spaCy/spacy/cli/debug_data.py
2020-08-17 21:38:20 +02:00

646 lines
26 KiB
Python

from typing import List, Sequence, Dict, Any, Tuple, Optional
from pathlib import Path
from collections import Counter
import sys
import srsly
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 ..gold import Corpus, Example
from ..pipeline._parser_internals import nonproj
from ..language import Language
from .. import util
# Minimum number of expected occurrences of NER label in data to train new label
NEW_LABEL_THRESHOLD = 50
# Minimum number of expected occurrences of dependency labels
DEP_LABEL_THRESHOLD = 20
# Minimum number of expected examples to train a blank model
BLANK_MODEL_MIN_THRESHOLD = 100
BLANK_MODEL_THRESHOLD = 2000
@debug_cli.command(
"config",
context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
)
def debug_config_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),
code_path: Optional[Path] = Opt(None, "--code-path", "-c", help="Path to Python file with additional code (registered functions) to be imported"),
# fmt: on
):
"""Debug a config.cfg file and show validation errors. The command will
create all objects in the tree and validate them. Note that some config
validation errors are blocking and will prevent the rest of the config from
being resolved. This means that you may not see all validation errors at
once and some issues are only shown once previous errors have been fixed.
Similar as with the 'train' command, you can override settings from the config
as command line options. For instance, --training.batch_size 128 overrides
the value of "batch_size" in the block "[training]".
"""
overrides = parse_config_overrides(ctx.args)
import_code(code_path)
with show_validation_error(config_path):
config = util.load_config(config_path, overrides=overrides)
nlp, _ = util.load_model_from_config(config)
msg.good("Original config is valid")
@debug_cli.command(
"data", context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
)
@app.command(
"debug-data",
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 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"),
verbose: bool = Opt(False, "--verbose", "-V", help="Print additional information and explanations"),
no_format: bool = Opt(False, "--no-format", "-NF", help="Don't pretty-print the results"),
# fmt: on
):
"""
Analyze, debug and validate your training and development data. Outputs
useful stats, and can help you find problems like invalid entity annotations,
cyclic dependencies, low data labels and more.
"""
if ctx.command.name == "debug-data":
msg.warn(
"The debug-data command is now available via the 'debug data' "
"subcommand (without the hyphen). You can run python -m spacy debug "
"--help for an overview of the other available debugging commands."
)
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,
verbose=verbose,
no_format=no_format,
silent=False,
)
def debug_data(
train_path: Path,
dev_path: Path,
config_path: Path,
*,
config_overrides: Dict[str, Any] = {},
ignore_warnings: bool = False,
verbose: bool = False,
no_format: bool = True,
silent: bool = True,
):
msg = Printer(
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, config = util.load_model_from_config(cfg)
# Use original config here, not resolved version
sourced_components = get_sourced_components(cfg)
frozen_components = config["training"]["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(config["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(config["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 = ""
loading_dev_error_message = ""
with msg.loading("Loading corpus..."):
try:
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")
# 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(
train_dataset, factory_names, nlp, make_proj=False
)
gold_dev_data = _compile_gold(dev_dataset, factory_names, nlp, make_proj=True)
train_texts = gold_train_data["texts"]
dev_texts = gold_dev_data["texts"]
frozen_components = config["training"]["frozen_components"]
msg.divider("Training stats")
msg.text(f"Language: {config['nlp']['lang']}")
msg.text(f"Training pipeline: {', '.join(pipeline)}")
if resume_components:
msg.text(f"Components from other models: {', '.join(resume_components)}")
if frozen_components:
msg.text(f"Frozen components: {', '.join(frozen_components)}")
msg.text(f"{len(train_dataset)} training docs")
msg.text(f"{len(dev_dataset)} evaluation docs")
if not len(gold_dev_data):
msg.fail("No evaluation docs")
overlap = len(train_texts.intersection(dev_texts))
if overlap:
