* Avoid `TrainablePipe.finish_update` getting called twice during training
PR #12136 fixed an issue where the tok2vec pipe was updated before
gradient were accumulated. However, it introduced a new bug that cause
`finish_update` to be called twice when using the training loop. This
causes a fairly large slowdown.
The `Language.update` method accepts the `sgd` argument for passing an
optimizer. This argument has three possible values:
- `Optimizer`: use the given optimizer to finish pipe updates.
- `None`: use a default optimizer to finish pipe updates.
- `False`: do not finish pipe updates.
However, the latter option was not documented and not valid with the
existing type of `sgd`. I assumed that this was a remnant of earlier
spaCy versions and removed handling of `False`.
However, with that change, we are passing `None` to `Language.update`.
As a result, we were calling `finish_update` in both `Language.update`
and in the training loop after all subbatches are processed.
This change restores proper handling/use of `False`. Moreover, the role
of `False` is now documented and added to the type to avoid future
accidents.
* Fix typo
* Document defaults for `Language.update`
* `Language.update`: ensure that tok2vec gets updated
The components in a pipeline can be updated independently. However,
tok2vec implementations are an exception to this, since they depend on
listeners for their gradients. The update method of a tok2vec
implementation computes the tok2vec forward and passes this along with a
backprop function to the listeners. This backprop function accumulates
gradients for all the listeners. There are two ways in which the
accumulated gradients can be used to update the tok2vec weights:
1. Call the `finish_update` method of tok2vec *after* the `update`
method is called on all of the pipes that use a tok2vec listener.
2. Pass an optimizer to the `update` method of tok2vec. In this
case, tok2vec will give the last listener a special backprop
function that calls `finish_update` on the tok2vec.
Unfortunately, `Language.update` did neither of these. Instead, it
immediately called `finish_update` on every pipe after `update`. As a
result, the tok2vec weights are updated when no gradients have been
accumulated from listeners yet. And the gradients of the listeners are
only used in the next call to `Language.update` (when `finish_update` is
called on tok2vec again).
This change fixes this issue by passing the optimizer to the `update`
method of trainable pipes, leading to use of the second strategy
outlined above.
The main updating loop in `Language.update` is also simplified by using
the `TrainableComponent` protocol consistently.
* Train loop: `sgd` is `Optional[Optimizer]`, do not pass false
* Language.update: call pipe finish_update after all pipe updates
This does correct and fast updates if multiple components update the
same parameters.
* Add comment why we moved `finish_update` to a separate loop
* change logging call for spacy.LookupsDataLoader.v1
* substitutions in language and _util
* various more substitutions
* add string formatting guidelines to contribution guidelines
* Init
* fix tests
* Update spacy/errors.py
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Fix test_blank_languages
* Rename xx to mul in docs
* Format _util with black
* prettier formatting
---------
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Add `Language.distill`
This method is the distillation counterpart of `Language.update`. It
takes a teacher `Language` instance and distills the student pipes on
the teacher pipes.
* Apply suggestions from code review
Co-authored-by: Madeesh Kannan <shadeMe@users.noreply.github.com>
* Clarify that how Example is used in distillation
* Update transition parser distill docstring for examples argument
* Pass optimizer to `TrainablePipe.distill`
* Annotate pipe before update
As discussed internally, we want to let a pipe annotate before doing an
update with gold/silver data. Otherwise, the output may be (too)
informed by the gold/silver data.
* Rename `component_map` to `student_to_teacher`
* Better synopsis in `Language.distill` docstring
* `name` -> `student_name`
* Fix labels type in docstring
* Mark distill test as slow
* Fix `student_to_teacher` type in docs
---------
Co-authored-by: Madeesh Kannan <shadeMe@users.noreply.github.com>
* add test for running evaluate on an nlp pipeline with two distinct textcat components
* cleanup
* merge dicts instead of overwrite
* don't add more labels to the given set
* Revert "merge dicts instead of overwrite"
This reverts commit 89bee0ed77.
