spaCy/spacy/cli/train.py
adrianeboyd 44829950ba Fix Example details for train CLI / pipeline components (#4624)
* Switch to train_dataset() function in train CLI

* Fixes for pipe() methods in pipeline components

* Don't clobber `examples` variable with `as_example` in pipe() methods
* Remove unnecessary traversals of `examples`

* Update Parser.pipe() for Examples

* Add `as_examples` kwarg to `pipe()` with implementation to return
`Example`s

* Accept `Doc` or `Example` in `pipe()` with `_get_doc()` (copied from
`Pipe`)

* Fixes to Example implementation in spacy.gold

* Move `make_projective` from an attribute of Example to an argument of
`Example.get_gold_parses()`

* Head of 0 are not treated as unset

* Unset heads are set to self rather than `None` (which causes problems
while projectivizing)

* Check for `Doc` (not just not `None`) when creating GoldParses for
pre-merged example

* Don't clobber `examples` variable in `iter_gold_docs()`

* Add/modify gold tests for handling projectivity

* In JSON roundtrip compare results from `dev_dataset` rather than
`train_dataset` to avoid projectivization (and other potential
modifications)

* Add test for projective train vs. nonprojective dev versions of the
same `Doc`

* Handle ignore_misaligned as arg rather than attr

Move `ignore_misaligned` from an attribute of `Example` to an argument
to `Example.get_gold_parses()`, which makes it parallel to
`make_projective`.

Add test with old and new align that checks whether `ignore_misaligned`
errors are raised as expected (only for new align).

* Remove unused attrs from gold.pxd

Remove `ignore_misaligned` and `make_projective` from `gold.pxd`

* Refer to Example.goldparse in iter_gold_docs()

Use `Example.goldparse` in `iter_gold_docs()` instead of `Example.gold`
because a `None` `GoldParse` is generated with ignore_misaligned and
generating it on-the-fly can raise an unwanted AlignmentError

