizer = Field(..., title="The optimizer to use")
     logger: Logger = Field(..., title="The logger to track training progress")
     frozen_components: List[str] = Field(..., title="Pipeline components that shouldn't be updated during training")
@@ -273,36 +361,40 @@ class ConfigSchemaPretrain(BaseModel):
         arbitrary_types_allowed = True
 
 
+class ConfigSchemaInit(BaseModel):
+    # fmt: off
+    vocab_data: Optional[StrictStr] = Field(..., title="Path to JSON-formatted vocabulary file")
+    lookups: Optional[Lookups] = Field(..., title="Vocabulary lookups, e.g. lexeme normalization")
+    vectors: Optional[StrictStr] = Field(..., title="Path to vectors")
+    init_tok2vec: Optional[StrictStr] = Field(..., title="Path to pretrained tok2vec weights")
+    tokenizer: Dict[StrictStr, Any] = Field(..., help="Arguments to be passed into Tokenizer.initialize")
+    components: Dict[StrictStr, Dict[StrictStr, Any]] = Field(..., help="Arguments for Pipe.initialize methods of pipeline components, keyed by component")
+    # fmt: on
+
+    class Config:
+        extra = "forbid"
+        arbitrary_types_allowed = True
+
+
 class ConfigSchema(BaseModel):
     training: ConfigSchemaTraining
     nlp: ConfigSchemaNlp
     pretraining: Union[ConfigSchemaPretrain, ConfigSchemaPretrainEmpty] = {}
     components: Dict[str, Dict[str, Any]]
     corpora: Dict[str, Reader]
-
+    initialize: ConfigSchemaInit
-    @root_validator(allow_reuse=True)
-    def validate_config(cls, values):
-        """Perform additional validation for settings with dependencies."""
-        pt = values.get("pretraining")
-        if pt and not isinstance(pt, ConfigSchemaPretrainEmpty):
-            if pt.objective.get("type") == "vectors" and not values["nlp"].vectors:
-                err = "Need nlp.vectors if pretraining.objective.type is vectors"
-                raise ValueError(err)
-        return values
 
     class Config:
         extra = "allow"
         arbitrary_types_allowed = True
 
 
-class TrainingSchema(BaseModel):
+CONFIG_SCHEMAS = {
-    training: ConfigSchemaTraining
+    "nlp": ConfigSchemaNlp,
-    pretraining: Union[ConfigSchemaPretrain, ConfigSchemaPretrainEmpty] = {}
+    "training": ConfigSchemaTraining,
-    corpora: Dict[str, Reader]
+    "pretraining": ConfigSchemaPretrain,
-
+    "initialize": ConfigSchemaInit,
-    class Config:
+}
-        extra = "allow"
-        arbitrary_types_allowed = True
 
 
 # Project config Schema

spacy/scorer.py

@@ -32,9 +32,7 @@ class PRFScore:
 
     def __add__(self, other):
         return PRFScore(
-            tp=self.tp+other.tp,
+            tp=self.tp + other.tp, fp=self.fp + other.fp, fn=self.fn + other.fn
-            fp=self.fp+other.fp,
-            fn=self.fn+other.fn
         )
 
     def score_set(self, cand: set, gold: set) -> None:
@@ -485,7 +483,7 @@ class Scorer:
                     (pred_ent.start_char, pred_ent.end_char), None
                 )
                 label = gold_span.label_
-                if not label in f_per_type:
+                if label not in f_per_type:
                     f_per_type[label] = PRFScore()
                 gold = gold_span.kb_id_
                 # only evaluating entities that overlap between gold and pred,
@@ -632,7 +630,6 @@ def get_ner_prf(examples: Iterable[Example]) -> Dict[str, PRFScore]:
             continue
         golds = {(e.label_, e.start, e.end) for e in eg.y.ents}
         align_x2y = eg.alignment.x2y
-        preds = set()
         for pred_ent in eg.x.ents:
             if pred_ent.label_ not in scores:
                 scores[pred_ent.label_] = PRFScore()

spacy/tests/conftest.py

@@ -272,22 +272,35 @@ def zh_tokenizer_char():
 def zh_tokenizer_jieba():
     pytest.importorskip("jieba")
     config = {
+        "nlp": {
+            "tokenizer": {
                 "@tokenizers": "spacy.zh.ChineseTokenizer",
                 "segmenter": "jieba",
             }
-    nlp = get_lang_class("zh").from_config({"nlp": {"tokenizer": config}})
+        }
+    }
+    nlp = get_lang_class("zh").from_config(config)
     return nlp.tokenizer
 
 
 @pytest.fixture(scope="session")
 def zh_tokenizer_pkuseg():
     pytest.importorskip("pkuseg")
+    pytest.importorskip("pickle5")
     config = {
+        "nlp": {
+            "tokenizer": {
                 "@tokenizers": "spacy.zh.ChineseTokenizer",
                 "segmenter": "pkuseg",
+            }
+        },
+        "initialize": {"tokenizer": {
                 "pkuseg_model": "default",
             }
-    nlp = get_lang_class("zh").from_config({"nlp": {"tokenizer": config}})
+        },
+    }
+    nlp = get_lang_class("zh").from_config(config)
+    nlp.initialize()
     return nlp.tokenizer
 
 

spacy/tests/doc/test_add_entities.py

@@ -26,7 +26,7 @@ def test_doc_add_entities_set_ents_iob(en_vocab):
     cfg = {"model": DEFAULT_NER_MODEL}
     model = registry.resolve(cfg, validate=True)["model"]
     ner = EntityRecognizer(en_vocab, model, **config)
-    ner.begin_training(lambda: [_ner_example(ner)])
+    ner.initialize(lambda: [_ner_example(ner)])
     ner(doc)
 
     doc.ents = [("ANIMAL", 3, 4)]
@@ -48,7 +48,7 @@ def test_ents_reset(en_vocab):
     cfg = {"model": DEFAULT_NER_MODEL}
     model = registry.resolve(cfg, validate=True)["model"]
     ner = EntityRecognizer(en_vocab, model, **config)
-    ner.begin_training(lambda: [_ner_example(ner)])
+    ner.initialize(lambda: [_ner_example(ner)])
     ner(doc)
     orig_iobs = [t.ent_iob_ for t in doc]
     doc.ents = list(doc.ents)

spacy/tests/doc/test_array.py

@@ -46,9 +46,9 @@ def test_doc_array_morph(en_vocab):
     words = ["Eat", "blue", "ham"]
     morph = ["Feat=V", "Feat=J", "Feat=N"]
     doc = Doc(en_vocab, words=words, morphs=morph)
-    assert morph[0] == doc[0].morph_
+    assert morph[0] == str(doc[0].morph)
-    assert morph[1] == doc[1].morph_
+    assert morph[1] == str(doc[1].morph)
-    assert morph[2] == doc[2].morph_
+    assert morph[2] == str(doc[2].morph)
 
     feats_array = doc.to_array((ORTH, MORPH))
     assert feats_array[0][1] == doc[0].morph.key

spacy/tests/doc/test_doc_api.py

@@ -19,7 +19,7 @@ def test_doc_api_init(en_vocab):
     assert [t.is_sent_start for t in doc] == [True, False, True, False]
     # heads override sent_starts
     doc = Doc(
-        en_vocab, words=words, sent_starts=[True] * 4, heads=heads, deps=["dep"] * 4,
+        en_vocab, words=words, sent_starts=[True] * 4, heads=heads, deps=["dep"] * 4
     )
     assert [t.is_sent_start for t in doc] == [True, False, True, False]
 