* Switch tests to separate scorer keys rather than merged dicts
* Revert unrelated edits
* Switch textcat scorers to v2
* formatting
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* make nlp.pipe() return None docs when no exceptions are (re-)raised during error handling
* Remove changes other than as_tuples test
* Only check warning count for one process
* Fix types
* Format
Co-authored-by: Xi Bai <xi.bai.ed@gmail.com>
* Raise an error when multiprocessing is used on a GPU
As reported in #5507, a confusing exception is thrown when
multiprocessing is used with a GPU model and the `fork` multiprocessing
start method:
cupy.cuda.runtime.CUDARuntimeError: cudaErrorInitializationError: initialization error
This change checks whether one of the models uses the GPU when
multiprocessing is used. If so, raise a friendly error message.
Even though multiprocessing can work on a GPU with the `spawn` method,
it quickly runs the GPU out-of-memory on real-world data. Also,
multiprocessing on a single GPU typically does not provide large
performance gains.
* Move GPU multiprocessing check to Language.pipe
* Warn rather than error when using multiprocessing with GPU models
* Improve GPU multiprocessing warning message.
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Reduce API assumptions
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Update spacy/language.py
* Update spacy/language.py
* Test that warning is thrown with GPU + multiprocessing
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* use language-matching to allow language code aliases
Signed-off-by: Elia Robyn Speer <elia@explosion.ai>
* link to "IETF language tags" in docs
Signed-off-by: Elia Robyn Speer <elia@explosion.ai>
* Make requirements consistent
Signed-off-by: Elia Robyn Speer <elia@explosion.ai>
* change "two-letter language ID" to "IETF language tag" in language docs
Signed-off-by: Elia Robyn Speer <elia@explosion.ai>
* use langcodes 3.2 and handle language-tag errors better
Signed-off-by: Elia Robyn Speer <elia@explosion.ai>
* all unknown language codes are ImportErrors
Signed-off-by: Elia Robyn Speer <elia@explosion.ai>
Co-authored-by: Elia Robyn Speer <elia@explosion.ai>
* Accept Doc input in pipelines
Allow `Doc` input to `Language.__call__` and `Language.pipe`, which
skips `Language.make_doc` and passes the doc directly to the pipeline.
* ensure_doc helper function
* avoid running multiple processes on GPU
* Update spacy/tests/test_language.py
Co-authored-by: svlandeg <svlandeg@github.com>
* Don't use the same vocab for source models
The source models should not be loaded with the vocab from the current
pipeline because this loads the vectors from the source model into the
current vocab.
The strings are all copied in `Language.create_pipe_from_source`, so if
the vectors are configured correctly in the current pipeline, the
sourced component will work as expected. If there is a vector mismatch,
a warning is shown. (It's not possible to inspect whether the vectors
are actually used by the component, so a warning is the best option.)
* Update comment on source model loading
* Handle errors while multiprocessing
Handle errors while multiprocessing without hanging.
* Return the traceback for errors raised while processing a batch, which
can be handled by the top-level error handler
* Allow for shortened batches due to custom error handlers that ignore
errors and skip documents
* Define custom components at a higher level
* Also move up custom error handler
* Use simpler component for test
* Switch error type
* Adjust test
* Only call top-level error handler for exceptions
* Register custom test components within tests
Use global functions (so they can be pickled) but register the
components only within the individual tests.
* Set up CI for tests with GPU agent
* Update tests for enabled GPU
* Fix steps filename
* Add parallel build jobs as a setting
* Fix test requirements
* Fix install test requirements condition
* Fix pipeline models test
* Reset current ops in prefer/require testing
* Fix more tests
* Remove separate test_models test
* Fix regression 5551
* fix StaticVectors for GPU use
* fix vocab tests
* Fix regression test 5082
* Move azure steps to .github and reenable default pool jobs
* Consolidate/rename azure steps
Co-authored-by: svlandeg <sofie.vanlandeghem@gmail.com>
* add error handler for pipe methods
* add unit tests
* remove pipe method that are the same as their base class
* have Language keep track of a default error handler
* cleanup
* formatting
* small refactor
* add documentation
Add `initialize.before_init` and `initialize.after_init` callbacks to
the config. The `initialize.before_init` callback is a place to
implement one-time tokenizer customizations that are then saved with the
model.