msg.warn(f"{overlap} training examples also in evaluation data")
else:
msg.good("No overlap between training and evaluation data")
# TODO: make this feedback more fine-grained and report on updated
# components vs. blank components
if not resume_components and len(train_dataset) < BLANK_MODEL_THRESHOLD:
text = (
f"Low number of examples to train from a blank model ({len(train_dataset)})"
)
if len(train_dataset) < BLANK_MODEL_MIN_THRESHOLD:
msg.fail(text)
else:
msg.warn(text)
msg.text(
f"It's recommended to use at least {BLANK_MODEL_THRESHOLD} examples "
f"(minimum {BLANK_MODEL_MIN_THRESHOLD})",
show=verbose,
)
msg.divider("Vocab & Vectors")
n_words = gold_train_data["n_words"]
msg.info(
f"{n_words} total word(s) in the data ({len(gold_train_data['words'])} unique)"
)
if gold_train_data["n_misaligned_words"] > 0:
n_misaligned = gold_train_data["n_misaligned_words"]
msg.warn(f"{n_misaligned} misaligned tokens in the training data")
if gold_dev_data["n_misaligned_words"] > 0:
n_misaligned = gold_dev_data["n_misaligned_words"]
msg.warn(f"{n_misaligned} misaligned tokens in the dev data")
most_common_words = gold_train_data["words"].most_common(10)
msg.text(
f"10 most common words: {_format_labels(most_common_words, counts=True)}",
show=verbose,
)
if len(nlp.vocab.vectors):
msg.info(
f"{len(nlp.vocab.vectors)} vectors ({nlp.vocab.vectors.n_keys} "
f"unique keys, {nlp.vocab.vectors_length} dimensions)"
)
n_missing_vectors = sum(gold_train_data["words_missing_vectors"].values())
msg.warn(
"{} words in training data without vectors ({:0.2f}%)".format(
n_missing_vectors, n_missing_vectors / gold_train_data["n_words"],
),
)
msg.text(
"10 most common words without vectors: {}".format(
_format_labels(
gold_train_data["words_missing_vectors"].most_common(10),
counts=True,
)
),
show=verbose,
)
else:
msg.info("No word vectors present in the model")
if "ner" in factory_names:
# Get all unique NER labels present in the data
labels = set(
label for label in gold_train_data["ner"] if label not in ("O", "-", None)
)
label_counts = gold_train_data["ner"]
model_labels = _get_labels_from_model(nlp, "ner")
new_labels = [l for l in labels if l not in model_labels]
existing_labels = [l for l in labels if l in model_labels]
has_low_data_warning = False
has_no_neg_warning = False
has_ws_ents_error = False
has_punct_ents_warning = False
msg.divider("Named Entity Recognition")
msg.info(
f"{len(new_labels)} new label(s), {len(existing_labels)} existing label(s)"
)
missing_values = label_counts["-"]
msg.text(f"{missing_values} missing value(s) (tokens with '-' label)")
for label in new_labels:
if len(label) == 0:
msg.fail("Empty label found in new labels")
if new_labels:
labels_with_counts = [
(label, count)
for label, count in label_counts.most_common()
if label != "-"
]
labels_with_counts = _format_labels(labels_with_counts, counts=True)
msg.text(f"New: {labels_with_counts}", show=verbose)
if existing_labels:
msg.text(f"Existing: {_format_labels(existing_labels)}", show=verbose)
if gold_train_data["ws_ents"]:
msg.fail(f"{gold_train_data['ws_ents']} invalid whitespace entity spans")
has_ws_ents_error = True
if gold_train_data["punct_ents"]:
msg.warn(f"{gold_train_data['punct_ents']} entity span(s) with punctuation")
has_punct_ents_warning = True
for label in new_labels:
if label_counts[label] <= NEW_LABEL_THRESHOLD:
msg.warn(
f"Low number of examples for new label '{label}' ({label_counts[label]})"
)
has_low_data_warning = True
with msg.loading("Analyzing label distribution..."):
neg_docs = _get_examples_without_label(train_dataset, label)
if neg_docs == 0:
msg.warn(f"No examples for texts WITHOUT new label '{label}'")
has_no_neg_warning = True
if not has_low_data_warning:
msg.good("Good amount of examples for all labels")
if not has_no_neg_warning:
msg.good("Examples without occurrences available for all labels")
if not has_ws_ents_error:
msg.good("No entities consisting of or starting/ending with whitespace")
if not has_punct_ents_warning:
msg.good("No entities consisting of or starting/ending with punctuation")
if has_low_data_warning:
msg.text(
f"To train a new entity type, your data should include at "
f"least {NEW_LABEL_THRESHOLD} instances of the new label",
show=verbose,
)
if has_no_neg_warning:
msg.text(
"Training data should always include examples of entities "
"in context, as well as examples without a given entity "
"type.",
show=verbose,
)
if has_ws_ents_error:
msg.text(
"As of spaCy v2.1.0, entity spans consisting of or starting/ending "
"with whitespace characters are considered invalid."