@@ -319,15 +319,13 @@ def test_doc_from_array_morph(en_vocab):
     words = ["I", "live", "in", "New", "York", "."]
     morphs = ["Feat1=A", "Feat1=B", "Feat1=C", "Feat1=A|Feat2=D", "Feat2=E", "Feat3=F"]
     # fmt: on
-    doc = Doc(en_vocab, words=words)
+    doc = Doc(en_vocab, words=words, morphs=morphs)
-    for i, morph in enumerate(morphs):
-        doc[i].morph_ = morph
     attrs = [MORPH]
     arr = doc.to_array(attrs)
     new_doc = Doc(en_vocab, words=words)
     new_doc.from_array(attrs, arr)
-    assert [t.morph_ for t in new_doc] == morphs
+    assert [str(t.morph) for t in new_doc] == morphs
-    assert [t.morph_ for t in doc] == [t.morph_ for t in new_doc]
+    assert [str(t.morph) for t in doc] == [str(t.morph) for t in new_doc]
 
 
 def test_doc_api_from_docs(en_tokenizer, de_tokenizer):
@@ -423,7 +421,7 @@ def test_has_annotation(en_vocab):
 
     doc[0].tag_ = "A"
     doc[0].pos_ = "X"
-    doc[0].morph_ = "Feat=Val"
+    doc[0].set_morph("Feat=Val")
     doc[0].lemma_ = "a"
     doc[0].dep_ = "dep"
     doc[0].head = doc[1]
@@ -435,7 +433,7 @@ def test_has_annotation(en_vocab):
 
     doc[1].tag_ = "A"
     doc[1].pos_ = "X"
-    doc[1].morph_ = ""
+    doc[1].set_morph("")
     doc[1].lemma_ = "a"
     doc[1].dep_ = "dep"
     doc.ents = [Span(doc, 0, 2, label="HELLO")]
@@ -533,5 +531,78 @@ def test_doc_ents_setter():
     assert [e.label_ for e in doc.ents] == ["HELLO", "WORLD"]
     vocab = Vocab()
     ents = [("HELLO", 0, 2), (vocab.strings.add("WORLD"), 3, 5)]
+    ents = ["B-HELLO", "I-HELLO", "O", "B-WORLD", "I-WORLD"]
     doc = Doc(vocab, words=words, ents=ents)
     assert [e.label_ for e in doc.ents] == ["HELLO", "WORLD"]
+
+
+def test_doc_morph_setter(en_tokenizer, de_tokenizer):
+    doc1 = en_tokenizer("a b")
+    doc1b = en_tokenizer("c d")
+    doc2 = de_tokenizer("a b")
+
+    # unset values can be copied
+    doc1[0].morph = doc1[1].morph
+    assert doc1[0].morph.key == 0
+    assert doc1[1].morph.key == 0
+
+    # morph values from the same vocab can be copied
+    doc1[0].set_morph("Feat=Val")
+    doc1[1].morph = doc1[0].morph
+    assert doc1[0].morph == doc1[1].morph
+
+    # ... also across docs
+    doc1b[0].morph = doc1[0].morph
+    assert doc1[0].morph == doc1b[0].morph
+
+    doc2[0].set_morph("Feat2=Val2")
+
+    # the morph value must come from the same vocab
+    with pytest.raises(ValueError):
+        doc1[0].morph = doc2[0].morph
+
+
+def test_doc_init_iob():
+    """Test ents validation/normalization in Doc.__init__"""
+    words = ["a", "b", "c", "d", "e"]
+    ents = ["O"] * len(words)
+    doc = Doc(Vocab(), words=words, ents=ents)
+    assert doc.ents == ()
+
+    ents = ["B-PERSON", "I-PERSON", "O", "I-PERSON", "I-PERSON"]
+    doc = Doc(Vocab(), words=words, ents=ents)
+    assert len(doc.ents) == 2
+
+    ents = ["B-PERSON", "I-PERSON", "O", "I-PERSON", "I-GPE"]
+    doc = Doc(Vocab(), words=words, ents=ents)
+    assert len(doc.ents) == 3
+
+    # None is missing
+    ents = ["B-PERSON", "I-PERSON", "O", None, "I-GPE"]
+    doc = Doc(Vocab(), words=words, ents=ents)
+    assert len(doc.ents) == 2
+
+    # empty tag is missing
+    ents = ["", "B-PERSON", "O", "B-PERSON", "I-PERSON"]
+    doc = Doc(Vocab(), words=words, ents=ents)
+    assert len(doc.ents) == 2
+
+    # invalid IOB
+    ents = ["Q-PERSON", "I-PERSON", "O", "I-PERSON", "I-GPE"]
+    with pytest.raises(ValueError):
+        doc = Doc(Vocab(), words=words, ents=ents)
+
+    # no dash
+    ents = ["OPERSON", "I-PERSON", "O", "I-PERSON", "I-GPE"]
+    with pytest.raises(ValueError):
+        doc = Doc(Vocab(), words=words, ents=ents)
+
+    # no ent type
+    ents = ["O", "B-", "O", "I-PERSON", "I-GPE"]
+    with pytest.raises(ValueError):
+        doc = Doc(Vocab(), words=words, ents=ents)
+
+    # not strings or None
+    ents = [0, "B-", "O", "I-PERSON", "I-GPE"]
+    with pytest.raises(ValueError):
+        doc = Doc(Vocab(), words=words, ents=ents)

spacy/tests/doc/test_morphanalysis.py

@@ -4,13 +4,13 @@ import pytest
 @pytest.fixture
 def i_has(en_tokenizer):
     doc = en_tokenizer("I has")
-    doc[0].morph_ = {"PronType": "prs"}
+    doc[0].set_morph({"PronType": "prs"})
-    doc[1].morph_ = {
+    doc[1].set_morph({
         "VerbForm": "fin",
         "Tense": "pres",
         "Number": "sing",
         "Person": "three",
-    }
+    })
 
     return doc
 
@@ -47,20 +47,20 @@ def test_morph_get(i_has):
 def test_morph_set(i_has):
     assert i_has[0].morph.get("PronType") == ["prs"]
     # set by string
-    i_has[0].morph_ = "PronType=unk"
+    i_has[0].set_morph("PronType=unk")
     assert i_has[0].morph.get("PronType") == ["unk"]
     # set by string, fields are alphabetized
-    i_has[0].morph_ = "PronType=123|NounType=unk"
+    i_has[0].set_morph("PronType=123|NounType=unk")
-    assert i_has[0].morph_ == "NounType=unk|PronType=123"
+    assert str(i_has[0].morph) == "NounType=unk|PronType=123"
     # set by dict
-    i_has[0].morph_ = {"AType": "123", "BType": "unk"}
+    i_has[0].set_morph({"AType": "123", "BType": "unk"})
-    assert i_has[0].morph_ == "AType=123|BType=unk"
+    assert str(i_has[0].morph) == "AType=123|BType=unk"
     # set by string with multiple values, fields and values are alphabetized
-    i_has[0].morph_ = "BType=c|AType=b,a"
+    i_has[0].set_morph("BType=c|AType=b,a")
-    assert i_has[0].morph_ == "AType=a,b|BType=c"
+    assert str(i_has[0].morph) == "AType=a,b|BType=c"
     # set by dict with multiple values, fields and values are alphabetized
-    i_has[0].morph_ = {"AType": "b,a", "BType": "c"}
+    i_has[0].set_morph({"AType": "b,a", "BType": "c"})
-    assert i_has[0].morph_ == "AType=a,b|BType=c"
+    assert str(i_has[0].morph) == "AType=a,b|BType=c"
 
 
 def test_morph_str(i_has):
@@ -72,25 +72,25 @@ def test_morph_property(tokenizer):
     doc = tokenizer("a dog")
 
     # set through token.morph_
-    doc[0].morph_ = "PronType=prs"
+    doc[0].set_morph("PronType=prs")
-    assert doc[0].morph_ == "PronType=prs"
+    assert str(doc[0].morph) == "PronType=prs"
     assert doc.to_array(["MORPH"])[0] != 0
 
     # unset with token.morph
-    doc[0].morph = 0
+    doc[0].set_morph(None)
     assert doc.to_array(["MORPH"])[0] == 0
 
     # empty morph is equivalent to "_"
-    doc[0].morph_ = ""
+    doc[0].set_morph("")
-    assert doc[0].morph_ == ""
+    assert str(doc[0].morph) == ""
     assert doc.to_array(["MORPH"])[0] == tokenizer.vocab.strings["_"]
 
     # "_" morph is also equivalent to empty morph
-    doc[0].morph_ = "_"
+    doc[0].set_morph("_")
-    assert doc[0].morph_ == ""
+    assert str(doc[0].morph) == ""
     assert doc.to_array(["MORPH"])[0] == tokenizer.vocab.strings["_"]
 
