Remove the non-working `--use-chars` option from the train CLI. The
implementation of the option across component types and the CLI settings
could be fixed, but the `CharacterEmbed` model does not work on GPU in
v2 so it's better to remove it.
* Improve tag map initialization and updating
Generalize tag map initialization and updating so that a provided tag
map can be loaded correctly in the CLI.
* normalize provided tag map as necessary
* use the same method for initializing and overwriting the tag map
* Reinitialize cache after loading new tag map
Reinitialize the cache with the right size after loading a new tag map.
* Use cosine loss in Cloze multitask
* Fix char_embed for gpu
* Call resume_training for base model in train CLI
* Fix bilstm_depth default in pretrain command
* Implement character-based pretraining objective
* Use chars loss in ClozeMultitask
* Add method to decode predicted characters
* Fix number characters
* Rescale gradients for mlm
* Fix char embed+vectors in ml
* Fix pipes
* Fix pretrain args
* Move get_characters_loss
* Fix import
* Fix import
* Mention characters loss option in pretrain
* Remove broken 'self attention' option in pretrain
* Revert "Remove broken 'self attention' option in pretrain"
This reverts commit 56b820f6af.
* Document 'characters' objective of pretrain
* Reduce stored lexemes data, move feats to lookups
* Move non-derivable lexemes features (`norm / cluster / prob`) to
`spacy-lookups-data` as lookups
* Get/set `norm` in both lookups and `LexemeC`, serialize in lookups
* Remove `cluster` and `prob` from `LexemesC`, get/set/serialize in
lookups only
* Remove serialization of lexemes data as `vocab/lexemes.bin`
* Remove `SerializedLexemeC`
* Remove `Lexeme.to_bytes/from_bytes`
* Modify normalization exception loading:
* Always create `Vocab.lookups` table `lexeme_norm` for
normalization exceptions
* Load base exceptions from `lang.norm_exceptions`, but load
language-specific exceptions from lookups
* Set `lex_attr_getter[NORM]` including new lookups table in
`BaseDefaults.create_vocab()` and when deserializing `Vocab`
* Remove all cached lexemes when deserializing vocab to override
existing normalizations with the new normalizations (as a replacement
for the previous step that replaced all lexemes data with the
deserialized data)
* Skip English normalization test
Skip English normalization test because the data is now in
`spacy-lookups-data`.
* Remove norm exceptions
Moved to spacy-lookups-data.
* Move norm exceptions test to spacy-lookups-data
* Load extra lookups from spacy-lookups-data lazily
Load extra lookups (currently for cluster and prob) lazily from the
entry point `lg_extra` as `Vocab.lookups_extra`.
* Skip creating lexeme cache on load
To improve model loading times, do not create the full lexeme cache when
loading. The lexemes will be created on demand when processing.
* Identify numeric values in Lexeme.set_attrs()
With the removal of a special case for `PROB`, also identify `float` to
avoid trying to convert it with the `StringStore`.
* Skip lexeme cache init in from_bytes
* Unskip and update lookups tests for python3.6+
* Update vocab pickle to include lookups_extra
* Update vocab serialization tests
Check strings rather than lexemes since lexemes aren't initialized
automatically, account for addition of "_SP".
* Re-skip lookups test because of python3.5
* Skip PROB/float values in Lexeme.set_attrs
* Convert is_oov from lexeme flag to lex in vectors
Instead of storing `is_oov` as a lexeme flag, `is_oov` reports whether
the lexeme has a vector.
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* `debug-data`: determine coverage of provided vectors
* `evaluate`: support `blank:lg` model to make it possible to just evaluate
tokenization
* `init-model`: add option to truncate vectors to N most frequent vectors
from word2vec file
* `train`:
* if training on GPU, only run evaluation/timing on CPU in the first
iteration
* if training is aborted, exit with a non-0 exit status
* Fixed typo in cli warning
Fixed a typo in the warning for the provision of exactly two labels, which have not been designated as binary, to textcat.
* Create and signed contributor form
* Omit per_type scores from model-best calculations
The addition of per_type scores to the included metrics (#4911) causes
errors when they're compared while determining the best model, so omit
them for this `max()` comparison.