)
if has_punct_ents_warning:
msg.text(
"Entity spans consisting of or starting/ending "
"with punctuation can not be trained with a noise level > 0."
)
if "textcat" in factory_names:
msg.divider("Text Classification")
labels = [label for label in gold_train_data["cats"]]
model_labels = _get_labels_from_model(nlp, "textcat")
new_labels = [l for l in labels if l not in model_labels]
existing_labels = [l for l in labels if l in model_labels]
msg.info(
f"Text Classification: {len(new_labels)} new label(s), "
f"{len(existing_labels)} existing label(s)"
)
if new_labels:
labels_with_counts = _format_labels(
gold_train_data["cats"].most_common(), counts=True
)
msg.text(f"New: {labels_with_counts}", show=verbose)
if existing_labels:
msg.text(f"Existing: {_format_labels(existing_labels)}", show=verbose)
if set(gold_train_data["cats"]) != set(gold_dev_data["cats"]):
msg.fail(
f"The train and dev labels are not the same. "
f"Train labels: {_format_labels(gold_train_data['cats'])}. "
f"Dev labels: {_format_labels(gold_dev_data['cats'])}."
)
if gold_train_data["n_cats_multilabel"] > 0:
msg.info(
"The train data contains instances without "
"mutually-exclusive classes. Use '--textcat-multilabel' "
"when training."
)
if gold_dev_data["n_cats_multilabel"] == 0:
msg.warn(
"Potential train/dev mismatch: the train data contains "
"instances without mutually-exclusive classes while the "
"dev data does not."
)
else:
msg.info(
"The train data contains only instances with "
"mutually-exclusive classes."
)
if gold_dev_data["n_cats_multilabel"] > 0:
msg.fail(
"Train/dev mismatch: the dev data contains instances "
"without mutually-exclusive classes while the train data "
"contains only instances with mutually-exclusive classes."
)
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
msg.info(f"{len(labels)} label(s) in data ({len(tag_map)} label(s) in tag map)")
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
msg.divider("Dependency Parsing")
# profile sentence length
msg.info(
f"Found {gold_train_data['n_sents']} sentence(s) with an average "
f"length of {gold_train_data['n_words'] / gold_train_data['n_sents']:.1f} words."
)
# check for documents with multiple sentences
sents_per_doc = gold_train_data["n_sents"] / len(gold_train_data["texts"])
if sents_per_doc < 1.1:
msg.warn(
f"The training data contains {sents_per_doc:.2f} sentences per "
f"document. When there are very few documents containing more "
f"than one sentence, the parser will not learn how to segment "
f"longer texts into sentences."