     # set through existing hash with token.morph
     tokenizer.vocab.strings.add("Feat=Val")
-    doc[0].morph = tokenizer.vocab.strings.add("Feat=Val")
+    doc[0].set_morph(tokenizer.vocab.strings.add("Feat=Val"))
-    assert doc[0].morph_ == "Feat=Val"
+    assert str(doc[0].morph) == "Feat=Val"

spacy/tests/doc/test_retokenize_merge.py

@@ -21,11 +21,11 @@ def test_doc_retokenize_merge(en_tokenizer):
     assert doc[4].text == "the beach boys"
     assert doc[4].text_with_ws == "the beach boys "
     assert doc[4].tag_ == "NAMED"
-    assert doc[4].morph_ == "Number=Plur"
+    assert str(doc[4].morph) == "Number=Plur"
     assert doc[5].text == "all night"
     assert doc[5].text_with_ws == "all night"
     assert doc[5].tag_ == "NAMED"
-    assert doc[5].morph_ == "Number=Plur"
+    assert str(doc[5].morph) == "Number=Plur"
 
 
 def test_doc_retokenize_merge_children(en_tokenizer):
@@ -201,6 +201,12 @@ def test_doc_retokenize_spans_entity_merge(en_tokenizer):
     heads = [1, 2, 2, 4, 6, 4, 2, 8, 6, 8, 9, 8, 8, 14, 12, 2, 15]
     tags = ["NNP", "NNP", "VBZ", "DT", "VB", "RP", "NN", "WP", "VBZ", "IN", "NNP", "CC", "VBZ", "NNP", "NNP", ".", "SP"]
     ents = [("PERSON", 0, 2), ("GPE", 10, 11), ("PERSON", 13, 15)]
+    ents = ["O"] * len(heads)
+    ents[0] = "B-PERSON"
+    ents[1] = "I-PERSON"
+    ents[10] = "B-GPE"
+    ents[13] = "B-PERSON"
+    ents[14] = "I-PERSON"
     # fmt: on
     tokens = en_tokenizer(text)
     doc = Doc(
@@ -269,7 +275,11 @@ def test_doc_retokenize_spans_entity_merge_iob(en_vocab):
     # if there is a parse, span.root provides default values
     words = ["a", "b", "c", "d", "e", "f", "g", "h", "i"]
     heads = [0, 0, 3, 0, 0, 0, 5, 0, 0]
-    ents = [("ent-de", 3, 5), ("ent-fg", 5, 7)]
+    ents = ["O"] * len(words)
+    ents[3] = "B-ent-de"
+    ents[4] = "I-ent-de"
+    ents[5] = "B-ent-fg"
+    ents[6] = "I-ent-fg"
     deps = ["dep"] * len(words)
     en_vocab.strings.add("ent-de")
     en_vocab.strings.add("ent-fg")
@@ -292,7 +302,11 @@ def test_doc_retokenize_spans_entity_merge_iob(en_vocab):
     # check that B is preserved if span[start] is B
     words = ["a", "b", "c", "d", "e", "f", "g", "h", "i"]
     heads = [0, 0, 3, 4, 0, 0, 5, 0, 0]
-    ents = [("ent-de", 3, 5), ("ent-de", 5, 7)]
+    ents = ["O"] * len(words)
+    ents[3] = "B-ent-de"
+    ents[4] = "I-ent-de"
+    ents[5] = "B-ent-de"
+    ents[6] = "I-ent-de"
     deps = ["dep"] * len(words)
     doc = Doc(en_vocab, words=words, heads=heads, deps=deps, ents=ents)
     with doc.retokenize() as retokenizer:

spacy/tests/doc/test_retokenize_split.py

@@ -27,11 +27,11 @@ def test_doc_retokenize_split(en_vocab):
     assert doc[0].text == "Los"
     assert doc[0].head.text == "Angeles"
     assert doc[0].idx == 0
-    assert doc[0].morph_ == "Number=Sing"
+    assert str(doc[0].morph) == "Number=Sing"
     assert doc[1].idx == 3
     assert doc[1].text == "Angeles"
     assert doc[1].head.text == "start"
-    assert doc[1].morph_ == "Number=Sing"
+    assert str(doc[1].morph) == "Number=Sing"
     assert doc[2].text == "start"
     assert doc[2].head.text == "."
     assert doc[3].text == "."

spacy/tests/doc/test_to_json.py

@@ -9,7 +9,7 @@ def doc(en_vocab):
     tags = ["VBP", "NN", "NN"]
     heads = [0, 0, 0]
     deps = ["ROOT", "dobj", "dobj"]
-    ents = [("ORG", 1, 2)]
+    ents = ["O", "B-ORG", "O"]
     return Doc(
         en_vocab, words=words, pos=pos, tags=tags, heads=heads, deps=deps, ents=ents
     )

spacy/tests/lang/de/test_noun_chunks.py

@@ -2,8 +2,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed_de(de_tokenizer):
-    """Test that noun_chunks raises Value Error for 'de' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'de' language if Doc is not parsed."""
-    """
     doc = de_tokenizer("Er lag auf seinem")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/el/test_noun_chunks.py

@@ -2,8 +2,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed_el(el_tokenizer):
-    """Test that noun_chunks raises Value Error for 'el' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'el' language if Doc is not parsed."""
-    """
     doc = el_tokenizer("είναι χώρα της νοτιοανατολικής")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/en/test_noun_chunks.py

@@ -7,8 +7,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed(en_tokenizer):
-    """Test that noun_chunks raises Value Error for 'en' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'en' language if Doc is not parsed."""
-    """
     doc = en_tokenizer("This is a sentence")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/es/test_noun_chunks.py

@@ -2,8 +2,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed_es(es_tokenizer):
-    """Test that noun_chunks raises Value Error for 'es' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'es' language if Doc is not parsed."""
-    """
     doc = es_tokenizer("en Oxford este verano")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/fa/test_noun_chunks.py

@@ -2,8 +2,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed_fa(fa_tokenizer):
-    """Test that noun_chunks raises Value Error for 'fa' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'fa' language if Doc is not parsed."""
-    """
 
     doc = fa_tokenizer("این یک جمله نمونه می باشد.")
     with pytest.raises(ValueError):

spacy/tests/lang/fr/test_exceptions.py

@@ -36,9 +36,7 @@ def test_fr_tokenizer_infix_exceptions(fr_tokenizer, text):
     assert len(tokens) == 1
 
 
-@pytest.mark.parametrize(
+@pytest.mark.parametrize("text", ["janv.", "juill.", "Dr.", "av.", "sept."])
-    "text", ["janv.", "juill.", "Dr.", "av.", "sept."],
-)
 def test_fr_tokenizer_handles_abbr(fr_tokenizer, text):
     tokens = fr_tokenizer(text)
     assert len(tokens) == 1

spacy/tests/lang/fr/test_noun_chunks.py

@@ -2,8 +2,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed_fr(fr_tokenizer):
-    """Test that noun_chunks raises Value Error for 'fr' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'fr' language if Doc is not parsed."""
-    """
     doc = fr_tokenizer("trouver des travaux antérieurs")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/id/test_noun_chunks.py

@@ -2,8 +2,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed_id(id_tokenizer):
-    """Test that noun_chunks raises Value Error for 'id' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'id' language if Doc is not parsed."""
-    """
     doc = id_tokenizer("sebelas")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/ja/test_tokenizer.py

@@ -112,7 +112,7 @@ def test_ja_tokenizer_split_modes(ja_tokenizer, text, len_a, len_b, len_c):
 
 
 @pytest.mark.parametrize(
-    "text,sub_tokens_list_a,sub_tokens_list_b,sub_tokens_list_c", SUB_TOKEN_TESTS,
+    "text,sub_tokens_list_a,sub_tokens_list_b,sub_tokens_list_c", SUB_TOKEN_TESTS
 )
 def test_ja_tokenizer_sub_tokens(
     ja_tokenizer, text, sub_tokens_list_a, sub_tokens_list_b, sub_tokens_list_c

spacy/tests/lang/nb/test_noun_chunks.py

@@ -2,8 +2,7 @@ import pytest
 
 
 def test_noun_chunks_is_parsed_nb(nb_tokenizer):
-    """Test that noun_chunks raises Value Error for 'nb' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'nb' language if Doc is not parsed."""
-    """
     doc = nb_tokenizer("Smørsausen brukes bl.a. til")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/ne/test_text.py

@@ -8,7 +8,7 @@ def test_ne_tokenizer_handlers_long_text(ne_tokenizer):
 
 
 @pytest.mark.parametrize(
-    "text,length", [("समय जान कति पनि बेर लाग्दैन ।", 7), ("म ठूलो हुँदै थिएँ ।", 5)],
+    "text,length", [("समय जान कति पनि बेर लाग्दैन ।", 7), ("म ठूलो हुँदै थिएँ ।", 5)]
 )
 def test_ne_tokenizer_handles_cnts(ne_tokenizer, text, length):
     tokens = ne_tokenizer(text)

spacy/tests/lang/sa/test_text.py

@@ -10,7 +10,7 @@ def test_sa_tokenizer_handles_long_text(sa_tokenizer):
 @pytest.mark.parametrize(
     "text,length",
     [
-        ("श्री भगवानुवाच पश्य मे पार्थ रूपाणि शतशोऽथ सहस्रशः।", 9,),
+        ("श्री भगवानुवाच पश्य मे पार्थ रूपाणि शतशोऽथ सहस्रशः।", 9),
         ("गुणान् सर्वान् स्वभावो मूर्ध्नि वर्तते ।", 6),
     ],
 )

spacy/tests/lang/sv/test_noun_chunks.py

@@ -3,8 +3,7 @@ from spacy.tokens import Doc
 
 
 def test_noun_chunks_is_parsed_sv(sv_tokenizer):
-    """Test that noun_chunks raises Value Error for 'sv' language if Doc is not parsed.
+    """Test that noun_chunks raises Value Error for 'sv' language if Doc is not parsed."""
-    """
     doc = sv_tokenizer("Studenten läste den bästa boken")
     with pytest.raises(ValueError):
         list(doc.noun_chunks)

spacy/tests/lang/test_lemmatizers.py

@@ -17,16 +17,31 @@ def test_lemmatizer_initialize(lang, capfd):
     @registry.misc("lemmatizer_init_lookups")
     def lemmatizer_init_lookups():
         lookups = Lookups()
-        lookups.add_table("lemma_lookup", {"cope": "cope"})
+        lookups.add_table("lemma_lookup", {"cope": "cope", "x": "y"})
         lookups.add_table("lemma_index", {"verb": ("cope", "cop")})
         lookups.add_table("lemma_exc", {"verb": {"coping": ("cope",)}})
         lookups.add_table("lemma_rules", {"verb": [["ing", ""]]})
         return lookups
 