* Update test for ignore_misaligned
2019-11-23 14:32:15 +01:00

648 lines
27 KiB
Python

# coding: utf8
from __future__ import unicode_literals, division, print_function
import plac
import os
from pathlib import Path
from thinc.neural._classes.model import Model
from timeit import default_timer as timer
import shutil
import srsly
from wasabi import msg
import contextlib
import random
from .._ml import create_default_optimizer
from ..attrs import PROB, IS_OOV, CLUSTER, LANG
from ..gold import GoldCorpus
from ..compat import path2str
from .. import util
from .. import about
@plac.annotations(
# fmt: off
lang=("Model language", "positional", None, str),
output_path=("Output directory to store model in", "positional", None, Path),
train_path=("Location of JSON-formatted training data", "positional", None, Path),
dev_path=("Location of JSON-formatted development data", "positional", None, Path),
raw_text=("Path to jsonl file with unlabelled text documents.", "option", "rt", Path),
base_model=("Name of model to update (optional)", "option", "b", str),
pipeline=("Comma-separated names of pipeline components", "option", "p", str),
vectors=("Model to load vectors from", "option", "v", str),
n_iter=("Number of iterations", "option", "n", int),
n_early_stopping=("Maximum number of training epochs without dev accuracy improvement", "option", "ne", int),
n_examples=("Number of examples", "option", "ns", int),
use_gpu=("Use GPU", "option", "g", int),
version=("Model version", "option", "V", str),
meta_path=("Optional path to meta.json to use as base.", "option", "m", Path),
init_tok2vec=("Path to pretrained weights for the token-to-vector parts of the models. See 'spacy pretrain'. Experimental.", "option", "t2v", Path),
parser_multitasks=("Side objectives for parser CNN, e.g. 'dep' or 'dep,tag'", "option", "pt", str),
entity_multitasks=("Side objectives for NER CNN, e.g. 'dep' or 'dep,tag'", "option", "et", str),
noise_level=("Amount of corruption for data augmentation", "option", "nl", float),
orth_variant_level=("Amount of orthography variation for data augmentation", "option", "ovl", float),
eval_beam_widths=("Beam widths to evaluate, e.g. 4,8", "option", "bw", str),
gold_preproc=("Use gold preprocessing", "flag", "G", bool),
learn_tokens=("Make parser learn gold-standard tokenization", "flag", "T", bool),
textcat_multilabel=("Textcat classes aren't mutually exclusive (multilabel)", "flag", "TML", bool),
textcat_arch=("Textcat model architecture", "option", "ta", str),
textcat_positive_label=("Textcat positive label for binary classes with two labels", "option", "tpl", str),
verbose=("Display more information for debug", "flag", "VV", bool),
debug=("Run data diagnostics before training", "flag", "D", bool),
# fmt: on
)
def train(
lang,
output_path,
train_path,
dev_path,
raw_text=None,
base_model=None,
pipeline="tagger,parser,ner",
vectors=None,
n_iter=30,
n_early_stopping=None,
n_examples=0,
use_gpu=-1,
version="0.0.0",
meta_path=None,
init_tok2vec=None,
parser_multitasks="",
entity_multitasks="",
noise_level=0.0,
orth_variant_level=0.0,
eval_beam_widths="",
gold_preproc=False,
learn_tokens=False,
textcat_multilabel=False,
textcat_arch="bow",
textcat_positive_label=None,
verbose=False,
debug=False,
):
"""
Train or update a spaCy model. Requires data to be formatted in spaCy's
JSON format. To convert data from other formats, use the `spacy convert`
command.
"""
# temp fix to avoid import issues cf https://github.com/explosion/spaCy/issues/4200
import tqdm
util.fix_random_seed()
util.set_env_log(verbose)
# Make sure all files and paths exists if they are needed
train_path = util.ensure_path(train_path)
dev_path = util.ensure_path(dev_path)
meta_path = util.ensure_path(meta_path)
output_path = util.ensure_path(output_path)
if raw_text is not None:
raw_text = list(srsly.read_jsonl(raw_text))
if not train_path or not train_path.exists():
msg.fail("Training data not found", train_path, exits=1)
if not dev_path or not dev_path.exists():
msg.fail("Development data not found", dev_path, exits=1)