* 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
* WIP: Concept for modifying nlp object before and after init
* Make callbacks return nlp object
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* Raise if callbacks don't return correct type
* Rename, update types, add after_pipeline_creation
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* Update with WIP
* Update with WIP
* Update with pipeline serialization
* Update types and pipe factories
* Add deep merge, tidy up and add tests
* Fix pipe creation from config
* Don't validate default configs on load
* Update spacy/language.py
Co-authored-by: Ines Montani <ines@ines.io>
* Adjust factory/component meta error
* Clean up factory args and remove defaults
* Add test for failing empty dict defaults
* Update pipeline handling and methods
* provide KB as registry function instead of as object
* small change in test to make functionality more clear
* update example script for EL configuration
* Fix typo
* Simplify test
* Simplify test
* splitting pipes.pyx into separate files
* moving default configs to each component file
* fix batch_size type
* removing default values from component constructors where possible (TODO: test 4725)
* skip instead of xfail
* Add test for config -> nlp with multiple instances
* pipeline.pipes -> pipeline.pipe
* Tidy up, document, remove kwargs
* small cleanup/generalization for Tok2VecListener
* use DEFAULT_UPSTREAM field
* revert to avoid circular imports
* Fix tests
* Replace deprecated arg
* Make model dirs require config
* fix pickling of keyword-only arguments in constructor
* WIP: clean up and integrate full config
* Add helper to handle function args more reliably
Now also includes keyword-only args
* Fix config composition and serialization
* Improve config debugging and add visual diff
* Remove unused defaults and fix type
* Remove pipeline and factories from meta
* Update spacy/default_config.cfg
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Update spacy/default_config.cfg
* small UX edits
* avoid printing stack trace for debug CLI commands
* Add support for language-specific factories
* specify the section of the config which holds the model to debug
* WIP: add Language.from_config
* Update with language data refactor WIP
* Auto-format
* Add backwards-compat handling for Language.factories
* Update morphologizer.pyx
* Fix morphologizer
* Update and simplify lemmatizers
* Fix Japanese tests
* Port over tagger changes
* Fix Chinese and tests
* Update to latest Thinc
* WIP: xfail first Russian lemmatizer test
* Fix component-specific overrides
* fix nO for output layers in debug_model
* Fix default value
* Fix tests and don't pass objects in config
* Fix deep merging
* Fix lemma lookup data registry
Only load the lookups if an entry is available in the registry (and if spacy-lookups-data is installed)
* Add types
* Add Vocab.from_config
* Fix typo
* Fix tests
* Make config copying more elegant
* Fix pipe analysis
* Fix lemmatizers and is_base_form
* WIP: move language defaults to config
* Fix morphology type
* Fix vocab
* Remove comment
* Update to latest Thinc
* Add morph rules to config
* Tidy up
* Remove set_morphology option from tagger factory
* Hack use_gpu
* Move [pipeline] to top-level block and make [nlp.pipeline] list
Allows separating component blocks from component order – otherwise, ordering the config would mean a changed component order, which is bad. Also allows initial config to define more components and not use all of them
* Fix use_gpu and resume in CLI
* Auto-format
* Remove resume from config
* Fix formatting and error
* [pipeline] -> [components]
* Fix types
* Fix tagger test: requires set_morphology?
