SpaCy's HashEmbedCNN layer performs convolutions over tokens to produce
contextualized embeddings using a `MaxoutWindowEncoder` layer. These
convolutions are implemented using Thinc's `expand_window` layer, which
concatenates `window_size` neighboring sequence items on either side of
the sequence item being processed. This is repeated across `depth`
convolutional layers.
For example, consider the sequence "ABCDE" and a `MaxoutWindowEncoder`
layer with a context window of 1 and a depth of 2. We'll focus on the
token "C". We can visually represent the contextual embedding produced
for "C" as:
```mermaid
flowchart LR
A0(A<sub>0</sub>)
B0(B<sub>0</sub>)
C0(C<sub>0</sub>)
D0(D<sub>0</sub>)
E0(E<sub>0</sub>)
B1(B<sub>1</sub>)
C1(C<sub>1</sub>)
D1(D<sub>1</sub>)
C2(C<sub>2</sub>)
A0 --> B1
B0 --> B1
C0 --> B1
B0 --> C1
C0 --> C1
D0 --> C1
C0 --> D1
D0 --> D1
E0 --> D1
B1 --> C2
C1 --> C2
D1 --> C2
```
Described in words, this graph shows that before the first layer of the
convolution, the "receptive field" centered at each token consists only
of that same token. That is to say, that we have a receptive field of 1.
The first layer of the convolution adds one neighboring token on either
side to the receptive field. Since this is done on both sides, the
receptive field increases by 2, giving the first layer a receptive field
of 3. The second layer of the convolutions adds an _additional_
neighboring token on either side to the receptive field, giving a final
receptive field of 5.
However, this doesn't match the formula currently given in the docs,
which read:
> The receptive field of the CNN will be
> `depth * (window_size * 2 + 1)`, so a 4-layer network with a window
> size of `2` will be sensitive to 20 words at a time.
Substituting in our depth of 2 and window size of 1, this formula gives
us a receptive field of:
```
depth * (window_size * 2 + 1)
= 2 * (1 * 2 + 1)
= 2 * (2 + 1)
= 2 * 3
= 6
```
This not only doesn't match our computations from above, it's also an
even number! This is suspicious, since the receptive field is supposed
to be centered on a token, and not between tokens. Generally, this
formula results in an even number for any even value of `depth`.
The error in this formula is that the adjustment for the center token
is multiplied by the depth, when it should occur only once. The
corrected formula, `depth * window_size * 2 + 1`, gives the correct
value for our small example from above:
```
depth * window_size * 2 + 1
= 2 * 1 * 2 + 1
= 4 + 1
= 5
```
These changes update the docs to correct the receptive field formula and
the example receptive field size.
* Use isort with Black profile
* isort all the things
* Fix import cycles as a result of import sorting
* Add DOCBIN_ALL_ATTRS type definition
* Add isort to requirements
* Remove isort from build dependencies check
* Typo
* span finder integrated into spacy from experimental
* black
* isort
* black
* default spankey constant
* black
* Update spacy/pipeline/spancat.py
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* rename
* rename
* max_length and min_length as Optional[int] and strict checking
* black
* mypy fix for integer type infinity
* revert line order
* implement all comparison operators for inf int
* avoid two for loops over all docs by not precomputing
* interleave thresholding with span creation
* black
* revert to not interleaving (relized its faster)
* black
* Update spacy/errors.py
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* update dosctring
* enforce that the gold and predicted documents have the same text
* new error for ensuring reference and predicted texts are the same
* remove todo
* adjust test
* black
* handle misaligned tokenization
* return correct variable
* failing overfit test
* only use a single spans_key like in spancat
* black
* remove debug lines
* typo
* remove comment
* remove near duplicate reduntant method
* use the 'spans_key' variable name everywhere
* Update spacy/pipeline/span_finder.py
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* flaky test fix suggestion, hand set bias terms
* only test suggester and test result exhaustively
* make it clear that the span_finder_suggester is more general (not specific to span_finder)
* Update spacy/tests/pipeline/test_span_finder.py
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Apply suggestions from code review
* remove question comment
* move preset_spans_suggester test to spancat tests
* Add docs and unify default configs for spancat and span finder
* Add `allow_overlap=True` to span finder scorer
* Fix offset bug in set_annotations
* Ignore labels in span finder scorer
* Format
* Add span_finder to quickstart template
* Move settings to self.cfg, store min/max unset as None
* Remove debugging
* Update docstrings and docs
* Update spacy/pipeline/span_finder.py
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Fix imports
---------
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Make empty_kb() configurable.