if meta_path is not None and not meta_path.exists():
msg.fail("Can't find model meta.json", meta_path, exits=1)
meta = srsly.read_json(meta_path) if meta_path else {}
if output_path.exists() and [p for p in output_path.iterdir() if p.is_dir()]:
msg.warn(
"Output directory is not empty",
"This can lead to unintended side effects when saving the model. "
"Please use an empty directory or a different path instead. If "
"the specified output path doesn't exist, the directory will be "
"created for you.",
)
if not output_path.exists():
output_path.mkdir()
# Take dropout and batch size as generators of values -- dropout
# starts high and decays sharply, to force the optimizer to explore.
# Batch size starts at 1 and grows, so that we make updates quickly
# at the beginning of training.
dropout_rates = util.decaying(
util.env_opt("dropout_from", 0.2),
util.env_opt("dropout_to", 0.2),
util.env_opt("dropout_decay", 0.0),
)
batch_sizes = util.compounding(
util.env_opt("batch_from", 100.0),
util.env_opt("batch_to", 1000.0),
util.env_opt("batch_compound", 1.001),
)
if not eval_beam_widths:
eval_beam_widths = [1]
else:
eval_beam_widths = [int(bw) for bw in eval_beam_widths.split(",")]
if 1 not in eval_beam_widths:
eval_beam_widths.append(1)
eval_beam_widths.sort()
has_beam_widths = eval_beam_widths != [1]
# Set up the base model and pipeline. If a base model is specified, load
# the model and make sure the pipeline matches the pipeline setting. If
# training starts from a blank model, intitalize the language class.
pipeline = [p.strip() for p in pipeline.split(",")]
msg.text("Training pipeline: {}".format(pipeline))
if base_model:
msg.text("Starting with base model '{}'".format(base_model))
nlp = util.load_model(base_model)
if nlp.lang != lang:
msg.fail(
"Model language ('{}') doesn't match language specified as "
"`lang` argument ('{}') ".format(nlp.lang, lang),
exits=1,
)
nlp.disable_pipes([p for p in nlp.pipe_names if p not in pipeline])
for pipe in pipeline:
if pipe not in nlp.pipe_names:
if pipe == "parser":
pipe_cfg = {"learn_tokens": learn_tokens}
elif pipe == "textcat":
pipe_cfg = {
"exclusive_classes": not textcat_multilabel,
"architecture": textcat_arch,
"positive_label": textcat_positive_label,
}
else:
pipe_cfg = {}
nlp.add_pipe(nlp.create_pipe(pipe, config=pipe_cfg))
else:
if pipe == "textcat":
textcat_cfg = nlp.get_pipe("textcat").cfg
base_cfg = {
"exclusive_classes": textcat_cfg["exclusive_classes"],
"architecture": textcat_cfg["architecture"],
"positive_label": textcat_cfg["positive_label"],
}
pipe_cfg = {
"exclusive_classes": not textcat_multilabel,
"architecture": textcat_arch,
"positive_label": textcat_positive_label,
}
if base_cfg != pipe_cfg:
msg.fail(
"The base textcat model configuration does"
"not match the provided training options. "
"Existing cfg: {}, provided cfg: {}".format(
base_cfg, pipe_cfg
),
exits=1,
)
else:
msg.text("Starting with blank model '{}'".format(lang))
lang_cls = util.get_lang_class(lang)
nlp = lang_cls()
for pipe in pipeline:
if pipe == "parser":
pipe_cfg = {"learn_tokens": learn_tokens}
elif pipe == "textcat":
pipe_cfg = {
"exclusive_classes": not textcat_multilabel,
"architecture": textcat_arch,
"positive_label": textcat_positive_label,
}
else:
pipe_cfg = {}
nlp.add_pipe(nlp.create_pipe(pipe, config=pipe_cfg))
if vectors:
msg.text("Loading vector from model '{}'".format(vectors))
_load_vectors(nlp, vectors)
# Multitask objectives
multitask_options = [("parser", parser_multitasks), ("ner", entity_multitasks)]
for pipe_name, multitasks in multitask_options:
if multitasks:
if pipe_name not in pipeline:
msg.fail(
"Can't use multitask objective without '{}' in the "
"pipeline".format(pipe_name)
)
pipe = nlp.get_pipe(pipe_name)
for objective in multitasks.split(","):
pipe.add_multitask_objective(objective)
# Prepare training corpus
msg.text("Counting training words (limit={})".format(n_examples))
corpus = GoldCorpus(train_path, dev_path, limit=n_examples)
n_train_words = corpus.count_train()
if base_model:
# Start with an existing model, use default optimizer
optimizer = create_default_optimizer(Model.ops)