)
# profile labels
labels_train = [label for label in gold_train_data["deps"]]
labels_train_unpreprocessed = [
label for label in gold_train_unpreprocessed_data["deps"]
]
labels_dev = [label for label in gold_dev_data["deps"]]
if gold_train_unpreprocessed_data["n_nonproj"] > 0:
n_nonproj = gold_train_unpreprocessed_data["n_nonproj"]
msg.info(f"Found {n_nonproj} nonprojective train sentence(s)")
if gold_dev_data["n_nonproj"] > 0:
n_nonproj = gold_dev_data["n_nonproj"]
msg.info(f"Found {n_nonproj} nonprojective dev sentence(s)")
msg.info(f"{len(labels_train_unpreprocessed)} label(s) in train data")
msg.info(f"{len(labels_train)} label(s) in projectivized train data")
labels_with_counts = _format_labels(
gold_train_unpreprocessed_data["deps"].most_common(), counts=True
)
msg.text(labels_with_counts, show=verbose)
# rare labels in train
for label in gold_train_unpreprocessed_data["deps"]:
if gold_train_unpreprocessed_data["deps"][label] <= DEP_LABEL_THRESHOLD:
msg.warn(
f"Low number of examples for label '{label}' "
f"({gold_train_unpreprocessed_data['deps'][label]})"
)
has_low_data_warning = True
# rare labels in projectivized train
rare_projectivized_labels = []
for label in gold_train_data["deps"]:
if gold_train_data["deps"][label] <= DEP_LABEL_THRESHOLD and "||" in label:
rare_projectivized_labels.append(
f"{label}: {gold_train_data['deps'][label]}"
)
if len(rare_projectivized_labels) > 0:
msg.warn(
f"Low number of examples for {len(rare_projectivized_labels)} "
"label(s) in the projectivized dependency trees used for "
"training. You may want to projectivize labels such as punct "
"before training in order to improve parser performance."
)
msg.warn(
f"Projectivized labels with low numbers of examples: ",
", ".join(rare_projectivized_labels),
show=verbose,
)
has_low_data_warning = True
# labels only in train
if set(labels_train) - set(labels_dev):
msg.warn(
"The following labels were found only in the train data:",
", ".join(set(labels_train) - set(labels_dev)),
show=verbose,
)
# labels only in dev
if set(labels_dev) - set(labels_train):
msg.warn(
"The following labels were found only in the dev data:",
", ".join(set(labels_dev) - set(labels_train)),
show=verbose,
)
if has_low_data_warning:
msg.text(
f"To train a parser, your data should include at "
f"least {DEP_LABEL_THRESHOLD} instances of each label.",
show=verbose,
)
# multiple root labels
if len(gold_train_unpreprocessed_data["roots"]) > 1:
msg.warn(
f"Multiple root labels "
f"({', '.join(gold_train_unpreprocessed_data['roots'])}) "
f"found in training data. spaCy's parser uses a single root "
f"label ROOT so this distinction will not be available."