* Format.
* Update docs.
* Be more specific in KB serialization test.
* Update KB serialization tests. Update docs.
* Remove doc update for batched candidate generation.
* Fix serialization of subclassed KB in tests.
* Format.
* Update docstring.
* Update docstring.
* Switch from pickle to json for custom field serialization.
* Handle docs with no entities
If a whole batch contains no entities it won't make it to the model, but
it's possible for individual Docs to have no entities. Before this
commit, those Docs would cause an error when attempting to concatenate
arrays because the dimensions didn't match.
It turns out the process of preparing the Ragged at the end of the span
maker forward was a little different from list2ragged, which just uses
the flatten function directly. Letting list2ragged do the conversion
avoids the dimension issue.
This did not come up before because in NEL demo projects it's typical
for data with no entities to be discarded before it reaches the NEL
component.
This includes a simple direct test that shows the issue and checks it's
resolved. It doesn't check if there are any downstream changes, so a
more complete test could be added. A full run was tested by adding an
example with no entities to the Emerson sample project.
* Add a blank instance to default training data in tests
Rather than adding a specific test, since not failing on instances with
no entities is basic functionality, it makes sense to add it to the
default set.
* Fix without modifying architecture
If the architecture is modified this would have to be a new version, but
this change isn't big enough to merit that.
* Fix TODO about typing
Fix was simple: just request an array2f.
* Add type ignore
Maxout has a more restrictive type than the residual layer expects (only
Floats2d vs any Floats).
* Various cleanup
This moves a lot of lines around but doesn't change any functionality.
Details:
1. use `continue` to reduce indentation
2. move sentence doc building inside conditional since it's otherwise
unused
3. reduces some temporary assignments
* Make changes to typing
* Correction
* Format with black
* Corrections based on review
* Bumped Thinc dependency version
* Bumped blis requirement
* Correction for older Python versions
* Update spacy/ml/models/textcat.py
Co-authored-by: Daniël de Kok <me@github.danieldk.eu>
* Corrections based on review feedback
* Readd deleted docstring line
Co-authored-by: Daniël de Kok <me@github.danieldk.eu>
* Tagger: use unnormalized probabilities for inference
Using unnormalized softmax avoids use of the relatively expensive exp function,
which can significantly speed up non-transformer models (e.g. I got a speedup
of 27% on a German tagging + parsing pipeline).
* Add spacy.Tagger.v2 with configurable normalization
Normalization of probabilities is disabled by default to improve
performance.
* Update documentation, models, and tests to spacy.Tagger.v2
* Move Tagger.v1 to spacy-legacy
* docs/architectures: run prettier
* Unnormalized softmax is now a Softmax_v2 option
* Require thinc 8.0.14 and spacy-legacy 3.0.9
* Partial fix of entity linker batching
* Add import
* Better name
* Add `use_gold_ents` option, docs
* Change to v2, create stub v1, update docs etc.
* Fix error type
Honestly no idea what the right type to use here is.
ConfigValidationError seems wrong. Maybe a NotImplementedError?
* Make mypy happy
* Add hacky fix for init issue
* Add legacy pipeline entity linker
* Fix references to class name
* Add __init__.py for legacy
* Attempted fix for loss issue
* Remove placeholder V1
* formatting
* slightly more interesting train data
* Handle batches with no usable examples
This adds a test for batches that have docs but not entities, and a
check in the component that detects such cases and skips the update step
as thought the batch were empty.
* Remove todo about data verification
Check for empty data was moved further up so this should be OK now - the
case in question shouldn't be possible.
* Fix gradient calculation
The model doesn't know which entities are not in the kb, so it generates
embeddings for the context of all of them.
However, the loss does know which entities aren't in the kb, and it
ignores them, as there's no sensible gradient.
This has the issue that the gradient will not be calculated for some of
the input embeddings, which causes a dimension mismatch in backprop.
That should have caused a clear error, but with numpyops it was causing
nans to happen, which is another problem that should be addressed
separately.
This commit changes the loss to give a zero gradient for entities not in
the kb.
* add failing test for v1 EL legacy architecture
* Add nasty but simple working check for legacy arch
* Clarify why init hack works the way it does
* Clarify use_gold_ents use case
* Fix use gold ents related handling
* Add tests for no gold ents and fix other tests
* Use aligned ents function (not working)
This doesn't actually work because the "aligned" ents are gold-only. But
if I have a different function that returns the intersection, *then*
this will work as desired.