else:
# Start with a blank model, call begin_training
optimizer = nlp.begin_training(lambda: corpus.train_examples, device=use_gpu)
nlp._optimizer = None
# Load in pretrained weights
if init_tok2vec is not None:
components = _load_pretrained_tok2vec(nlp, init_tok2vec)
msg.text("Loaded pretrained tok2vec for: {}".format(components))
# Verify textcat config
if "textcat" in pipeline:
textcat_labels = nlp.get_pipe("textcat").cfg["labels"]
if textcat_positive_label and textcat_positive_label not in textcat_labels:
msg.fail(
"The textcat_positive_label (tpl) '{}' does not match any "
"label in the training data.".format(textcat_positive_label),
exits=1,
)
if textcat_positive_label and len(textcat_labels) != 2:
msg.fail(
"A textcat_positive_label (tpl) '{}' was provided for training "
"data that does not appear to be a binary classification "
"problem with two labels.".format(textcat_positive_label),
exits=1,
)
train_data = corpus.train_data(
nlp,
noise_level=noise_level,
gold_preproc=gold_preproc,
max_length=0,
ignore_misaligned=True,
)
train_labels = set()
if textcat_multilabel:
multilabel_found = False
for ex in train_data:
train_labels.update(ex.gold.cats.keys())
if list(ex.gold.cats.values()).count(1.0) != 1:
multilabel_found = True
if not multilabel_found and not base_model:
msg.warn(
"The textcat training instances look like they have "
"mutually-exclusive classes. Remove the flag "
"'--textcat-multilabel' to train a classifier with "
"mutually-exclusive classes."
)
if not textcat_multilabel:
for ex in train_data:
train_labels.update(ex.gold.cats.keys())
if list(ex.gold.cats.values()).count(1.0) != 1 and not base_model:
msg.warn(
"Some textcat training instances do not have exactly "
"one positive label. Modifying training options to "
"include the flag '--textcat-multilabel' for classes "
"that are not mutually exclusive."
)
nlp.get_pipe("textcat").cfg["exclusive_classes"] = False
textcat_multilabel = True
break
if base_model and set(textcat_labels) != train_labels:
msg.fail(
"Cannot extend textcat model using data with different "
"labels. Base model labels: {}, training data labels: "
"{}.".format(textcat_labels, list(train_labels)),
exits=1,
)
if textcat_multilabel:
msg.text(
"Textcat evaluation score: ROC AUC score macro-averaged across "
"the labels '{}'".format(", ".join(textcat_labels))
)
elif textcat_positive_label and len(textcat_labels) == 2:
msg.text(
"Textcat evaluation score: F1-score for the "
"label '{}'".format(textcat_positive_label)
)
elif len(textcat_labels) > 1:
if len(textcat_labels) == 2:
msg.warn(
"If the textcat component is a binary classifier with "
"exclusive classes, provide '--textcat_positive_label' for "
"an evaluation on the positive class."
)
msg.text(
"Textcat evaluation score: F1-score macro-averaged across "
"the labels '{}'".format(", ".join(textcat_labels))
)
else:
msg.fail(
"Unsupported textcat configuration. Use `spacy debug-data` "
"for more information."
)
# fmt: off
row_head, output_stats = _configure_training_output(pipeline, use_gpu, has_beam_widths)
row_widths = [len(w) for w in row_head]
row_settings = {"widths": row_widths, "aligns": tuple(["r" for i in row_head]), "spacing": 2}
# fmt: on
print("")
msg.row(row_head, **row_settings)
msg.row(["-" * width for width in row_settings["widths"]], **row_settings)
try:
iter_since_best = 0
best_score = 0.0
for i in range(n_iter):
train_data = corpus.train_dataset(
nlp,
noise_level=noise_level,
orth_variant_level=orth_variant_level,
gold_preproc=gold_preproc,
max_length=0,
ignore_misaligned=True,
)
if raw_text:
random.shuffle(raw_text)
raw_batches = util.minibatch(
(nlp.make_doc(rt["text"]) for rt in raw_text), size=8
)
words_seen = 0
with tqdm.tqdm(total=n_train_words, leave=False) as pbar:
losses = {}
for batch in util.minibatch_by_words(train_data, size=batch_sizes):
if not batch:
continue
nlp.update(
batch,
sgd=optimizer,
drop=next(dropout_rates),
losses=losses,
)
if raw_text:
# If raw text is available, perform 'rehearsal' updates,
# which use unlabelled data to reduce overfitting.
raw_batch = list(next(raw_batches))