* Add default speed data for interrupted train CLI
Add better speed meta defaults so that an interrupted iteration still
produces a best model.
Co-authored-by: Ines Montani <ines@ines.io>
* Fix model-final/model-best meta
* include speed and accuracy from final iteration
* combine with speeds from base model if necessary
* Include token_acc metric for all components
Improve train CLI with a provided base model so that you can:
* add a new component
* extend an existing component
* replace an existing component
When the final model and best model are saved, reenable any disabled
components and merge the meta information to include the full pipeline
and accuracy information for all components in the base model plus the
newly added components if needed.
* Flag to ignore examples with mismatched raw/gold text
After #4525, we're seeing some alignment failures on our OntoNotes data. I think we actually have fixes for most of these cases.
In general it's better to fix the data, but it seems good to allow the GoldCorpus class to just skip cases where the raw text doesn't
match up to the gold words. I think previously we were silently ignoring these cases.
* Try to fix test on Python 2.7
* Add doc.cats to spacy.gold at the paragraph level
Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in
the spacy JSON training format at the paragraph level.
* `spacy.gold.docs_to_json()` writes `docs.cats`
* `GoldCorpus` reads in cats in each `GoldParse`
* Update instances of gold_tuples to handle cats
Update iteration over gold_tuples / gold_parses to handle addition of
cats at the paragraph level.
* Add textcat to train CLI
* Add textcat options to train CLI
* Add textcat labels in `TextCategorizer.begin_training()`
* Add textcat evaluation to `Scorer`:
* For binary exclusive classes with provided label: F1 for label
* For 2+ exclusive classes: F1 macro average
* For multilabel (not exclusive): ROC AUC macro average (currently
relying on sklearn)
* Provide user info on textcat evaluation settings, potential
incompatibilities
* Provide pipeline to Scorer in `Language.evaluate` for textcat config
* Customize train CLI output to include only metrics relevant to current
pipeline
* Add textcat evaluation to evaluate CLI
* Fix handling of unset arguments and config params
Fix handling of unset arguments and model confiug parameters in Scorer
initialization.
* Temporarily add sklearn requirement
* Remove sklearn version number
* Improve Scorer handling of models without textcats
* Fixing Scorer handling of models without textcats
* Update Scorer output for python 2.7
* Modify inf in Scorer for python 2.7
* Auto-format
Also make small adjustments to make auto-formatting with black easier and produce nicer results
* Move error message to Errors
* Update documentation
* Add cats to annotation JSON format [ci skip]
* Fix tpl flag and docs [ci skip]
* Switch to internal roc_auc_score
Switch to internal `roc_auc_score()` adapted from scikit-learn.
* Add AUCROCScore tests and improve errors/warnings
* Add tests for AUCROCScore and roc_auc_score
* Add missing error for only positive/negative values
* Remove unnecessary warnings and errors
* Make reduced roc_auc_score functions private
Because most of the checks and warnings have been stripped for the
internal functions and access is only intended through `ROCAUCScore`,
make the functions for roc_auc_score adapted from scikit-learn private.
* Check that data corresponds with multilabel flag
Check that the training instances correspond with the multilabel flag,
adding the multilabel flag if required.
* Add textcat score to early stopping check
* Add more checks to debug-data for textcat
* Add example training data for textcat
* Add more checks to textcat train CLI
* Check configuration when extending base model
* Fix typos
* Update textcat example data
* Provide licensing details and licenses for data
* Remove two labels with no positive instances from jigsaw-toxic-comment
data.
Co-authored-by: Ines Montani <ines@ines.io>
* Prevent subtok label if not learning tokens
The parser introduces the subtok label to mark tokens that should be
merged during post-processing. Previously this happened even if we did
not have the --learn-tokens flag set. This patch passes the config
through to the parser, to prevent the problem.
* Make merge_subtokens a parser post-process if learn_subtokens
* Fix train script
* Add test for 3830: subtok problem
* Fix handlign of non-subtok in parser training
* Add early stopping
* Add return_score option to evaluate
* Fix missing str to path conversion
* Fix import + old python compatibility
* Fix bad beam_width setting during cpu evaluation in spacy train with gpu option turned on
