* Tidy up train-from-config a bit
* Fix accidentally quadratic perf in TokenAnnotation.brackets
When we're reading in the gold data, we had a nested loop where
we looped over the brackets for each token, looking for brackets
that start on that word. This is accidentally quadratic, because
we have one bracket per word (for the POS tags). So we had
an O(N**2) behaviour here that ended up being pretty slow.
To solve this I'm indexing the brackets by their starting word
on the TokenAnnotations object, and having a property to provide
the previous view.
* Fixes
* setting KB in the EL constructor, similar to how the model is passed on
* removing wikipedia example files - moved to projects
* throw an error when nlp.update is called with 2 positional arguments
* rewriting the config logic in create pipe to accomodate for other objects (e.g. KB) in the config
* update config files with new parameters
* avoid training pipeline components that don't have a model (like sentencizer)
* various small fixes + UX improvements
* small fixes
* set thinc to 8.0.0a9 everywhere
* remove outdated comment
* make disable_pipes deprecated in favour of the new toggle_pipes
* rewrite disable_pipes statements
* update documentation
* remove bin/wiki_entity_linking folder
* one more fix
* remove deprecated link to documentation
* few more doc fixes
* add note about name change to the docs
* restore original disable_pipes
* small fixes
* fix typo
* fix error number to W096
* rename to select_pipes
* also make changes to the documentation
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* Draft layer for BILUO actions
* Fixes to biluo layer
* WIP on BILUO layer
* Add tests for BILUO layer
* Format
* Fix transitions
* Update test
* Link in the simple_ner
* Update BILUO tagger
* Update __init__
* Import simple_ner
* Update test
* Import
* Add files
* Add config
* Fix label passing for BILUO and tagger
* Fix label handling for simple_ner component
* Update simple NER test
* Update config
* Hack train script
* Update BILUO layer
* Fix SimpleNER component
* Update train_from_config
* Add biluo_to_iob helper
* Add IOB layer
* Add IOBTagger model
* Update biluo layer
* Update SimpleNER tagger
* Update BILUO
* Read random seed in train-from-config
* Update use of normal_init
* Fix normalization of gradient in SimpleNER
* Update IOBTagger
* Remove print
* Tweak masking in BILUO
* Add dropout in SimpleNER
* Update thinc
* Tidy up simple_ner
* Fix biluo model
* Unhack train-from-config
* Update setup.cfg and requirements
* Add tb_framework.py for parser model
* Try to avoid memory leak in BILUO
* Move ParserModel into spacy.ml, avoid need for subclass.
* Use updated parser model
* Remove incorrect call to model.initializre in PrecomputableAffine
* Update parser model
* Avoid divide by zero in tagger
* Add extra dropout layer in tagger
* Refine minibatch_by_words function to avoid oom
* Fix parser model after refactor
* Try to avoid div-by-zero in SimpleNER
* Fix infinite loop in minibatch_by_words
* Use SequenceCategoricalCrossentropy in Tagger
* Fix parser model when hidden layer
* Remove extra dropout from tagger
* Add extra nan check in tagger
* Fix thinc version
* Update tests and imports
* Fix test
* Update test
* Update tests
* Fix tests
* Fix test
Co-authored-by: Ines Montani <ines@ines.io>
Previously, pipelines with shared tok2vec weights would call the
tok2vec backprop callback multiple times, once for each pipeline
component. This caused errors for PyTorch, and was inefficient.
Instead, accumulate the gradient for all but one component, and just
call the callback once.
* Add pos and morph scoring to Scorer
Add pos, morph, and morph_per_type to `Scorer`. Report pos and morph
accuracy in `spacy evaluate`.
* Update morphologizer for v3
* switch to tagger-based morphologizer
* use `spacy.HashCharEmbedCNN` for morphologizer defaults
* add `Doc.is_morphed` flag
* Add morphologizer to train CLI
* Add basic morphologizer pipeline tests
* Add simple morphologizer training example
* Remove subword_features from CharEmbed models
Remove `subword_features` argument from `spacy.HashCharEmbedCNN.v1` and
`spacy.HashCharEmbedBiLSTM.v1` since in these cases `subword_features`
is always `False`.
* Rename setting in morphologizer example
Use `with_pos_tags` instead of `without_pos_tags`.
* Fix kwargs for spacy.HashCharEmbedBiLSTM.v1
* Remove defaults for spacy.HashCharEmbedBiLSTM.v1
Remove default `nM/nC` for `spacy.HashCharEmbedBiLSTM.v1`.
* Set random seed for textcat overfitting test
* bring back default build_text_classifier method
* remove _set_dims_ hack in favor of proper dim inference
* add tok2vec initialize to unit test
* small fixes
* add unit test for various textcat config settings
* logistic output layer does not have nO
* fix window_size setting
* proper fix
* fix W initialization
* Update textcat training example
* Use ml_datasets
* Convert training data to `Example` format
* Use `n_texts` to set proportionate dev size
* fix _init renaming on latest thinc
* avoid setting a non-existing dim
* update to thinc==8.0.0a2
* add BOW and CNN defaults for easy testing
* various experiments with train_textcat script, fix softmax activation in textcat bow
* allow textcat train script to work on other datasets as well
* have dataset as a parameter
* train textcat from config, with example config
* add config for training textcat
* formatting
* fix exclusive_classes
* fixing BOW for GPU
* bump thinc to 8.0.0a3 (not published yet so CI will fail)
* add in link_vectors_to_models which got deleted
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Check whether doc is instantiated
When creating docs to pair with gold parses, modify test to check
whether a doc is unset rather than whether it contains tokens.
* Restore test of evaluate on an empty doc
* Set a minimal gold.orig for the scorer
Without a minimal gold.orig the scorer can't evaluate empty docs. This
is the v3 equivalent of #4925.
* 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>
UD_Danish-DDT has (as far as I can tell) hallucinated periods after
abbreviations, so the changes are an artifact of the corpus and not due
to anything meaningful about Danish tokenization.
