From 7ebba8640298c8f111ecaa73e769e993b981b76f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Dani=C3=ABl=20de=20Kok?= Date: Thu, 21 Dec 2023 11:00:06 +0100 Subject: [PATCH] Add TextCatReduce.v1 (#13181) * Add TextCatReduce.v1 This is a textcat classifier that pools the vectors generated by a tok2vec implementation and then applies a classifier to the pooled representation. Three reductions are supported for pooling: first, max, and mean. When multiple reductions are enabled, the reductions are concatenated before providing them to the classification layer. This model is a generalization of the TextCatCNN model, which only supports mean reductions and is a bit of a misnomer, because it can also be used with transformers. This change also reimplements TextCatCNN.v2 using the new TextCatReduce.v1 layer. * Doc fixes Co-authored-by: Sofie Van Landeghem * Fully specify `TextCatCNN` <-> `TextCatReduce` equivalence * Move TextCatCNN docs to legacy, in prep for moving to spacy-legacy * Add back a test for TextCatCNN.v2 * Replace TextCatCNN in pipe configurations and templates * Add an infobox to the `TextCatReduce` section with an `TextCatCNN` anchor * Add last reduction (`use_reduce_last`) * Remove non-working TextCatCNN Netlify redirect * Revert layer changes for the quickstart * Revert one more quickstart change * Remove unused import * Fix docstring * Fix setting name in error message --------- Co-authored-by: Sofie Van Landeghem Co-authored-by: Adriane Boyd --- spacy/errors.py | 3 + spacy/ml/models/textcat.py | 122 +++++++++++++++++++-------- spacy/pipeline/textcat.py | 6 +- spacy/pipeline/textcat_multilabel.py | 6 +- spacy/tests/pipeline/test_textcat.py | 15 ++-- spacy/tests/test_models.py | 15 ++++ website/docs/api/architectures.mdx | 88 ++++++++++--------- website/docs/api/legacy.mdx | 50 ++++++++++- 8 files changed, 223 insertions(+), 82 deletions(-) diff --git a/spacy/errors.py b/spacy/errors.py index 093c65f3d..b6108dd0f 100644 --- a/spacy/errors.py +++ b/spacy/errors.py @@ -984,6 +984,9 @@ class Errors(metaclass=ErrorsWithCodes): E1055 = ("The 'replace_listener' callback expects {num_params} parameters, " "but only callbacks with one or three parameters are supported") E1056 = ("The `TextCatBOW` architecture expects a length of at least 1, was {length}.") + E1057 = ("The `TextCatReduce` architecture must be used with at least one " + "reduction. Please enable one of `use_reduce_first`, " + "`use_reduce_last`, `use_reduce_max` or `use_reduce_mean`.") # Deprecated model shortcuts, only used in errors and warnings diff --git a/spacy/ml/models/textcat.py b/spacy/ml/models/textcat.py index e6d1f030f..93929bd4e 100644 --- a/spacy/ml/models/textcat.py +++ b/spacy/ml/models/textcat.py @@ -17,6 +17,9 @@ from thinc.api import ( clone, concatenate, list2ragged, + reduce_first, + reduce_last, + reduce_max, reduce_mean, reduce_sum, residual, @@ -49,39 +52,15 @@ def build_simple_cnn_text_classifier( outputs sum to 1. If exclusive_classes=False, a logistic non-linearity is applied instead, so that outputs are in the range [0, 1]. """ - fill_defaults = {"b": 0, "W": 0} - with Model.define_operators({">>": chain}): - cnn = tok2vec >> list2ragged() >> reduce_mean() - nI = tok2vec.maybe_get_dim("nO") - if exclusive_classes: - output_layer = Softmax(nO=nO, nI=nI) - fill_defaults["b"] = NEG_VALUE - resizable_layer: Model = resizable( - output_layer, - resize_layer=partial( - resize_linear_weighted, fill_defaults=fill_defaults - ), - ) - model = cnn >> resizable_layer - else: - output_layer = Linear(nO=nO, nI=nI) - resizable_layer = resizable( - output_layer, - resize_layer=partial( - resize_linear_weighted, fill_defaults=fill_defaults - ), - ) - model = cnn >> resizable_layer >> Logistic() - model.set_ref("output_layer", output_layer) - model.attrs["resize_output"] = partial( - resize_and_set_ref, - resizable_layer=resizable_layer, - ) - model.set_ref("tok2vec", tok2vec) - if nO is not None: - model.set_dim("nO", cast(int, nO)) - model.attrs["multi_label"] = not exclusive_classes - return model + return build_reduce_text_classifier( + tok2vec=tok2vec, + exclusive_classes=exclusive_classes, + use_reduce_first=False, + use_reduce_last=False, + use_reduce_max=False, + use_reduce_mean=True, + nO=nO, + ) def resize_and_set_ref(model, new_nO, resizable_layer): @@ -230,3 +209,80 @@ def build_text_classifier_lowdata( model = model >> Dropout(dropout) model = model >> Logistic() return model + + +@registry.architectures("spacy.TextCatReduce.v1") +def build_reduce_text_classifier( + tok2vec: Model, + exclusive_classes: bool, + use_reduce_first: bool, + use_reduce_last: bool, + use_reduce_max: bool, + use_reduce_mean: bool, + nO: Optional[int] = None, +) -> Model[List[Doc], Floats2d]: + """Build a model that classifies pooled `Doc` representations. + + Pooling is performed using reductions. Reductions are concatenated when + multiple reductions are used. + + tok2vec (Model): the tok2vec layer to pool over. + exclusive_classes (bool): Whether or not classes are mutually exclusive. + use_reduce_first (bool): Pool by using the hidden representation of the + first token of a `Doc`. + use_reduce_last (bool): Pool by using the hidden representation of the + last token of a `Doc`. + use_reduce_max (bool): Pool by taking the maximum values of the hidden + representations of a `Doc`. + use_reduce_mean (bool): Pool by taking the mean of all hidden + representations of a `Doc`. + nO (Optional[int]): Number of classes. + """ + + fill_defaults = {"b": 0, "W": 0} + reductions = [] + if use_reduce_first: + reductions.append(reduce_first()) + if use_reduce_last: + reductions.append(reduce_last()) + if use_reduce_max: + reductions.append(reduce_max()) + if use_reduce_mean: + reductions.append(reduce_mean()) + + if not len(reductions): + raise ValueError(Errors.E1057) + + with Model.define_operators({">>": chain}): + cnn = tok2vec >> list2ragged() >> concatenate(*reductions) + nO_tok2vec = tok2vec.maybe_get_dim("nO") + nI = nO_tok2vec * len(reductions) if nO_tok2vec is not None else None + if exclusive_classes: + output_layer = Softmax(nO=nO, nI=nI) + fill_defaults["b"] = NEG_VALUE + resizable_layer: Model = resizable( + output_layer, + resize_layer=partial( + resize_linear_weighted, fill_defaults=fill_defaults + ), + ) + model = cnn >> resizable_layer + else: + output_layer = Linear(nO=nO, nI=nI) + resizable_layer = resizable( + output_layer, + resize_layer=partial( + resize_linear_weighted, fill_defaults=fill_defaults + ), + ) + model = cnn >> resizable_layer >> Logistic() + model.set_ref("output_layer", output_layer) + model.attrs["resize_output"] = partial( + resize_and_set_ref, + resizable_layer=resizable_layer, + ) + model.set_ref("tok2vec", tok2vec) + if nO is not None: + model.set_dim("nO", cast(int, nO)) + model.attrs["multi_label"] = not exclusive_classes + return model diff --git a/spacy/pipeline/textcat.py b/spacy/pipeline/textcat.py index 43a335c4a..ae227017a 100644 --- a/spacy/pipeline/textcat.py +++ b/spacy/pipeline/textcat.py @@ -55,8 +55,12 @@ no_output_layer = false single_label_cnn_config = """ [model] -@architectures = "spacy.TextCatCNN.v2" +@architectures = "spacy.TextCatReduce.v1" exclusive_classes = true +use_reduce_first = false +use_reduce_last = false +use_reduce_max = false +use_reduce_mean = true [model.tok2vec] @architectures = "spacy.HashEmbedCNN.v2" diff --git a/spacy/pipeline/textcat_multilabel.py b/spacy/pipeline/textcat_multilabel.py index c917cc610..2f8d5e604 100644 --- a/spacy/pipeline/textcat_multilabel.py +++ b/spacy/pipeline/textcat_multilabel.py @@ -53,8 +53,12 @@ no_output_layer = false multi_label_cnn_config = """ [model] -@architectures = "spacy.TextCatCNN.v2" +@architectures = "spacy.TextCatReduce.v1" exclusive_classes = false +use_reduce_first = false +use_reduce_last = false +use_reduce_max = false +use_reduce_mean = true [model.tok2vec] @architectures = "spacy.HashEmbedCNN.v2" diff --git a/spacy/tests/pipeline/test_textcat.py b/spacy/tests/pipeline/test_textcat.py index 147ea4900..5dff8d124 100644 --- a/spacy/tests/pipeline/test_textcat.py +++ b/spacy/tests/pipeline/test_textcat.py @@ -457,8 +457,8 @@ def test_no_resize(name, textcat_config): ("textcat_multilabel", {"@architectures": "spacy.TextCatBOW.v3", "exclusive_classes": False, "no_output_layer": False, "ngram_size": 3}), ("textcat_multilabel", {"@architectures": "spacy.TextCatBOW.v3", "exclusive_classes": False, "no_output_layer": True, "ngram_size": 3}), # CNN - ("textcat", {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True}), - ("textcat_multilabel", {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False}), + ("textcat", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True, "use_reduce_first": True, "use_reduce_last": True, "use_reduce_max": True, "use_reduce_mean": True}), + ("textcat_multilabel", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False, "use_reduce_first": True, "use_reduce_last": True, "use_reduce_max": True, "use_reduce_mean": True}), ], ) # fmt: on @@ -485,9 +485,9 @@ def test_resize(name, textcat_config): ("textcat", {"@architectures": "spacy.TextCatBOW.v3", "exclusive_classes": True, "no_output_layer": True, "ngram_size": 3}), ("textcat_multilabel", {"@architectures": "spacy.TextCatBOW.v3", "exclusive_classes": False, "no_output_layer": False, "ngram_size": 3}), ("textcat_multilabel", {"@architectures": "spacy.TextCatBOW.v3", "exclusive_classes": False, "no_output_layer": True, "ngram_size": 3}), - # CNN - ("textcat", {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True}), - ("textcat_multilabel", {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False}), + # REDUCE + ("textcat", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True, "use_reduce_first": True, "use_reduce_last": True, "use_reduce_max": True, "use_reduce_mean": True}), + ("textcat_multilabel", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False, "use_reduce_first": True, "use_reduce_last": True, "use_reduce_max": True, "use_reduce_mean": True}), ], ) # fmt: on @@ -701,9 +701,12 @@ def test_overfitting_IO_multi(): # ENSEMBLE V2 ("textcat_multilabel", TRAIN_DATA_MULTI_LABEL, {"@architectures": "spacy.TextCatEnsemble.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "linear_model": {"@architectures": "spacy.TextCatBOW.v3", "exclusive_classes": False, "ngram_size": 1, "no_output_layer": False}}), ("textcat", TRAIN_DATA_SINGLE_LABEL, {"@architectures": "spacy.TextCatEnsemble.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "linear_model": {"@architectures": "spacy.TextCatBOW.v3", "exclusive_classes": True, "ngram_size": 5, "no_output_layer": False}}), - # CNN V2 + # CNN V2 (legacy) ("textcat", TRAIN_DATA_SINGLE_LABEL, {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True}), ("textcat_multilabel", TRAIN_DATA_MULTI_LABEL, {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False}), + # REDUCE V1 + ("textcat", TRAIN_DATA_SINGLE_LABEL, {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True, "use_reduce_first": True, "use_reduce_last": True, "use_reduce_max": True, "use_reduce_mean": True}), + ("textcat_multilabel", TRAIN_DATA_MULTI_LABEL, {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False, "use_reduce_first": True, "use_reduce_last": True, "use_reduce_max": True, "use_reduce_mean": True}), ], ) # fmt: on diff --git a/spacy/tests/test_models.py