* Add callback to copy vocab/tokenizer from model
Add callback `spacy.copy_from_base_model.v1` to copy the tokenizer
settings and/or vocab (including vectors) from a base model.
* Move spacy.copy_from_base_model.v1 to spacy.training.callbacks
* Add documentation
* Modify to specify model as tokenizer and vocab params
* Update sent_starts in Example.from_dict
Update `sent_starts` for `Example.from_dict` so that `Optional[bool]`
values have the same meaning as for `Token.is_sent_start`.
Use `Optional[bool]` as the type for sent start values in the docs.
* Use helper function for conversion to ternary ints
* extend span scorer with consider_label and allow_overlap
* unit test for spans y2x overlap
* add score_spans unit test
* docs for new fields in scorer.score_spans
* rename to include_label
* spell out if-else for clarity
* rename to 'labeled'
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
Data in the JSON format is split into sentences, and each sentence is
saved with is_sent_start flags. Currently the flags are 1 for the first
token and 0 for the others. When deserialized this results in a pattern
of True, None, None, None... which makes single-sentence documents look
as though they haven't had sentence boundaries set.
Since items saved in JSON format have been split into sentences already,
the is_sent_start values should all be True or False.
* Support infinite generators for training corpora
Support a training corpus with an infinite generator in the `spacy
train` training loop:
* Revert `create_train_batches` to the state where an infinite generator
can be used as the in the first epoch of exactly one epoch without
resulting in a memory leak (`max_epochs != 1` will still result in a
memory leak)
* Move the shuffling for the first epoch into the corpus reader,
renaming it to `spacy.Corpus.v2`.
* Switch to training option for shuffling in memory
Training loop:
* Add option `training.shuffle_train_corpus_in_memory` that controls
whether the corpus is loaded in memory once and shuffled in the training
loop
* Revert changes to `create_train_batches` and rename to
`create_train_batches_with_shuffling` for use with `spacy.Corpus.v1` and
a corpus that should be loaded in memory
* Add `create_train_batches_without_shuffling` for a corpus that
should not be shuffled in the training loop: the corpus is merely
batched during training
Corpus readers:
* Restore `spacy.Corpus.v1`
* Add `spacy.ShuffledCorpus.v1` for a corpus shuffled in memory in the
reader instead of the training loop
* In combination with `shuffle_train_corpus_in_memory = False`, each
epoch could result in a different augmentation
* Refactor create_train_batches, validation
* Rename config setting to `training.shuffle_train_corpus`
* Refactor to use a single `create_train_batches` method with a
`shuffle` option
* Only validate `get_examples` in initialize step if:
* labels are required
* labels are not provided
* Switch back to max_epochs=-1 for streaming train corpus
* Use first 100 examples for stream train corpus init
* Always check validate_get_examples in initialize
See here:
https://github.com/explosion/spaCy/discussions/7463
Still need to check if there are any side effects of listeners being
present but not in the pipeline, but this commit will silence the
warnings.
* Fix aborted/skipped augmentation for `spacy.orth_variants.v1` if
lowercasing was enabled for an example
* Simplify `spacy.orth_variants.v1` for `Example` vs. `GoldParse`
* Preserve reference tokenization in `spacy.lower_case.v1`
* initialize NLP with train corpus
* add more pretraining tests
* more tests
* function to fetch tok2vec layer for pretraining
* clarify parameter name
* test different objectives
* formatting
* fix check for static vectors when using vectors objective
* clarify docs
* logger statement
* fix init_tok2vec and proc.initialize order
* test training after pretraining
* add init_config tests for pretraining
* pop pretraining block to avoid config validation errors
* custom errors
* Fix patience for identical scores
Fix training patience so that the earliest best step is chosen for
identical max scores.
* Restore break, remove print
* Explicitly define best_step for clarity
* Allow output_path to be None during training
* Fix cat scoring (?)
* Improve error message for weighted None score
* Improve messages
So we can call this in other places etc.
* FIx output path check
* Use latest wasabi
* Revert "Improve error message for weighted None score"
This reverts commit 7059926763.
* Exclude None scores from final score by default
It's otherwise very difficult to keep track of the score weights if we modify a config programmatically, source components etc.
* Update warnings and use logger.warning
* warn when frozen components break listener pattern
* few notes in the documentation
* update arg name
* formatting
* cleanup
* specify listeners return type
Validate both `[initialize]` and `[training]` in `debug data` and
`nlp.initialize()` with separate config validation error blocks that
indicate which block of the config is being validated.
* Fix memory issues in Language.evaluate
Reset annotation in predicted docs before evaluating and store all data
in `examples`.
* Minor refactor to docs generator init
* Fix generator expression
* Fix final generator check
* Refactor pipeline loop
* Handle examples generator in Language.evaluate
* Add test with generator
* Use make_doc
* Switch converters to generator functions
To reduce the memory usage when converting large corpora, refactor the
convert methods to be generator functions.
* Update tests
* Get basic beam tests working
* Get basic beam tests working
* Compile _beam_utils
* Remove prints
* Test beam density
* Beam parser seems to train
* Draft beam NER
* Upd beam
* Add hypothesis as dev dependency
* Implement missing is-gold-parse method
* Implement early update
* Fix state hashing
* Fix test
* Fix test
* Default to non-beam in parser constructor
* Improve oracle for beam
* Start refactoring beam
* Update test
* Refactor beam
* Update nn
* Refactor beam and weight by cost
* Update ner beam settings
* Update test
* Add __init__.pxd
* Upd test
* Fix test
* Upd test
* Fix test
* Remove ring buffer history from StateC
* WIP change arc-eager transitions
* Add state tests
* Support ternary sent start values
* Fix arc eager
* Fix NER
* Pass oracle cut size for beam
* Fix ner test
* Fix beam
* Improve StateC.clone
* Improve StateClass.borrow
* Work directly with StateC, not StateClass
* Remove print statements
* Fix state copy
* Improve state class
* Refactor parser oracles
* Fix arc eager oracle
* Fix arc eager oracle
* Use a vector to implement the stack
* Refactor state data structure
* Fix alignment of sent start
* Add get_aligned_sent_starts method
* Add test for ae oracle when bad sentence starts
* Fix sentence segment handling
* Avoid Reduce that inserts illegal sentence
* Update preset SBD test
* Fix test
* Remove prints
* Fix sent starts in Example
* Improve python API of StateClass
* Tweak comments and debug output of arc eager
* Upd test
* Fix state test
* Fix state test
* define new architectures for the pretraining objective
* add loss function as attr of the omdel
* cleanup
* cleanup
* shorten name
* fix typo
* remove unused error
* When checking for token alignments, check not only that the tokens are
identical but that the character positions are both at the start of a
token.
It's possible for the tokens to be identical even though the two
tokens aren't aligned one-to-one in a case like `["a'", "''"]` vs.
`["a", "''", "'"]`, where the middle tokens are identical but should not
be aligned on the token level at character position 2 since it's the
start of one token but the middle of another.
* Use the lowercased version of the token texts to create the
character-to-token alignment because lowercasing can change the string
length (e.g., for `İ`, see the not-a-bug bug report:
https://bugs.python.org/issue34723)