-    """Test that languages can be initialized."""
+    # Test that languages can be initialized
     nlp = get_lang_class(lang)()
-    nlp.add_pipe("lemmatizer", config={"lookups": {"@misc": "lemmatizer_init_lookups"}})
+    lemmatizer = nlp.add_pipe("lemmatizer", config={"mode": "lookup"})
+    assert not lemmatizer.lookups.tables
+    nlp.config["initialize"]["components"]["lemmatizer"] = {
+        "lookups": {"@misc": "lemmatizer_init_lookups"}
+    }
+    with pytest.raises(ValueError):
+        nlp("x")
+    nlp.initialize()
+    assert lemmatizer.lookups.tables
+    doc = nlp("x")
     # Check for stray print statements (see #3342)
-    doc = nlp("test")  # noqa: F841
     captured = capfd.readouterr()
     assert not captured.out
+    assert doc[0].lemma_ == "y"
+
+    # Test initialization by calling .initialize() directly
+    nlp = get_lang_class(lang)()
+    lemmatizer = nlp.add_pipe("lemmatizer", config={"mode": "lookup"})
+    lemmatizer.initialize(lookups=lemmatizer_init_lookups())
+    assert nlp("x")[0].lemma_ == "y"

spacy/tests/lang/zh/test_serialize.py

@@ -27,9 +27,18 @@ def test_zh_tokenizer_serialize_jieba(zh_tokenizer_jieba):
 
 @pytest.mark.slow
 def test_zh_tokenizer_serialize_pkuseg_with_processors(zh_tokenizer_pkuseg):
-    nlp = Chinese(
+    config = {
-        meta={
+        "nlp": {
-            "tokenizer": {"config": {"segmenter": "pkuseg", "pkuseg_model": "medicine"}}
+            "tokenizer": {
+                "@tokenizers": "spacy.zh.ChineseTokenizer",
+                "segmenter": "pkuseg",
             }
-    )
+        },
+        "initialize": {"tokenizer": {
+                "pkuseg_model": "medicine",
+            }
+        },
+    }
+    nlp = Chinese.from_config(config)
+    nlp.initialize()
     zh_tokenizer_serialize(nlp.tokenizer)

spacy/tests/matcher/test_matcher_api.py

@@ -236,13 +236,13 @@ def test_matcher_subset_value_operator(en_vocab):
     matcher.add("M", [pattern])
     doc = Doc(en_vocab, words=["a", "b", "c"])
     assert len(matcher(doc)) == 3
-    doc[0].morph_ = "Feat=Val"
+    doc[0].set_morph("Feat=Val")
     assert len(matcher(doc)) == 3
-    doc[0].morph_ = "Feat=Val|Feat2=Val2"
+    doc[0].set_morph("Feat=Val|Feat2=Val2")
     assert len(matcher(doc)) == 3
-    doc[0].morph_ = "Feat=Val|Feat2=Val2|Feat3=Val3"
+    doc[0].set_morph("Feat=Val|Feat2=Val2|Feat3=Val3")
     assert len(matcher(doc)) == 2
-    doc[0].morph_ = "Feat=Val|Feat2=Val2|Feat3=Val3|Feat4=Val4"
+    doc[0].set_morph("Feat=Val|Feat2=Val2|Feat3=Val3|Feat4=Val4")
     assert len(matcher(doc)) == 2
 
     # IS_SUBSET acts like "IN" for attrs other than MORPH
@@ -268,11 +268,11 @@ def test_matcher_superset_value_operator(en_vocab):
     matcher.add("M", [pattern])
     doc = Doc(en_vocab, words=["a", "b", "c"])
     assert len(matcher(doc)) == 0
-    doc[0].morph_ = "Feat=Val|Feat2=Val2"
+    doc[0].set_morph("Feat=Val|Feat2=Val2")
     assert len(matcher(doc)) == 0
-    doc[0].morph_ = "Feat=Val|Feat2=Val2|Feat3=Val3"
+    doc[0].set_morph("Feat=Val|Feat2=Val2|Feat3=Val3")
     assert len(matcher(doc)) == 1
-    doc[0].morph_ = "Feat=Val|Feat2=Val2|Feat3=Val3|Feat4=Val4"
+    doc[0].set_morph("Feat=Val|Feat2=Val2|Feat3=Val3|Feat4=Val4")
     assert len(matcher(doc)) == 1
 
     # IS_SUPERSET with more than one value only matches for MORPH
@@ -310,9 +310,9 @@ def test_matcher_morph_handling(en_vocab):
     doc = Doc(en_vocab, words=["a", "b", "c"])
     assert len(matcher(doc)) == 0
 
-    doc[0].morph_ = "Feat2=Val2|Feat1=Val1"
+    doc[0].set_morph("Feat2=Val2|Feat1=Val1")
     assert len(matcher(doc)) == 2
-    doc[0].morph_ = "Feat1=Val1|Feat2=Val2"
+    doc[0].set_morph("Feat1=Val1|Feat2=Val2")
     assert len(matcher(doc)) == 2
 
     # multiple values are split
@@ -324,9 +324,9 @@ def test_matcher_morph_handling(en_vocab):
     doc = Doc(en_vocab, words=["a", "b", "c"])
     assert len(matcher(doc)) == 0
 
-    doc[0].morph_ = "Feat2=Val2,Val3|Feat1=Val1"
+    doc[0].set_morph("Feat2=Val2,Val3|Feat1=Val1")
     assert len(matcher(doc)) == 1
-    doc[0].morph_ = "Feat1=Val1,Val3|Feat2=Val2"
+    doc[0].set_morph("Feat1=Val1,Val3|Feat2=Val2")
     assert len(matcher(doc)) == 2
 
 
@@ -405,7 +405,7 @@ def test_attr_pipeline_checks(en_vocab):
     doc2 = Doc(en_vocab, words=["Test"])
     doc2[0].tag_ = "TAG"
     doc2[0].pos_ = "X"
-    doc2[0].morph_ = "Feat=Val"
+    doc2[0].set_morph("Feat=Val")
     doc2[0].lemma_ = "LEMMA"
     doc3 = Doc(en_vocab, words=["Test"])
     # DEP requires DEP

spacy/tests/matcher/test_phrase_matcher.py

@@ -190,7 +190,7 @@ def test_phrase_matcher_validation(en_vocab):
     doc2 = Doc(en_vocab, words=["Test"])
     doc2[0].tag_ = "TAG"
     doc2[0].pos_ = "X"
-    doc2[0].morph_ = "Feat=Val"
+    doc2[0].set_morph("Feat=Val")
     doc3 = Doc(en_vocab, words=["Test"])
     matcher = PhraseMatcher(en_vocab, validate=True)
     with pytest.warns(UserWarning):
@@ -217,7 +217,7 @@ def test_attr_pipeline_checks(en_vocab):
     doc2 = Doc(en_vocab, words=["Test"])
     doc2[0].tag_ = "TAG"
     doc2[0].pos_ = "X"
-    doc2[0].morph_ = "Feat=Val"
+    doc2[0].set_morph("Feat=Val")
     doc2[0].lemma_ = "LEMMA"
     doc3 = Doc(en_vocab, words=["Test"])
     # DEP requires DEP

spacy/tests/parser/test_add_label.py

@@ -35,7 +35,7 @@ def test_init_parser(parser):
 def _train_parser(parser):
     fix_random_seed(1)
     parser.add_label("left")
-    parser.begin_training(lambda: [_parser_example(parser)], **parser.cfg)
+    parser.initialize(lambda: [_parser_example(parser)])
     sgd = Adam(0.001)
 
     for i in range(5):
@@ -87,7 +87,7 @@ def test_add_label_deserializes_correctly():
     ner1.add_label("C")
     ner1.add_label("B")
     ner1.add_label("A")
-    ner1.begin_training(lambda: [_ner_example(ner1)])
+    ner1.initialize(lambda: [_ner_example(ner1)])
     ner2 = EntityRecognizer(Vocab(), model, **config)
 