nlp.rehearse(raw_batch, sgd=optimizer, losses=losses)
docs = [ex.doc for ex in batch]
if not int(os.environ.get("LOG_FRIENDLY", 0)):
pbar.update(sum(len(doc) for doc in docs))
words_seen += sum(len(doc) for doc in docs)
with nlp.use_params(optimizer.averages):
util.set_env_log(False)
epoch_model_path = output_path / ("model%d" % i)
nlp.to_disk(epoch_model_path)
nlp_loaded = util.load_model_from_path(epoch_model_path)
for beam_width in eval_beam_widths:
for name, component in nlp_loaded.pipeline:
if hasattr(component, "cfg"):
component.cfg["beam_width"] = beam_width
dev_dataset = list(
corpus.dev_dataset(
nlp_loaded,
gold_preproc=gold_preproc,
ignore_misaligned=True,
)
)
nwords = sum(len(ex.doc) for ex in dev_dataset)
start_time = timer()
scorer = nlp_loaded.evaluate(dev_dataset, verbose=verbose)
end_time = timer()
if use_gpu < 0:
gpu_wps = None
cpu_wps = nwords / (end_time - start_time)
else:
gpu_wps = nwords / (end_time - start_time)
with Model.use_device("cpu"):
nlp_loaded = util.load_model_from_path(epoch_model_path)
for name, component in nlp_loaded.pipeline:
if hasattr(component, "cfg"):
component.cfg["beam_width"] = beam_width
dev_dataset = list(
corpus.dev_dataset(
nlp_loaded,
gold_preproc=gold_preproc,
ignore_misaligned=True,
)
)
start_time = timer()
scorer = nlp_loaded.evaluate(dev_dataset, verbose=verbose)
end_time = timer()
cpu_wps = nwords / (end_time - start_time)
acc_loc = output_path / ("model%d" % i) / "accuracy.json"
srsly.write_json(acc_loc, scorer.scores)
# Update model meta.json
meta["lang"] = nlp.lang
meta["pipeline"] = nlp.pipe_names
meta["spacy_version"] = ">=%s" % about.__version__
if beam_width == 1:
meta["speed"] = {
"nwords": nwords,
"cpu": cpu_wps,
"gpu": gpu_wps,
}
meta["accuracy"] = scorer.scores
else:
meta.setdefault("beam_accuracy", {})
meta.setdefault("beam_speed", {})
meta["beam_accuracy"][beam_width] = scorer.scores
meta["beam_speed"][beam_width] = {
"nwords": nwords,
"cpu": cpu_wps,
"gpu": gpu_wps,
}
meta["vectors"] = {
"width": nlp.vocab.vectors_length,
"vectors": len(nlp.vocab.vectors),
"keys": nlp.vocab.vectors.n_keys,
"name": nlp.vocab.vectors.name,
}
meta.setdefault("name", "model%d" % i)
meta.setdefault("version", version)
meta["labels"] = nlp.meta["labels"]
meta_loc = output_path / ("model%d" % i) / "meta.json"
srsly.write_json(meta_loc, meta)
util.set_env_log(verbose)
progress = _get_progress(
i,
losses,
scorer.scores,
output_stats,
beam_width=beam_width if has_beam_widths else None,
cpu_wps=cpu_wps,
gpu_wps=gpu_wps,
)
if i == 0 and "textcat" in pipeline:
textcats_per_cat = scorer.scores.get("textcats_per_cat", {})
for cat, cat_score in textcats_per_cat.items():
if cat_score.get("roc_auc_score", 0) < 0:
msg.warn(
"Textcat ROC AUC score is undefined due to "
"only one value in label '{}'.".format(cat)
)
msg.row(progress, **row_settings)
# Early stopping
if n_early_stopping is not None:
current_score = _score_for_model(meta)
if current_score < best_score:
iter_since_best += 1
else:
iter_since_best = 0
best_score = current_score
if iter_since_best >= n_early_stopping:
msg.text(
"Early stopping, best iteration "
"is: {}".format(i - iter_since_best)
)
msg.text(
"Best score = {}; Final iteration "
"score = {}".format(best_score, current_score)
)
break
finally:
with nlp.use_params(optimizer.averages):
final_model_path = output_path / "model-final"
nlp.to_disk(final_model_path)
msg.good("Saved model to output directory", final_model_path)
with msg.loading("Creating best model..."):
best_model_path = _collate_best_model(meta, output_path, nlp.pipe_names)
msg.good("Created best model", best_model_path)
def _score_for_model(meta):
""" Returns mean score between tasks in pipeline that can be used for early stopping. """
mean_acc = list()
pipes = meta["pipeline"]
acc = meta["accuracy"]
if "tagger" in pipes:
mean_acc.append(acc["tags_acc"])
if "parser" in pipes:
mean_acc.append((acc["uas"] + acc["las"]) / 2)
if "ner" in pipes:
mean_acc.append((acc["ents_p"] + acc["ents_r"] + acc["ents_f"]) / 3)
if "textcat" in pipes:
