In the reference implementations, there's usually a function to build a
ffnn of arbitrary depth, consisting of a stack of Linear >> Relu >>
Dropout. In practice the depth is always 1 in coref-hoi, but in earlier
iterations of the model, which are more similar to our model here (since
we aren't using attention or even necessarily BERT), using a small depth
like 2 was common. This hard-codes a stack of 2.
In brief tests this allows similar performance to the unstacked version
with much smaller embedding sizes.
The depth of the stack could be made into a hyperparameter.
This generall means fewer spans are considered, which makes individual
steps in training faster but can make training take longer to find the
good spans.
Not necessary for convergence, but in coref-hoi this seems to add a few
f1 points.
Note that there are two width-related features in coref-hoi. This is a
"prior" that is added to mention scores. The other width related feature
is appended to the span embedding representation for other layers to
reference.
This rewrites the loss to not use the Thinc crossentropy code at all.
The main difference here is that the negative predictions are being
masked out (= marginalized over), but negative gradient is still being
reflected.
I'm still not sure this is exactly right but models seem to train
reliably now.
The calculation of this in the coref-hoi code is hard to follow. Based
on comments and variable names it sounds like it's using the doc length,
but it might actually be the number of mentions? Number of mentions
should be much larger and seems more correct, but might want to revisit
this.
I think this was technically incorrect but harmless. The reason the code
here is different than the reference in coref-hoi is that the indices
there are such that they get +1 at the end of processing, while the code
here handles indices directly.
* Draft spancat model
* Add spancat model
* Add test for extract_spans
* Add extract_spans layer
* Upd extract_spans
* Add spancat model
* Add test for spancat model
* Upd spancat model
* Update spancat component
* Upd spancat
* Update spancat model
* Add quick spancat test
* Import SpanCategorizer
* Fix SpanCategorizer component
* Import SpanGroup
* Fix span extraction
* Fix import
* Fix import
* Upd model
* Update spancat models
* Add scoring, update defaults
* Update and add docs
* Fix type
* Update spacy/ml/extract_spans.py
* Auto-format and fix import
* Fix comment
* Fix type
* Fix type
* Update website/docs/api/spancategorizer.md
* Fix comment
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Better defense
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Fix labels list
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Update spacy/ml/extract_spans.py
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Update spacy/pipeline/spancat.py
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Set annotations during update
* Set annotations in spancat
* fix imports in test
* Update spacy/pipeline/spancat.py
* replace MaxoutLogistic with LinearLogistic
* fix config
* various small fixes
* remove set_annotations parameter in update
* use our beloved tupley format with recent support for doc.spans
* bugfix to allow renaming the default span_key (scores weren't showing up)
* use different key in docs example
* change defaults to better-working parameters from project (WIP)
* register spacy.extract_spans.v1 for legacy purposes
* Upd dev version so can build wheel
* layers instead of architectures for smaller building blocks
* Update website/docs/api/spancategorizer.md
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Update website/docs/api/spancategorizer.md
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Include additional scores from overrides in combined score weights
* Parameterize spans key in scoring
Parameterize the `SpanCategorizer` `spans_key` for scoring purposes so
that it's possible to evaluate multiple `spancat` components in the same
pipeline.
* Use the (intentionally very short) default spans key `sc` in the
`SpanCategorizer`
* Adjust the default score weights to include the default key
* Adjust the scorer to use `spans_{spans_key}` as the prefix for the
returned score
* Revert addition of `attr_name` argument to `score_spans` and adjust
the key in the `getter` instead.
Note that for `spancat` components with a custom `span_key`, the score
weights currently need to be modified manually in
`[training.score_weights]` for them to be available during training. To
suppress the default score weights `spans_sc_p/r/f` during training, set
them to `null` in `[training.score_weights]`.
* Update website/docs/api/scorer.md
* Fix scorer for spans key containing underscore
* Increment version
* Add Spans to Evaluate CLI (#8439)
* Add Spans to Evaluate CLI
* Change to spans_key
* Add spans per_type output
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Fix spancat GPU issues (#8455)
* Fix GPU issues
* Require thinc >=8.0.6
* Switch to glorot_uniform_init
* Fix and test ngram suggester
* Include final ngram in doc for all sizes
* Fix ngrams for docs of the same length as ngram size
* Handle batches of docs that result in no ngrams
* Add tests
Co-authored-by: Ines Montani <ines@ines.io>
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
Co-authored-by: svlandeg <sofie.vanlandeghem@gmail.com>
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
Co-authored-by: Nirant <NirantK@users.noreply.github.com>
The call here was creating a float64 array, which was turning many
downstream scores into float64s. Later on these values were assigned to
a float32 array in backprop, and numerical underflow caused things to go
to zero.
That's almost certainly not the only reason things go to zero, but it is
incorrect.
* implement textcat resizing for TextCatCNN
* resizing textcat in-place
* simplify code
* ensure predictions for old textcat labels remain the same after resizing (WIP)
* fix for softmax
* store softmax as attr
* fix ensemble weight copy and cleanup
* restructure slightly
* adjust documentation, update tests and quickstart templates to use latest versions
* extend unit test slightly
* revert unnecessary edits
* fix typo
* ensemble architecture won't be resizable for now
* use resizable layer (WIP)
* revert using resizable layer
* resizable container while avoid shape inference trouble
* cleanup
* ensure model continues training after resizing
* use fill_b parameter
* use fill_defaults
* resize_layer callback
* format
* bump thinc to 8.0.4
* bump spacy-legacy to 3.0.6
At a few points in the code it's normal to get a "2d" array where each
row is a single entry. Calling squeeze will make that a proper 1d
array... unless it's just one entry, in which case it turns into a 0d
scalar. That's not what we want; flatten() provides the desired
behavior.
`make_clean_doc` is not needed and was removed.
`logsumexp` may be needed if I misunderstood the loss calculation, so I
left it in for now with a note.