     # the second model needs to be resized before we can call from_bytes

spacy/tests/parser/test_ner.py

@@ -202,7 +202,7 @@ def test_train_empty():
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
     ner = nlp.add_pipe("ner", last=True)
     ner.add_label("PERSON")
-    nlp.begin_training()
+    nlp.initialize()
     for itn in range(2):
         losses = {}
         batches = util.minibatch(train_examples, size=8)
@@ -213,7 +213,7 @@ def test_train_empty():
 def test_overwrite_token():
     nlp = English()
     nlp.add_pipe("ner")
-    nlp.begin_training()
+    nlp.initialize()
     # The untrained NER will predict O for each token
     doc = nlp("I live in New York")
     assert [token.ent_iob_ for token in doc] == ["O", "O", "O", "O", "O"]
@@ -235,7 +235,7 @@ def test_empty_ner():
     nlp = English()
     ner = nlp.add_pipe("ner")
     ner.add_label("MY_LABEL")
-    nlp.begin_training()
+    nlp.initialize()
     doc = nlp("John is watching the news about Croatia's elections")
     # if this goes wrong, the initialization of the parser's upper layer is probably broken
     result = ["O", "O", "O", "O", "O", "O", "O", "O", "O"]
@@ -254,7 +254,7 @@ def test_ruler_before_ner():
     # 2: untrained NER - should set everything else to O
     untrained_ner = nlp.add_pipe("ner")
     untrained_ner.add_label("MY_LABEL")
-    nlp.begin_training()
+    nlp.initialize()
     doc = nlp("This is Antti Korhonen speaking in Finland")
     expected_iobs = ["B", "O", "O", "O", "O", "O", "O"]
     expected_types = ["THING", "", "", "", "", "", ""]
@@ -269,7 +269,7 @@ def test_ner_before_ruler():
     # 1: untrained NER - should set everything to O
     untrained_ner = nlp.add_pipe("ner", name="uner")
     untrained_ner.add_label("MY_LABEL")
-    nlp.begin_training()
+    nlp.initialize()
 
     # 2 : Entity Ruler - should set "this" to B and keep everything else O
     patterns = [{"label": "THING", "pattern": "This"}]
@@ -290,7 +290,7 @@ def test_block_ner():
     nlp.add_pipe("blocker", config={"start": 2, "end": 5})
     untrained_ner = nlp.add_pipe("ner")
     untrained_ner.add_label("MY_LABEL")
-    nlp.begin_training()
+    nlp.initialize()
     doc = nlp("This is Antti L Korhonen speaking in Finland")
     expected_iobs = ["O", "O", "B", "B", "B", "O", "O", "O"]
     expected_types = ["", "", "", "", "", "", "", ""]
@@ -307,7 +307,7 @@ def test_overfitting_IO():
         train_examples.append(Example.from_dict(nlp.make_doc(text), annotations))
         for ent in annotations.get("entities"):
             ner.add_label(ent[2])
-    optimizer = nlp.begin_training()
+    optimizer = nlp.initialize()
 
     for i in range(50):
         losses = {}
@@ -340,13 +340,13 @@ def test_ner_warns_no_lookups(caplog):
     assert not len(nlp.vocab.lookups)
     nlp.add_pipe("ner")
     with caplog.at_level(logging.DEBUG):
-        nlp.begin_training()
+        nlp.initialize()
         assert "W033" in caplog.text
     caplog.clear()
     nlp.vocab.lookups.add_table("lexeme_norm")
     nlp.vocab.lookups.get_table("lexeme_norm")["a"] = "A"
     with caplog.at_level(logging.DEBUG):
-        nlp.begin_training()
+        nlp.initialize()
         assert "W033" not in caplog.text
 
 

spacy/tests/parser/test_parse.py

@@ -191,7 +191,7 @@ def test_overfitting_IO():
         train_examples.append(Example.from_dict(nlp.make_doc(text), annotations))
         for dep in annotations.get("deps", []):
             parser.add_label(dep)
-    optimizer = nlp.begin_training()
+    optimizer = nlp.initialize()
     for i in range(100):
         losses = {}
         nlp.update(train_examples, sgd=optimizer, losses=losses)

spacy/tests/parser/test_preset_sbd.py

@@ -34,7 +34,7 @@ def parser(vocab):
     parser.cfg["hidden_width"] = 32
     # parser.add_label('right')
     parser.add_label("left")
-    parser.begin_training(lambda: [_parser_example(parser)], **parser.cfg)
+    parser.initialize(lambda: [_parser_example(parser)])
     sgd = Adam(0.001)
 
     for i in range(10):

spacy/tests/pipeline/test_attributeruler.py

@@ -69,9 +69,9 @@ def test_attributeruler_init(nlp, pattern_dicts):
         a.add(**p)
     doc = nlp("This is a test.")
     assert doc[2].lemma_ == "the"
-    assert doc[2].morph_ == "Case=Nom|Number=Plur"
+    assert str(doc[2].morph) == "Case=Nom|Number=Plur"
     assert doc[3].lemma_ == "cat"
-    assert doc[3].morph_ == "Case=Nom|Number=Sing"
+    assert str(doc[3].morph) == "Case=Nom|Number=Sing"
     assert doc.has_annotation("LEMMA")
     assert doc.has_annotation("MORPH")
 
@@ -81,9 +81,9 @@ def test_attributeruler_init_patterns(nlp, pattern_dicts):
     nlp.add_pipe("attribute_ruler", config={"pattern_dicts": pattern_dicts})
     doc = nlp("This is a test.")
     assert doc[2].lemma_ == "the"
-    assert doc[2].morph_ == "Case=Nom|Number=Plur"
+    assert str(doc[2].morph) == "Case=Nom|Number=Plur"
     assert doc[3].lemma_ == "cat"
-    assert doc[3].morph_ == "Case=Nom|Number=Sing"
+    assert str(doc[3].morph) == "Case=Nom|Number=Sing"
     assert doc.has_annotation("LEMMA")
     assert doc.has_annotation("MORPH")
     nlp.remove_pipe("attribute_ruler")
@@ -94,9 +94,9 @@ def test_attributeruler_init_patterns(nlp, pattern_dicts):
     )
     doc = nlp("This is a test.")
     assert doc[2].lemma_ == "the"
-    assert doc[2].morph_ == "Case=Nom|Number=Plur"
+    assert str(doc[2].morph) == "Case=Nom|Number=Plur"
     assert doc[3].lemma_ == "cat"
-    assert doc[3].morph_ == "Case=Nom|Number=Sing"
+    assert str(doc[3].morph) == "Case=Nom|Number=Sing"
     assert doc.has_annotation("LEMMA")
     assert doc.has_annotation("MORPH")
 
@@ -106,9 +106,9 @@ def test_attributeruler_score(nlp, pattern_dicts):
     nlp.add_pipe("attribute_ruler", config={"pattern_dicts": pattern_dicts})
     doc = nlp("This is a test.")
     assert doc[2].lemma_ == "the"
-    assert doc[2].morph_ == "Case=Nom|Number=Plur"
+    assert str(doc[2].morph) == "Case=Nom|Number=Plur"
     assert doc[3].lemma_ == "cat"
-    assert doc[3].morph_ == "Case=Nom|Number=Sing"
+    assert str(doc[3].morph) == "Case=Nom|Number=Sing"
 
     dev_examples = [
         Example.from_dict(
@@ -150,10 +150,10 @@ def test_attributeruler_tag_map(nlp, tag_map):
     for i in range(len(doc)):
         if i == 4:
             assert doc[i].pos_ == "PUNCT"
-            assert doc[i].morph_ == "PunctType=peri"
+            assert str(doc[i].morph) == "PunctType=peri"
         else:
             assert doc[i].pos_ == ""
-            assert doc[i].morph_ == ""
+            assert str(doc[i].morph) == ""
 
 
 def test_attributeruler_morph_rules(nlp, morph_rules):
@@ -168,11 +168,11 @@ def test_attributeruler_morph_rules(nlp, morph_rules):
     for i in range(len(doc)):
         if i != 2:
             assert doc[i].pos_ == ""
-            assert doc[i].morph_ == ""
+            assert str(doc[i].morph) == ""
         else:
             assert doc[2].pos_ == "DET"
             assert doc[2].lemma_ == "a"
-            assert doc[2].morph_ == "Case=Nom"
+            assert str(doc[2].morph) == "Case=Nom"
 