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
Co-authored-by: svlandeg <sofie.vanlandeghem@gmail.com>
Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
* remove _convert_examples
* fix test_gold, raise TypeError if tuples are used instead of Example's
* throwing proper errors when the wrong type of objects are passed
* fix deprectated format in tests
* fix deprectated format in parser tests
* fix tests for NEL, morph, senter, tagger, textcat
* update regression tests with new Example format
* use make_doc
* more fixes to nlp.update calls
* few more small fixes for rehearse and evaluate
* only import ml_datasets if really necessary
* Update errors
* Remove beam for now (maybe)
Remove beam_utils
Update setup.py
Remove beam
* Remove GoldParse
WIP on removing goldparse
Get ArcEager compiling after GoldParse excise
Update setup.py
Get spacy.syntax compiling after removing GoldParse
Rename NewExample -> Example and clean up
Clean html files
Start updating tests
Update Morphologizer
* fix error numbers
* fix merge conflict
* informative error when calling to_array with wrong field
* fix error catching
* fixing language and scoring tests
* start testing get_aligned
* additional tests for new get_aligned function
* Draft create_gold_state for arc_eager oracle
* Fix import
* Fix import
* Remove TokenAnnotation code from nonproj
* fixing NER one-to-many alignment
* Fix many-to-one IOB codes
* fix test for misaligned
* attempt to fix cases with weird spaces
* fix spaces
* test_gold_biluo_different_tokenization works
* allow None as BILUO annotation
* fixed some tests + WIP roundtrip unit test
* add spaces to json output format
* minibatch utiltiy can deal with strings, docs or examples
* fix augment (needs further testing)
* various fixes in scripts - needs to be further tested
* fix test_cli
* cleanup
* correct silly typo
* add support for MORPH in to/from_array, fix morphologizer overfitting test
* fix tagger
* fix entity linker
* ensure test keeps working with non-linked entities
* pipe() takes docs, not examples
* small bug fix
* textcat bugfix
* throw informative error when running the components with the wrong type of objects
* fix parser tests to work with example (most still failing)
* fix BiluoPushDown parsing entities
* small fixes
* bugfix tok2vec
* fix renames and simple_ner labels
* various small fixes
* prevent writing dummy values like deps because that could interfer with sent_start values
* fix the fix
* implement split_sent with aligned SENT_START attribute
* test for split sentences with various alignment issues, works
* Return ArcEagerGoldParse from ArcEager
* Update parser and NER gold stuff
* Draft new GoldCorpus class
* add links to to_dict
* clean up
* fix test checking for variants
* Fix oracles
* Start updating converters
* Move converters under spacy.gold
* Move things around
* Fix naming
* Fix name
* Update converter to produce DocBin
* Update converters
* Allow DocBin to take list of Doc objects.
* Make spacy convert output docbin
* Fix import
* Fix docbin
* Fix compile in ArcEager
* Fix import
* Serialize all attrs by default
* Update converter
* Remove jsonl converter
* Add json2docs converter
* Draft Corpus class for DocBin
* Work on train script
* Update Corpus
* Update DocBin
* Allocate Doc before starting to add words
* Make doc.from_array several times faster
* Update train.py
* Fix Corpus
* Fix parser model
* Start debugging arc_eager oracle
* Update header
* Fix parser declaration
* Xfail some tests
* Skip tests that cause crashes
* Skip test causing segfault
* Remove GoldCorpus
* Update imports
* Update after removing GoldCorpus
* Fix module name of corpus
* Fix mimport
* Work on parser oracle
* Update arc_eager oracle
* Restore ArcEager.get_cost function
* Update transition system
* Update test_arc_eager_oracle
* Remove beam test
* Update test
* Unskip
* Unskip tests
* add links to to_dict
* clean up
* fix test checking for variants
* Allow DocBin to take list of Doc objects.
* Fix compile in ArcEager
* Serialize all attrs by default
Move converters under spacy.gold
Move things around
Fix naming
Fix name
Update converter to produce DocBin
Update converters
Make spacy convert output docbin
Fix import
Fix docbin
Fix import
Update converter
Remove jsonl converter
Add json2docs converter
* Allocate Doc before starting to add words
* Make doc.from_array several times faster
* Start updating converters
* Work on train script
* Draft Corpus class for DocBin
Update Corpus
Fix Corpus
* Update DocBin
Add missing strings when serializing
* Update train.py
* Fix parser model
* Start debugging arc_eager oracle
* Update header
* Fix parser declaration
* Xfail some tests
Skip tests that cause crashes
Skip test causing segfault
* Remove GoldCorpus
Update imports
Update after removing GoldCorpus