b/spacy/tests/test_models.py index e6692ad92..5228b4544 100644 --- a/spacy/tests/test_models.py +++ b/spacy/tests/test_models.py @@ -26,6 +26,7 @@ from spacy.ml.models import ( build_Tok2Vec_model, ) from spacy.ml.staticvectors import StaticVectors +from spacy.util import registry def get_textcat_bow_kwargs(): @@ -284,3 +285,17 @@ def test_spancat_model_forward_backward(nO=5): Y, backprop = model((docs, spans), is_train=True) assert Y.shape == (spans.dataXd.shape[0], nO) backprop(Y) + + +def test_textcat_reduce_invalid_args(): + textcat_reduce = registry.architectures.get("spacy.TextCatReduce.v1") + tok2vec = make_test_tok2vec() + with pytest.raises(ValueError, match=r"must be used with at least one reduction"): + textcat_reduce( + tok2vec=tok2vec, + exclusive_classes=False, + use_reduce_first=False, + use_reduce_last=False, + use_reduce_max=False, + use_reduce_mean=False, + ) diff --git a/website/docs/api/architectures.mdx b/website/docs/api/architectures.mdx index 9d8b3ddfa..63f723a28 100644 --- a/website/docs/api/architectures.mdx +++ b/website/docs/api/architectures.mdx @@ -1018,46 +1018,6 @@ but used an internal `tok2vec` instead of taking it as argument: -### spacy.TextCatCNN.v2 {id="TextCatCNN"} - -> #### Example Config -> -> ```ini -> [model] -> @architectures = "spacy.TextCatCNN.v2" -> exclusive_classes = false -> nO = null -> -> [model.tok2vec] -> @architectures = "spacy.HashEmbedCNN.v2" -> pretrained_vectors = null -> width = 96 -> depth = 4 -> embed_size = 2000 -> window_size = 1 -> maxout_pieces = 3 -> subword_features = true -> ``` - -A neural network model where token vectors are calculated using a CNN. The -vectors are mean pooled and used as features in a feed-forward network. This -architecture is usually less accurate than the ensemble, but runs faster. - -| Name | Description | -| ------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| `exclusive_classes` | Whether or not categories are mutually exclusive. ~~bool~~ | -| `tok2vec` | The [`tok2vec`](#tok2vec) layer of the model. ~~Model~~ | -| `nO` | Output dimension, determined by the number of different labels. If not set, the [`TextCategorizer`](/api/textcategorizer) component will set it when `initialize` is called. ~~Optional[int]~~ | -| **CREATES** | The model using the architecture. ~~Model[List[Doc], Floats2d]~~ | - - - -[TextCatCNN.v1](/api/legacy#TextCatCNN_v1) had the exact same signature, but was -not yet resizable. Since v2, new labels can be added to this component, even -after training. - - - ### spacy.TextCatBOW.v3 {id="TextCatBOW"} > #### Example Config @@ -1096,6 +1056,54 @@ the others, but may not be as accurate, especially if texts are short. +### spacy.TextCatReduce.v1 {id="TextCatReduce"} + +> #### Example Config +> +> ```ini +> [model] +> @architectures = "spacy.TextCatReduce.v1" +> exclusive_classes = false +> use_reduce_first = false +> use_reduce_last = false +> use_reduce_max = false +> use_reduce_mean = true +> nO = null +> +> [model.tok2vec] +> @architectures = "spacy.HashEmbedCNN.v2" +> pretrained_vectors = null +> width = 96 +> depth = 4 +> embed_size = 2000 +> window_size = 1 +> maxout_pieces = 3 +> subword_features = true +> ``` + +A classifier that pools token hidden representations of each `Doc` using first, +max or mean reduction and then applies a classification layer. Reductions are +concatenated when multiple reductions are used. + + + +`TextCatReduce` is a generalization of the older +[`TextCatCNN`](/api/legacy#TextCatCNN_v2) model. `TextCatCNN` always uses a mean +reduction, whereas `TextCatReduce` also supports first/max reductions. + + + +| Name | Description | +| ------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | +| `exclusive_classes` | Whether or not categories are mutually exclusive. ~~bool~~ | +| `tok2vec` | The [`tok2vec`](#tok2vec) layer of the model. ~~Model~~ | +| `use_reduce_first` | Pool by using the hidden representation of the first token of a `Doc`. ~~bool~~ | +| `use_reduce_last` | Pool by using the hidden representation of the last token of a `Doc`. ~~bool~~ | +| `use_reduce_max` | Pool by taking the maximum values of the hidden representations of a `Doc`. ~~bool~~ | +| `use_reduce_mean` | Pool by taking the mean of all hidden representations of a `Doc`. ~~bool~~ | +| `nO` | Output dimension, determined by the number of different labels. If not set, the [`TextCategorizer`](/api/textcategorizer) component will set it when `initialize` is called. ~~Optional[int]~~ | +| **CREATES** | The model using the architecture. ~~Model[List[Doc], Floats2d]~~ | + ## Span classification architectures {id="spancat",source="spacy/ml/models/spancat.py"} ### spacy.SpanCategorizer.v1 {id="SpanCategorizer"} diff --git a/website/docs/api/legacy.mdx b/website/docs/api/legacy.mdx index 32111ce92..b44df5387 100644 --- a/website/docs/api/legacy.mdx +++ b/website/docs/api/legacy.mdx @@ -162,7 +162,10 @@ network has an internal CNN Tok2Vec layer and uses attention. Since `spacy.TextCatCNN.v2`, this architecture has become resizable, which means that you can add labels to a previously trained textcat. `TextCatCNN` v1 did not -yet support that. +yet support that. `TextCatCNN` has been replaced by the more general +[`TextCatReduce`](/api/architectures#TextCatReduce) layer. `TextCatCNN` is +identical to `TextCatReduce` with `use_reduce_mean=true`, +`use_reduce_first=false`, `reduce_last=false` and `use_reduce_max=false`. > #### Example Config > @@ -194,6 +197,51 @@ architecture is usually less accurate than the ensemble, but runs faster. | `nO` | Output dimension, determined by the number of different labels. If not set, the [`TextCategorizer`](/api/textcategorizer) component will set it when `initialize` is called. ~~Optional[int]~~ | | **CREATES** | The model using the architecture. ~~Model[List[Doc], Floats2d]~~ | +### spacy.TextCatCNN.v2 {id="TextCatCNN_v2"} + +> #### Example Config +> +> ```ini +> [model] +> @architectures = "spacy.TextCatCNN.v2" +> exclusive_classes = false +> nO = null +> +> [model.tok2vec] +> @architectures = "spacy.HashEmbedCNN.v2" +> pretrained_vectors = null +> width = 96 +> depth = 4 +> embed_size = 2000 +> window_size = 1 +> maxout_pieces = 3 +> subword_features = true +> ``` + +A neural network model where token vectors are calculated using a CNN. The +vectors are mean pooled and used as features in a feed-forward network. This +architecture is usually less accurate than the ensemble, but runs faster. + +`TextCatCNN` has been replaced by the more general +[`TextCatReduce`](/api/architectures#TextCatReduce) layer. `TextCatCNN` is +identical to `TextCatReduce` with `use_reduce_mean=true`, +`use_reduce_first=false`, `reduce_last=false` and `use_reduce_max=false`. + +| Name | Description | +| ------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | +| `exclusive_classes` | Whether or not categories are mutually exclusive. ~~bool~~ | +| `tok2vec` | The [`tok2vec`](#tok2vec) layer of the model. ~~Model~~ | +| `nO` | Output dimension, determined by the number of different labels. If not set, the [`TextCategorizer`](/api/textcategorizer) component will set it when `initialize` is called. ~~Optional[int]~~ | +| **CREATES** | The model using the architecture. ~~Model[List[Doc], Floats2d]~~ | + + + +[TextCatCNN.v1](/api/legacy#TextCatCNN_v1) had the exact same signature, but was +not yet resizable. Since v2, new labels can be added to this component, even +after training. + + + ### spacy.TextCatBOW.v1 {id="TextCatBOW_v1"} Since `spacy.TextCatBOW.v2`, this architecture has become resizable, which means