 
 def test_attributeruler_indices(nlp):
@@ -194,14 +194,14 @@ def test_attributeruler_indices(nlp):
     for i in range(len(doc)):
         if i == 1:
             assert doc[i].lemma_ == "was"
-            assert doc[i].morph_ == "Case=Nom|Number=Sing"
+            assert str(doc[i].morph) == "Case=Nom|Number=Sing"
         elif i == 2:
             assert doc[i].lemma_ == "the"
-            assert doc[i].morph_ == "Case=Nom|Number=Plur"
+            assert str(doc[i].morph) == "Case=Nom|Number=Plur"
         elif i == 3:
             assert doc[i].lemma_ == "cat"
         else:
-            assert doc[i].morph_ == ""
+            assert str(doc[i].morph) == ""
     # raises an error when trying to modify a token outside of the match
     a.add([[{"ORTH": "a"}, {"ORTH": "test"}]], {"LEMMA": "cat"}, index=2)
     with pytest.raises(ValueError):

spacy/tests/pipeline/test_entity_linker.py

@@ -134,7 +134,7 @@ def test_kb_undefined(nlp):
     """Test that the EL can't train without defining a KB"""
     entity_linker = nlp.add_pipe("entity_linker", config={})
     with pytest.raises(ValueError):
-        entity_linker.begin_training(lambda: [])
+        entity_linker.initialize(lambda: [])
 
 
 def test_kb_empty(nlp):
@@ -143,7 +143,7 @@ def test_kb_empty(nlp):
     entity_linker = nlp.add_pipe("entity_linker", config=config)
     assert len(entity_linker.kb) == 0
     with pytest.raises(ValueError):
-        entity_linker.begin_training(lambda: [])
+        entity_linker.initialize(lambda: [])
 
 
 def test_kb_serialize(nlp):
@@ -254,14 +254,12 @@ def test_vocab_serialization(nlp):
     mykb = KnowledgeBase(nlp.vocab, entity_vector_length=1)
 
     # adding entities
-    q1_hash = mykb.add_entity(entity="Q1", freq=27, entity_vector=[1])
+    mykb.add_entity(entity="Q1", freq=27, entity_vector=[1])
     q2_hash = mykb.add_entity(entity="Q2", freq=12, entity_vector=[2])
-    q3_hash = mykb.add_entity(entity="Q3", freq=5, entity_vector=[3])
+    mykb.add_entity(entity="Q3", freq=5, entity_vector=[3])
 
     # adding aliases
-    douglas_hash = mykb.add_alias(
+    mykb.add_alias(alias="douglas", entities=["Q2", "Q3"], probabilities=[0.4, 0.1])
-        alias="douglas", entities=["Q2", "Q3"], probabilities=[0.4, 0.1]
-    )
     adam_hash = mykb.add_alias(alias="adam", entities=["Q2"], probabilities=[0.9])
 
     candidates = mykb.get_alias_candidates("adam")
@@ -360,7 +358,7 @@ def test_preserving_links_asdoc(nlp):
     ruler.add_patterns(patterns)
     el_config = {"kb_loader": {"@misc": "myLocationsKB.v1"}, "incl_prior": False}
     entity_linker = nlp.add_pipe("entity_linker", config=el_config, last=True)
-    nlp.begin_training()
+    nlp.initialize()
     assert entity_linker.model.get_dim("nO") == vector_length
 
     # test whether the entity links are preserved by the `as_doc()` function
@@ -463,7 +461,7 @@ def test_overfitting_IO():
     )
 
     # train the NEL pipe
-    optimizer = nlp.begin_training(get_examples=lambda: train_examples)
+    optimizer = nlp.initialize(get_examples=lambda: train_examples)
     assert entity_linker.model.get_dim("nO") == vector_length
     assert entity_linker.model.get_dim("nO") == entity_linker.kb.entity_vector_length
 

spacy/tests/pipeline/test_initialize.py

@@ -0,0 +1,69 @@
+import pytest
+from spacy.language import Language
+from spacy.lang.en import English
+from spacy.training import Example
+from thinc.api import ConfigValidationError
+from pydantic import StrictBool
+
+
+def test_initialize_arguments():
+    name = "test_initialize_arguments"
+
+    class CustomTokenizer:
+        def __init__(self, tokenizer):
+            self.tokenizer = tokenizer
+            self.from_initialize = None
+
+        def __call__(self, text):
+            return self.tokenizer(text)
+
+        def initialize(self, get_examples, nlp, custom: int):
+            self.from_initialize = custom
+
+    class Component:
+        def __init__(self):
+            self.from_initialize = None
+
+        def initialize(
+            self, get_examples, nlp, custom1: str, custom2: StrictBool = False
+        ):
+            self.from_initialize = (custom1, custom2)
+
+    Language.factory(name, func=lambda nlp, name: Component())
+
+    nlp = English()
+    nlp.tokenizer = CustomTokenizer(nlp.tokenizer)
+    example = Example.from_dict(nlp("x"), {})
+    get_examples = lambda: [example]
+    nlp.add_pipe(name)
+    # The settings here will typically come from the [initialize] block
+    init_cfg = {"tokenizer": {"custom": 1}, "components": {name: {}}}
+    nlp.config["initialize"].update(init_cfg)
+    with pytest.raises(ConfigValidationError) as e:
+        # Empty config for component, no required custom1 argument
+        nlp.initialize(get_examples)
+    errors = e.value.errors
+    assert len(errors) == 1
+    assert errors[0]["loc"] == ("custom1",)
+    assert errors[0]["type"] == "value_error.missing"
+    init_cfg = {
+        "tokenizer": {"custom": 1},
+        "components": {name: {"custom1": "x", "custom2": 1}},
+    }
+    nlp.config["initialize"].update(init_cfg)
+    with pytest.raises(ConfigValidationError) as e:
+        # Wrong type of custom 2
+        nlp.initialize(get_examples)
+    errors = e.value.errors
+    assert len(errors) == 1
+    assert errors[0]["loc"] == ("custom2",)
+    assert errors[0]["type"] == "value_error.strictbool"
+    init_cfg = {
+        "tokenizer": {"custom": 1},
+        "components": {name: {"custom1": "x"}},
+    }
+    nlp.config["initialize"].update(init_cfg)
+    nlp.initialize(get_examples)
+    assert nlp.tokenizer.from_initialize == 1
+    pipe = nlp.get_pipe(name)
+    assert pipe.from_initialize == ("x", False)

spacy/tests/pipeline/test_lemmatizer.py

@@ -8,61 +8,52 @@ from ..util import make_tempdir
 
 @pytest.fixture
 def nlp():
-    return English()
-
-
-@pytest.fixture
-def lemmatizer(nlp):
     @registry.misc("cope_lookups")
     def cope_lookups():
         lookups = Lookups()
-        lookups.add_table("lemma_lookup", {"cope": "cope"})
+        lookups.add_table("lemma_lookup", {"cope": "cope", "coped": "cope"})
         lookups.add_table("lemma_index", {"verb": ("cope", "cop")})
         lookups.add_table("lemma_exc", {"verb": {"coping": ("cope",)}})
         lookups.add_table("lemma_rules", {"verb": [["ing", ""]]})
         return lookups
 
-    lemmatizer = nlp.add_pipe(
+    nlp = English()
-        "lemmatizer", config={"mode": "rule", "lookups": {"@misc": "cope_lookups"}}
+    nlp.config["initialize"]["components"]["lemmatizer"] = {
-    )
+        "lookups": {"@misc": "cope_lookups"}
-    return lemmatizer
+    }
+    return nlp
 
 
 def test_lemmatizer_init(nlp):
-    @registry.misc("cope_lookups")
+    lemmatizer = nlp.add_pipe("lemmatizer", config={"mode": "lookup"})
-    def cope_lookups():
-        lookups = Lookups()
-        lookups.add_table("lemma_lookup", {"cope": "cope"})
-        lookups.add_table("lemma_index", {"verb": ("cope", "cop")})
-        lookups.add_table("lemma_exc", {"verb": {"coping": ("cope",)}})
-        lookups.add_table("lemma_rules", {"verb": [["ing", ""]]})
-        return lookups
-
-    lemmatizer = nlp.add_pipe(
-        "lemmatizer", config={"mode": "lookup", "lookups": {"@misc": "cope_lookups"}}
-    )
     assert isinstance(lemmatizer.lookups, Lookups)
+    assert not lemmatizer.lookups.tables
     assert lemmatizer.mode == "lookup"
+    with pytest.raises(ValueError):
+        nlp("test")
+    nlp.initialize()
+    assert lemmatizer.lookups.tables
+    assert nlp("cope")[0].lemma_ == "cope"
+    assert nlp("coped")[0].lemma_ == "cope"
     # replace any tables from spacy-lookups-data
     lemmatizer.lookups = Lookups()
-    doc = nlp("coping")
     # lookup with no tables sets text as lemma
-    assert doc[0].lemma_ == "coping"
+    assert nlp("cope")[0].lemma_ == "cope"
-
+    assert nlp("coped")[0].lemma_ == "coped"
     nlp.remove_pipe("lemmatizer")
-
+    lemmatizer = nlp.add_pipe("lemmatizer", config={"mode": "lookup"})
-    @registry.misc("empty_lookups")
-    def empty_lookups():
-        return Lookups()
-
     with pytest.raises(ValueError):
-        nlp.add_pipe(
+        # Can't initialize without required tables
-            "lemmatizer",
+        lemmatizer.initialize(lookups=Lookups())
-            config={"mode": "lookup", "lookups": {"@misc": "empty_lookups"}},
+    lookups = Lookups()
-        )
+    lookups.add_table("lemma_lookup", {})
+    lemmatizer.initialize(lookups=lookups)
 