* Use proper matching ent check
This changes the process when gold ents are not used so that the
intersection of ents in the pred and gold is used.
* Move get_matching_ents to Example
* Use model attribute to check for legacy arch
* Rename flag
* bump spacy-legacy to lower 3.0.9
Co-authored-by: svlandeg <svlandeg@github.com>
* Auto-format code with black
* add black requirement to dev dependencies and pin to 22.x
* ignore black dependency for comparison with setup.cfg
Co-authored-by: explosion-bot <explosion-bot@users.noreply.github.com>
Co-authored-by: svlandeg <svlandeg@github.com>
* Use Vectors.shape rather than Vectors.data.shape
* Use Vectors.size rather than Vectors.data.size
* Add Vectors.to_ops to move data between different ops
* Add documentation for Vector.to_ops
* Replace use_ops("numpy") by use_ops("cpu") in the parser
This ensures that the best available CPU implementation is chosen
(e.g. Thinc Apple Ops on macOS).
* Run spaCy tests with apple-thinc-ops on macOS
* 🚨 Ignore all existing Mypy errors
* 🏗 Add Mypy check to CI
* Add types-mock and types-requests as dev requirements
* Add additional type ignore directives
* Add types packages to dev-only list in reqs test
* Add types-dataclasses for python 3.6
* Add ignore to pretrain
* 🏷 Improve type annotation on `run_command` helper
The `run_command` helper previously declared that it returned an
`Optional[subprocess.CompletedProcess]`, but it isn't actually possible
for the function to return `None`. These changes modify the type
annotation of the `run_command` helper and remove all now-unnecessary
`# type: ignore` directives.
* 🔧 Allow variable type redefinition in limited contexts
These changes modify how Mypy is configured to allow variables to have
their type automatically redefined under certain conditions. The Mypy
documentation contains the following example:
```python
def process(items: List[str]) -> None:
# 'items' has type List[str]
items = [item.split() for item in items]
# 'items' now has type List[List[str]]
...
```
This configuration change is especially helpful in reducing the number
of `# type: ignore` directives needed to handle the common pattern of:
* Accepting a filepath as a string
* Overwriting the variable using `filepath = ensure_path(filepath)`
These changes enable redefinition and remove all `# type: ignore`
directives rendered redundant by this change.
* 🏷 Add type annotation to converters mapping
* 🚨 Fix Mypy error in convert CLI argument verification
* 🏷 Improve type annotation on `resolve_dot_names` helper
* 🏷 Add type annotations for `Vocab` attributes `strings` and `vectors`
* 🏷 Add type annotations for more `Vocab` attributes
* 🏷 Add loose type annotation for gold data compilation
* 🏷 Improve `_format_labels` type annotation
* 🏷 Fix `get_lang_class` type annotation
* 🏷 Loosen return type of `Language.evaluate`
* 🏷 Don't accept `Scorer` in `handle_scores_per_type`
* 🏷 Add `string_to_list` overloads
* 🏷 Fix non-Optional command-line options
* 🙈 Ignore redefinition of `wandb_logger` in `loggers.py`
* ➕ Install `typing_extensions` in Python 3.8+
The `typing_extensions` package states that it should be used when
"writing code that must be compatible with multiple Python versions".
Since SpaCy needs to support multiple Python versions, it should be used
when newer `typing` module members are required. One example of this is
`Literal`, which is available starting with Python 3.8.
Previously SpaCy tried to import `Literal` from `typing`, falling back
to `typing_extensions` if the import failed. However, Mypy doesn't seem
to be able to understand what `Literal` means when the initial import
means. Therefore, these changes modify how `compat` imports `Literal` by
always importing it from `typing_extensions`.
These changes also modify how `typing_extensions` is installed, so that
it is a requirement for all Python versions, including those greater
than or equal to 3.8.
* 🏷 Improve type annotation for `Language.pipe`
These changes add a missing overload variant to the type signature of
`Language.pipe`. Additionally, the type signature is enhanced to allow
type checkers to differentiate between the two overload variants based
on the `as_tuple` parameter.