)
# these should not happen, but just in case
if gold_train_data["n_nonproj"] > 0:
msg.fail(
f"Found {gold_train_data['n_nonproj']} nonprojective "
f"projectivized train sentence(s)"
)
if gold_train_data["n_cycles"] > 0:
msg.fail(
f"Found {gold_train_data['n_cycles']} projectivized train sentence(s) with cycles"
)
msg.divider("Summary")
good_counts = msg.counts[MESSAGES.GOOD]
warn_counts = msg.counts[MESSAGES.WARN]
fail_counts = msg.counts[MESSAGES.FAIL]
if good_counts:
msg.good(f"{good_counts} {'check' if good_counts == 1 else 'checks'} passed")
if warn_counts:
msg.warn(f"{warn_counts} {'warning' if warn_counts == 1 else 'warnings'}")
if fail_counts:
msg.fail(f"{fail_counts} {'error' if fail_counts == 1 else 'errors'}")
sys.exit(1)
def _load_file(file_path: Path, msg: Printer) -> None:
file_name = file_path.parts[-1]
if file_path.suffix == ".json":
with msg.loading(f"Loading {file_name}..."):
data = srsly.read_json(file_path)
msg.good(f"Loaded {file_name}")
return data
elif file_path.suffix == ".jsonl":
with msg.loading(f"Loading {file_name}..."):
data = srsly.read_jsonl(file_path)
msg.good(f"Loaded {file_name}")
return data
msg.fail(
f"Can't load file extension {file_path.suffix}",
"Expected .json or .jsonl",
exits=1,
)
def _compile_gold(
examples: Sequence[Example],
factory_names: List[str],
nlp: Language,
make_proj: bool,
) -> Dict[str, Any]:
data = {
"ner": Counter(),
"cats": Counter(),
"tags": Counter(),
"deps": Counter(),
"words": Counter(),
"roots": Counter(),
"ws_ents": 0,
"punct_ents": 0,
"n_words": 0,
"n_misaligned_words": 0,
"words_missing_vectors": Counter(),
"n_sents": 0,
"n_nonproj": 0,
"n_cycles": 0,
"n_cats_multilabel": 0,
"texts": set(),
}
for eg in examples:
gold = eg.reference
doc = eg.predicted
valid_words = [x for x in gold if x is not None]
data["words"].update(valid_words)
data["n_words"] += len(valid_words)
data["n_misaligned_words"] += len(gold) - len(valid_words)
data["texts"].add(doc.text)
if len(nlp.vocab.vectors):
for word in valid_words:
if nlp.vocab.strings[word] not in nlp.vocab.vectors:
data["words_missing_vectors"].update([word])
if "ner" in factory_names:
for i, label in enumerate(eg.get_aligned_ner()):
if label is None:
continue
if label.startswith(("B-", "U-", "L-")) and doc[i].is_space:
# "Illegal" whitespace entity
data["ws_ents"] += 1
if label.startswith(("B-", "U-", "L-")) and doc[i].text in [
".",
"'",
"!",
"?",
",",
]:
# punctuation entity: could be replaced by whitespace when training with noise,
# so add a warning to alert the user to this unexpected side effect.
data["punct_ents"] += 1
if label.startswith(("B-", "U-")):
combined_label = label.split("-")[1]
data["ner"][combined_label] += 1
elif label == "-":
data["ner"]["-"] += 1
if "textcat" in factory_names:
data["cats"].update(gold.cats)
if list(gold.cats.values()).count(1.0) != 1:
data["n_cats_multilabel"] += 1
if "tagger" in factory_names:
tags = eg.get_aligned("TAG", as_string=True)
data["tags"].update([x for x in tags if x is not None])
if "parser" in factory_names:
aligned_heads, aligned_deps = eg.get_aligned_parse(projectivize=make_proj)
data["deps"].update([x for x in aligned_deps if x is not None])
for i, (dep, head) in enumerate(zip(aligned_deps, aligned_heads)):
if head == i:
data["roots"].update([dep])
data["n_sents"] += 1
if nonproj.is_nonproj_tree(aligned_heads):
data["n_nonproj"] += 1
if nonproj.contains_cycle(aligned_heads):
data["n_cycles"] += 1
return data
def _format_labels(labels: List[Tuple[str, int]], counts: bool = False) -> str:
if counts:
return ", ".join([f"'{l}' ({c})" for l, c in labels])
return ", ".join([f"'{l}'" for l in labels])
def _get_examples_without_label(data: Sequence[Example], label: str) -> int:
count = 0
for eg in data:
labels = [
label.split("-")[1]
for label in eg.get_aligned_ner()
if label not in ("O", "-", None)
]
if label not in labels:
count += 1
return count
def _get_labels_from_model(nlp: Language, pipe_name: str) -> Sequence[str]:
if pipe_name not in nlp.pipe_names:
return set()
pipe = nlp.get_pipe(pipe_name)
return pipe.labels