 
-def test_lemmatizer_config(nlp, lemmatizer):
+def test_lemmatizer_config(nlp):
+    lemmatizer = nlp.add_pipe("lemmatizer", config={"mode": "rule"})
+    nlp.initialize()
+
     doc = nlp.make_doc("coping")
     doc[0].pos_ = "VERB"
     assert doc[0].lemma_ == ""
@@ -78,20 +69,21 @@ def test_lemmatizer_config(nlp, lemmatizer):
     assert doc[0].lemma_ == "cope"
 
 
-def test_lemmatizer_serialize(nlp, lemmatizer):
+def test_lemmatizer_serialize(nlp):
-    @registry.misc("cope_lookups")
+    lemmatizer = nlp.add_pipe("lemmatizer", config={"mode": "rule"})
+    nlp.initialize()
+
     def cope_lookups():
         lookups = Lookups()
-        lookups.add_table("lemma_lookup", {"cope": "cope"})
+        lookups.add_table("lemma_lookup", {"cope": "cope", "coped": "cope"})
         lookups.add_table("lemma_index", {"verb": ("cope", "cop")})
         lookups.add_table("lemma_exc", {"verb": {"coping": ("cope",)}})
         lookups.add_table("lemma_rules", {"verb": [["ing", ""]]})
         return lookups
 
     nlp2 = English()
-    lemmatizer2 = nlp2.add_pipe(
+    lemmatizer2 = nlp2.add_pipe("lemmatizer", config={"mode": "rule"})
-        "lemmatizer", config={"mode": "rule", "lookups": {"@misc": "cope_lookups"}}
+    lemmatizer2.initialize(lookups=cope_lookups())
-    )
     lemmatizer2.from_bytes(lemmatizer.to_bytes())
     assert lemmatizer.to_bytes() == lemmatizer2.to_bytes()
     assert lemmatizer.lookups.tables == lemmatizer2.lookups.tables

spacy/tests/pipeline/test_morphologizer.py

@@ -33,7 +33,7 @@ def test_no_label():
     nlp = Language()
     nlp.add_pipe("morphologizer")
     with pytest.raises(ValueError):
-        nlp.begin_training()
+        nlp.initialize()
 
 
 def test_implicit_label():
@@ -42,7 +42,7 @@ def test_implicit_label():
     train_examples = []
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
-    nlp.begin_training(get_examples=lambda: train_examples)
+    nlp.initialize(get_examples=lambda: train_examples)
 
 
 def test_no_resize():
@@ -50,13 +50,13 @@ def test_no_resize():
     morphologizer = nlp.add_pipe("morphologizer")
     morphologizer.add_label("POS" + Morphology.FIELD_SEP + "NOUN")
     morphologizer.add_label("POS" + Morphology.FIELD_SEP + "VERB")
-    nlp.begin_training()
+    nlp.initialize()
     # this throws an error because the morphologizer can't be resized after initialization
     with pytest.raises(ValueError):
         morphologizer.add_label("POS" + Morphology.FIELD_SEP + "ADJ")
 
 
-def test_begin_training_examples():
+def test_initialize_examples():
     nlp = Language()
     morphologizer = nlp.add_pipe("morphologizer")
     morphologizer.add_label("POS" + Morphology.FIELD_SEP + "NOUN")
@@ -64,12 +64,12 @@ def test_begin_training_examples():
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
     # you shouldn't really call this more than once, but for testing it should be fine
-    nlp.begin_training()
+    nlp.initialize()
-    nlp.begin_training(get_examples=lambda: train_examples)
+    nlp.initialize(get_examples=lambda: train_examples)
-    with pytest.raises(TypeError):
-        nlp.begin_training(get_examples=lambda: None)
     with pytest.raises(ValueError):
-        nlp.begin_training(get_examples=train_examples)
+        nlp.initialize(get_examples=lambda: None)
+    with pytest.raises(ValueError):
+        nlp.initialize(get_examples=train_examples)
 
 
 def test_overfitting_IO():
@@ -79,7 +79,7 @@ def test_overfitting_IO():
     train_examples = []
     for inst in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(inst[0]), inst[1]))
-    optimizer = nlp.begin_training(get_examples=lambda: train_examples)
+    optimizer = nlp.initialize(get_examples=lambda: train_examples)
 
     for i in range(50):
         losses = {}
@@ -91,7 +91,7 @@ def test_overfitting_IO():
     doc = nlp(test_text)
     gold_morphs = ["Feat=N", "Feat=V", "", ""]
     gold_pos_tags = ["NOUN", "VERB", "ADJ", ""]
-    assert [t.morph_ for t in doc] == gold_morphs
+    assert [str(t.morph) for t in doc] == gold_morphs
     assert [t.pos_ for t in doc] == gold_pos_tags
 
     # Also test the results are still the same after IO
@@ -99,5 +99,5 @@ def test_overfitting_IO():
         nlp.to_disk(tmp_dir)
         nlp2 = util.load_model_from_path(tmp_dir)
         doc2 = nlp2(test_text)
-        assert [t.morph_ for t in doc2] == gold_morphs
+        assert [str(t.morph) for t in doc2] == gold_morphs
         assert [t.pos_ for t in doc2] == gold_pos_tags

spacy/tests/pipeline/test_senter.py

@@ -31,19 +31,19 @@ TRAIN_DATA = [
 ]
 
 
-def test_begin_training_examples():
+def test_initialize_examples():
     nlp = Language()
     nlp.add_pipe("senter")
     train_examples = []
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
     # you shouldn't really call this more than once, but for testing it should be fine
-    nlp.begin_training()
+    nlp.initialize()
-    nlp.begin_training(get_examples=lambda: train_examples)
+    nlp.initialize(get_examples=lambda: train_examples)
-    with pytest.raises(TypeError):
-        nlp.begin_training(get_examples=lambda: None)
     with pytest.raises(ValueError):
-        nlp.begin_training(get_examples=train_examples)
+        nlp.initialize(get_examples=lambda: None)
+    with pytest.raises(ValueError):
+        nlp.initialize(get_examples=train_examples)
 
 
 def test_overfitting_IO():
@@ -58,7 +58,7 @@ def test_overfitting_IO():
     train_examples[1].reference[11].is_sent_start = False
 
     nlp.add_pipe("senter")
-    optimizer = nlp.begin_training()
+    optimizer = nlp.initialize()
 
     for i in range(200):
         losses = {}

spacy/tests/pipeline/test_tagger.py

@@ -15,14 +15,14 @@ def test_label_types():
         tagger.add_label(9)
 
 
-def test_tagger_begin_training_tag_map():
+def test_tagger_initialize_tag_map():
-    """Test that Tagger.begin_training() without gold tuples does not clobber
+    """Test that Tagger.initialize() without gold tuples does not clobber
     the tag map."""
     nlp = Language()
     tagger = nlp.add_pipe("tagger")
     orig_tag_count = len(tagger.labels)
     tagger.add_label("A")
-    nlp.begin_training()
+    nlp.initialize()
     assert orig_tag_count + 1 == len(nlp.get_pipe("tagger").labels)
 
 
@@ -38,7 +38,7 @@ def test_no_label():
     nlp = Language()
     nlp.add_pipe("tagger")
     with pytest.raises(ValueError):
-        nlp.begin_training()
+        nlp.initialize()
 
 
 def test_no_resize():
@@ -47,7 +47,7 @@ def test_no_resize():
     tagger.add_label("N")
     tagger.add_label("V")
     assert tagger.labels == ("N", "V")
-    nlp.begin_training()
+    nlp.initialize()
     assert tagger.model.get_dim("nO") == 2
     # this throws an error because the tagger can't be resized after initialization
     with pytest.raises(ValueError):
@@ -60,10 +60,10 @@ def test_implicit_label():
     train_examples = []
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
-    nlp.begin_training(get_examples=lambda: train_examples)
+    nlp.initialize(get_examples=lambda: train_examples)
 