Fixes#8772
* ➖ Don't install `typing-extensions` in Python 3.8+
After more detailed analysis of how to implement Python version-specific
type annotations using SpaCy, it has been determined that by branching
on a comparison against `sys.version_info` can be statically analyzed by
Mypy well enough to enable us to conditionally use
`typing_extensions.Literal`. This means that we no longer need to
install `typing_extensions` for Python versions greater than or equal to
3.8! 🎉
These changes revert previous changes installing `typing-extensions`
regardless of Python version and modify how we import the `Literal` type
to ensure that Mypy treats it properly.
* resolve mypy errors for Strict pydantic types
* refactor code to avoid missing return statement
* fix types of convert CLI command
* avoid list-set confustion in debug_data
* fix typo and formatting
* small fixes to avoid type ignores
* fix types in profile CLI command and make it more efficient
* type fixes in projects CLI
* put one ignore back
* type fixes for render
* fix render types - the sequel
* fix BaseDefault in language definitions
* fix type of noun_chunks iterator - yields tuple instead of span
* fix types in language-specific modules
* 🏷 Expand accepted inputs of `get_string_id`
`get_string_id` accepts either a string (in which case it returns its
ID) or an ID (in which case it immediately returns the ID). These
changes extend the type annotation of `get_string_id` to indicate that
it can accept either strings or IDs.
* 🏷 Handle override types in `combine_score_weights`
The `combine_score_weights` function allows users to pass an `overrides`
mapping to override data extracted from the `weights` argument. Since it
allows `Optional` dictionary values, the return value may also include
`Optional` dictionary values.
These changes update the type annotations for `combine_score_weights` to
reflect this fact.
* 🏷 Fix tokenizer serialization method signatures in `DummyTokenizer`
* 🏷 Fix redefinition of `wandb_logger`
These changes fix the redefinition of `wandb_logger` by giving a
separate name to each `WandbLogger` version. For
backwards-compatibility, `spacy.train` still exports `wandb_logger_v3`
as `wandb_logger` for now.
* more fixes for typing in language
* type fixes in model definitions
* 🏷 Annotate `_RandomWords.probs` as `NDArray`
* 🏷 Annotate `tok2vec` layers to help Mypy
* 🐛 Fix `_RandomWords.probs` type annotations for Python 3.6
Also remove an import that I forgot to move to the top of the module 😅
* more fixes for matchers and other pipeline components
* quick fix for entity linker
* fixing types for spancat, textcat, etc
* bugfix for tok2vec
* type annotations for scorer
* add runtime_checkable for Protocol
* type and import fixes in tests
* mypy fixes for training utilities
* few fixes in util
* fix import
* 🐵 Remove unused `# type: ignore` directives
* 🏷 Annotate `Language._components`
* 🏷 Annotate `spacy.pipeline.Pipe`
* add doc as property to span.pyi
* small fixes and cleanup
* explicit type annotations instead of via comment
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
Co-authored-by: svlandeg <sofie.vanlandeghem@gmail.com>
Co-authored-by: svlandeg <svlandeg@github.com>
* 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>
* 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
* Replace negative rows with 0 in StaticVectors
Replace negative row indices with 0-vectors in `StaticVectors`.
* Increase versions related to StaticVectors
* Increase versions of all architctures and layers related to
`StaticVectors`
* Improve efficiency of 0-vector operations
Parallel `spacy-legacy` PR: https://github.com/explosion/spacy-legacy/pull/5
* Update config defaults to new versions
* Update docs
* 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 TorchBiLSTMEncoder documentation
* ensure the types of the encoding Tok2vec layers are correct
* update references from v1 to v2 for the new architectures
* add convenience method to determine tok2vec width in a model
* fix transformer tok2vec dimensions in TextCatEnsemble architecture
* init function should not be nested to avoid pickle issues
* define new architectures for the pretraining objective
* add loss function as attr of the omdel
* cleanup
* cleanup
* shorten name
* fix typo
* remove unused error
* small fix in example imports
* throw error when train_corpus or dev_corpus is not a string
* small fix in custom logger example
* limit macro_auc to labels with 2 annotations
* fix typo
* also create parents of output_dir if need be
* update documentation of textcat scores
* refactor TextCatEnsemble
* fix tests for new AUC definition
* bump to 3.0.0a42
* update docs
* rename to spacy.TextCatEnsemble.v2
* spacy.TextCatEnsemble.v1 in legacy
* cleanup
* small fix
* update to 3.0.0rc2
* fix import that got lost in merge
* cursed IDE
* fix two typos