Fix module name of corpus
Fix mimport
* Work on parser oracle
Update arc_eager oracle
Restore ArcEager.get_cost function
Update transition system
* Update tests
Remove beam test
Update test
Unskip
Unskip tests
* Add get_aligned_parse method in Example
Fix Example.get_aligned_parse
* Add kwargs to Corpus.dev_dataset to match train_dataset
* Update nonproj
* Use get_aligned_parse in ArcEager
* Add another arc-eager oracle test
* Remove Example.doc property
Remove Example.doc
Remove Example.doc
Remove Example.doc
Remove Example.doc
* Update ArcEager oracle
Fix Break oracle
* Debugging
* Fix Corpus
* Fix eg.doc
* Format
* small fixes
* limit arg for Corpus
* fix test_roundtrip_docs_to_docbin
* fix test_make_orth_variants
* fix add_label test
* Update tests
* avoid writing temp dir in json2docs, fixing 4402 test
* Update test
* Add missing costs to NER oracle
* Update test
* Work on Example.get_aligned_ner method
* Clean up debugging
* Xfail tests
* Remove prints
* Remove print
* Xfail some tests
* Replace unseen labels for parser
* Update test
* Update test
* Xfail test
* Fix Corpus
* fix imports
* fix docs_to_json
* various small fixes
* cleanup
* Support gold_preproc in Corpus
* Support gold_preproc
* Pass gold_preproc setting into corpus
* Remove debugging
* Fix gold_preproc
* Fix json2docs converter
* Fix convert command
* Fix flake8
* Fix import
* fix output_dir (converted to Path by typer)
* fix var
* bugfix: update states after creating golds to avoid out of bounds indexing
* Improve efficiency of ArEager oracle
* pull merge_sent into iob2docs to avoid Doc creation for each line
* fix asserts
* bugfix excl Span.end in iob2docs
* Support max_length in Corpus
* Fix arc_eager oracle
* Filter out uannotated sentences in NER
* Remove debugging in parser
* Simplify NER alignment
* Fix conversion of NER data
* Fix NER init_gold_batch
* Tweak efficiency of precomputable affine
* Update onto-json default
* Update gold test for NER
* Fix parser test
* Update test
* Add NER data test
* Fix convert for single file
* Fix test
* Hack scorer to avoid evaluating non-nered data
* Fix handling of NER data in Example
* Output unlabelled spans from O biluo tags in iob_utils
* Fix unset variable
* Return kept examples from init_gold_batch
* Return examples from init_gold_batch
* Dont return Example from init_gold_batch
* Set spaces on gold doc after conversion
* Add test
* Fix spaces reading
* Improve NER alignment
* Improve handling of missing values in NER
* Restore the 'cutting' in parser training
* Add assertion
* Print epochs
* Restore random cuts in parser/ner training
* Implement Doc.copy
* Implement Example.copy
* Copy examples at the start of Language.update
* Don't unset example docs
* Tweak parser model slightly
* attempt to fix _guess_spaces
* _add_entities_to_doc first, so that links don't get overwritten
* fixing get_aligned_ner for one-to-many
* fix indexing into x_text
* small fix biluo_tags_from_offsets
* Add onto-ner config
* Simplify NER alignment
* Fix NER scoring for partially annotated documents
* fix indexing into x_text
* fix test_cli failing tests by ignoring spans in doc.ents with empty label
* Fix limit
* Improve NER alignment
* Fix count_train
* Remove print statement
* fix tests, we're not having nothing but None
* fix clumsy fingers
* Fix tests
* Fix doc.ents
* Remove empty docs in Corpus and improve limit
* Update config
Co-authored-by: svlandeg <sofie.vanlandeghem@gmail.com>
* OrigAnnot class instead of gold.orig_annot list of zipped tuples
* from_orig to replace from_annot_tuples
* rename to RawAnnot
* some unit tests for GoldParse creation and internal format
* removing orig_annot and switching to lists instead of tuple
* rewriting tuples to use RawAnnot (+ debug statements, WIP)
* fix pop() changing the data
* small fixes
* pop-append fixes
* return RawAnnot for existing GoldParse to have uniform interface
* clean up imports
* fix merge_sents
* add unit test for 4402 with new structure (not working yet)
* introduce DocAnnot
* typo fixes
* add unit test for merge_sents
* rename from_orig to from_raw
* fixing unit tests
* fix nn parser
* read_annots to produce text, doc_annot pairs
* _make_golds fix
* rename golds_to_gold_annots
* small fixes
* fix encoding
* have golds_to_gold_annots use DocAnnot
* missed a spot
* merge_sents as function in DocAnnot
* allow specifying only part of the token-level annotations
* refactor with Example class + underlying dicts
* pipeline components to work with Example objects (wip)
* input checking
* fix yielding
* fix calls to update
* small fixes
* fix scorer unit test with new format
* fix kwargs order
* fixes for ud and conllu scripts
* fix reading data for conllu script
* add in proper errors (not fixed numbering yet to avoid merge conflicts)
* fixing few more small bugs
* fix EL script