mean_acc.append(acc["textcat_score"])
return sum(mean_acc) / len(mean_acc)
@contextlib.contextmanager
def _create_progress_bar(total):
# temp fix to avoid import issues cf https://github.com/explosion/spaCy/issues/4200
import tqdm
if int(os.environ.get("LOG_FRIENDLY", 0)):
yield
else:
pbar = tqdm.tqdm(total=total, leave=False)
yield pbar
def _load_vectors(nlp, vectors):
util.load_model(vectors, vocab=nlp.vocab)
for lex in nlp.vocab:
values = {}
for attr, func in nlp.vocab.lex_attr_getters.items():
# These attrs are expected to be set by data. Others should
# be set by calling the language functions.
if attr not in (CLUSTER, PROB, IS_OOV, LANG):
values[lex.vocab.strings[attr]] = func(lex.orth_)
lex.set_attrs(**values)
lex.is_oov = False
def _load_pretrained_tok2vec(nlp, loc):
"""Load pretrained weights for the 'token-to-vector' part of the component
models, which is typically a CNN. See 'spacy pretrain'. Experimental.
"""
with loc.open("rb") as file_:
weights_data = file_.read()
loaded = []
for name, component in nlp.pipeline:
if hasattr(component, "model") and hasattr(component.model, "tok2vec"):
component.tok2vec.from_bytes(weights_data)
loaded.append(name)
return loaded
def _collate_best_model(meta, output_path, components):
bests = {}
for component in components:
bests[component] = _find_best(output_path, component)
best_dest = output_path / "model-best"
shutil.copytree(path2str(output_path / "model-final"), path2str(best_dest))
for component, best_component_src in bests.items():
shutil.rmtree(path2str(best_dest / component))
shutil.copytree(
path2str(best_component_src / component), path2str(best_dest / component)
)
accs = srsly.read_json(best_component_src / "accuracy.json")
for metric in _get_metrics(component):
meta["accuracy"][metric] = accs[metric]
srsly.write_json(best_dest / "meta.json", meta)
return best_dest
def _find_best(experiment_dir, component):
accuracies = []
for epoch_model in experiment_dir.iterdir():
if epoch_model.is_dir() and epoch_model.parts[-1] != "model-final":
accs = srsly.read_json(epoch_model / "accuracy.json")
scores = [accs.get(metric, 0.0) for metric in _get_metrics(component)]
accuracies.append((scores, epoch_model))
if accuracies:
return max(accuracies)[1]
else:
return None
def _get_metrics(component):
if component == "parser":
return ("las", "uas", "token_acc")
elif component == "tagger":
return ("tags_acc",)
elif component == "ner":
return ("ents_f", "ents_p", "ents_r")
return ("token_acc",)
def _configure_training_output(pipeline, use_gpu, has_beam_widths):
row_head = ["Itn"]
output_stats = []
for pipe in pipeline:
if pipe == "tagger":
row_head.extend(["Tag Loss ", " Tag % "])
output_stats.extend(["tag_loss", "tags_acc"])
elif pipe == "parser":
row_head.extend(["Dep Loss ", " UAS ", " LAS "])
output_stats.extend(["dep_loss", "uas", "las"])
elif pipe == "ner":
row_head.extend(["NER Loss ", "NER P ", "NER R ", "NER F "])
output_stats.extend(["ner_loss", "ents_p", "ents_r", "ents_f"])
elif pipe == "textcat":
row_head.extend(["Textcat Loss", "Textcat"])
output_stats.extend(["textcat_loss", "textcat_score"])
row_head.extend(["Token %", "CPU WPS"])
output_stats.extend(["token_acc", "cpu_wps"])
if use_gpu >= 0:
row_head.extend(["GPU WPS"])
output_stats.extend(["gpu_wps"])
if has_beam_widths:
row_head.insert(1, "Beam W.")
return row_head, output_stats
def _get_progress(
itn, losses, dev_scores, output_stats, beam_width=None, cpu_wps=0.0, gpu_wps=0.0
):
scores = {}
for stat in output_stats:
scores[stat] = 0.0
scores["dep_loss"] = losses.get("parser", 0.0)
scores["ner_loss"] = losses.get("ner", 0.0)
scores["tag_loss"] = losses.get("tagger", 0.0)
scores["textcat_loss"] = losses.get("textcat", 0.0)
scores["cpu_wps"] = cpu_wps
scores["gpu_wps"] = gpu_wps or 0.0
scores.update(dev_scores)
formatted_scores = []
for stat in output_stats:
format_spec = "{:.3f}"
if stat.endswith("_wps"):
format_spec = "{:.0f}"
formatted_scores.append(format_spec.format(scores[stat]))
result = [itn + 1]
result.extend(formatted_scores)
if beam_width is not None:
result.insert(1, beam_width)
return result