 
-def test_begin_training_examples():
+def test_initialize_examples():
     nlp = Language()
     tagger = nlp.add_pipe("tagger")
     train_examples = []
@@ -72,16 +72,16 @@ def test_begin_training_examples():
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
     # you shouldn't really call this more than once, but for testing it should be fine
-    nlp.begin_training()
+    nlp.initialize()
-    nlp.begin_training(get_examples=lambda: train_examples)
+    nlp.initialize(get_examples=lambda: train_examples)
-    with pytest.raises(TypeError):
-        nlp.begin_training(get_examples=lambda: None)
-    with pytest.raises(TypeError):
-        nlp.begin_training(get_examples=lambda: train_examples[0])
     with pytest.raises(ValueError):
-        nlp.begin_training(get_examples=lambda: [])
+        nlp.initialize(get_examples=lambda: None)
+    with pytest.raises(TypeError):
+        nlp.initialize(get_examples=lambda: train_examples[0])
     with pytest.raises(ValueError):
-        nlp.begin_training(get_examples=train_examples)
+        nlp.initialize(get_examples=lambda: [])
+    with pytest.raises(ValueError):
+        nlp.initialize(get_examples=train_examples)
 
 
 def test_overfitting_IO():
@@ -91,7 +91,7 @@ def test_overfitting_IO():
     train_examples = []
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
-    optimizer = nlp.begin_training(get_examples=lambda: train_examples)
+    optimizer = nlp.initialize(get_examples=lambda: train_examples)
     assert tagger.model.get_dim("nO") == len(TAGS)
 
     for i in range(50):
@@ -122,4 +122,4 @@ def test_tagger_requires_labels():
     nlp = English()
     nlp.add_pipe("tagger")
     with pytest.raises(ValueError):
-        nlp.begin_training()
+        nlp.initialize()

spacy/tests/pipeline/test_textcat.py

@@ -9,10 +9,10 @@ from spacy.pipeline import TextCategorizer
 from spacy.tokens import Doc
 from spacy.pipeline.tok2vec import DEFAULT_TOK2VEC_MODEL
 from spacy.scorer import Scorer
+from spacy.training import Example
+from spacy.training.initialize import verify_textcat_config
 
 from ..util import make_tempdir
-from ...cli.train import verify_textcat_config
-from ...training import Example
 
 
 TRAIN_DATA = [
@@ -26,7 +26,7 @@ def test_simple_train():
     nlp = Language()
     textcat = nlp.add_pipe("textcat")
     textcat.add_label("answer")
-    nlp.begin_training()
+    nlp.initialize()
     for i in range(5):
         for text, answer in [
             ("aaaa", 1.0),
@@ -56,7 +56,7 @@ def test_textcat_learns_multilabel():
     textcat = TextCategorizer(nlp.vocab, width=8)
     for letter in letters:
         textcat.add_label(letter)
-    optimizer = textcat.begin_training(lambda: [])
+    optimizer = textcat.initialize(lambda: [])
     for i in range(30):
         losses = {}
         examples = [Example.from_dict(doc, {"cats": cats}) for doc, cat in docs]
@@ -86,7 +86,7 @@ def test_no_label():
     nlp = Language()
     nlp.add_pipe("textcat")
     with pytest.raises(ValueError):
-        nlp.begin_training()
+        nlp.initialize()
 
 
 def test_implicit_label():
@@ -95,7 +95,7 @@ def test_implicit_label():
     train_examples = []
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
-    nlp.begin_training(get_examples=lambda: train_examples)
+    nlp.initialize(get_examples=lambda: train_examples)
 
 
 def test_no_resize():
@@ -103,14 +103,14 @@ def test_no_resize():
     textcat = nlp.add_pipe("textcat")
     textcat.add_label("POSITIVE")
     textcat.add_label("NEGATIVE")
-    nlp.begin_training()
+    nlp.initialize()
     assert textcat.model.get_dim("nO") == 2
     # this throws an error because the textcat can't be resized after initialization
     with pytest.raises(ValueError):
         textcat.add_label("NEUTRAL")
 
 
-def test_begin_training_examples():
+def test_initialize_examples():
     nlp = Language()
     textcat = nlp.add_pipe("textcat")
     train_examples = []
@@ -119,12 +119,12 @@ def test_begin_training_examples():
         for label, value in annotations.get("cats").items():
             textcat.add_label(label)
     # you shouldn't really call this more than once, but for testing it should be fine
-    nlp.begin_training()
+    nlp.initialize()
-    nlp.begin_training(get_examples=lambda: train_examples)
+    nlp.initialize(get_examples=lambda: train_examples)
-    with pytest.raises(TypeError):
-        nlp.begin_training(get_examples=lambda: None)
     with pytest.raises(ValueError):
-        nlp.begin_training(get_examples=train_examples)
+        nlp.initialize(get_examples=lambda: None)
+    with pytest.raises(ValueError):
+        nlp.initialize(get_examples=train_examples)
 
 
 def test_overfitting_IO():
@@ -139,7 +139,7 @@ def test_overfitting_IO():
     train_examples = []
     for text, annotations in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(text), annotations))
-    optimizer = nlp.begin_training(get_examples=lambda: train_examples)
+    optimizer = nlp.initialize(get_examples=lambda: train_examples)
     assert textcat.model.get_dim("nO") == 2
 
     for i in range(50):
@@ -195,7 +195,7 @@ def test_textcat_configs(textcat_config):
         train_examples.append(Example.from_dict(nlp.make_doc(text), annotations))
         for label, value in annotations.get("cats").items():
             textcat.add_label(label)
-    optimizer = nlp.begin_training()
+    optimizer = nlp.initialize()
     for i in range(5):
         losses = {}
         nlp.update(train_examples, sgd=optimizer, losses=losses)
@@ -226,6 +226,7 @@ def test_positive_class_not_binary():
     with pytest.raises(ValueError):
         verify_textcat_config(nlp, pipe_config)
 
+
 def test_textcat_evaluation():
     train_examples = []
     nlp = English()
@@ -241,7 +242,9 @@ def test_textcat_evaluation():
     pred2.cats = {"winter": 1.0, "summer": 0.0, "spring": 0.0, "autumn": 1.0}
     train_examples.append(Example(pred2, ref2))
 
-    scores = Scorer().score_cats(train_examples, "cats", labels=["winter", "summer", "spring", "autumn"])
+    scores = Scorer().score_cats(
+        train_examples, "cats", labels=["winter", "summer", "spring", "autumn"]
+    )
     assert scores["cats_f_per_type"]["winter"]["p"] == 1 / 2
     assert scores["cats_f_per_type"]["winter"]["r"] == 1 / 1
     assert scores["cats_f_per_type"]["summer"]["p"] == 0

spacy/tests/pipeline/test_tok2vec.py

@@ -73,8 +73,7 @@ def test_tok2vec_configs(width, embed_arch, embed_config, encode_arch, encode_co
     encode_config["width"] = width
     docs = get_batch(3)
     tok2vec = build_Tok2Vec_model(
-        embed_arch(**embed_config),
+        embed_arch(**embed_config), encode_arch(**encode_config)
-        encode_arch(**encode_config)
     )
     tok2vec.initialize(docs)
     vectors, backprop = tok2vec.begin_update(docs)
@@ -88,7 +87,7 @@ def test_init_tok2vec():
     nlp = English()
     tok2vec = nlp.add_pipe("tok2vec")
     assert tok2vec.listeners == []
-    nlp.begin_training()
+    nlp.initialize()
     assert tok2vec.model.get_dim("nO")
 
 
@@ -154,7 +153,7 @@ def test_tok2vec_listener():
 
     # Check that the Tok2Vec component finds it listeners
     assert tok2vec.listeners == []
-    optimizer = nlp.begin_training(lambda: train_examples)
+    optimizer = nlp.initialize(lambda: train_examples)
     assert tok2vec.listeners == [tagger_tok2vec]
 
     for i in range(5):

spacy/tests/regression/test_issue1-1000.py

@@ -428,7 +428,7 @@ def test_issue999():
     for _, offsets in TRAIN_DATA:
         for start, end, label in offsets:
             ner.add_label(label)
-    nlp.begin_training()
+    nlp.initialize()
     for itn in range(20):
         random.shuffle(TRAIN_DATA)
         for raw_text, entity_offsets in TRAIN_DATA:

spacy/tests/regression/test_issue1501-2000.py

@@ -250,7 +250,7 @@ def test_issue1915():
     ner = nlp.add_pipe("ner")
     ner.add_label("answer")
     with pytest.raises(ValueError):
-        nlp.begin_training(**cfg)
+        nlp.initialize(**cfg)
 
 
 def test_issue1945():