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.

spacy/errors.py

@@ -984,6 +984,8 @@ 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_max` or `use_reduce_mean`.")
 
 
 # Deprecated model shortcuts, only used in errors and warnings

spacy/ml/models/textcat.py

@@ -17,6 +17,8 @@ from thinc.api import (
     clone,
     concatenate,
     list2ragged,
+    reduce_first,
+    reduce_max,
     reduce_mean,
     reduce_sum,
     residual,
@@ -49,39 +51,14 @@ 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}
+    return build_reduce_text_classifier(
-    with Model.define_operators({">>": chain}):
+        tok2vec=tok2vec,
-        cnn = tok2vec >> list2ragged() >> reduce_mean()
+        exclusive_classes=exclusive_classes,
-        nI = tok2vec.maybe_get_dim("nO")
+        use_reduce_first=False,
-        if exclusive_classes:
+        use_reduce_max=False,
-            output_layer = Softmax(nO=nO, nI=nI)
+        use_reduce_mean=True,
-            fill_defaults["b"] = NEG_VALUE
+        nO=nO,
-            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
 
 
 def resize_and_set_ref(model, new_nO, resizable_layer):
@@ -230,3 +207,75 @@ 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_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
+    multible 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_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_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

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", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True, "use_reduce_first": True, "use_reduce_max": True, "use_reduce_mean": True}),
-        ("textcat_multilabel", {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False}),
+        ("textcat_multilabel", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False, "use_reduce_first": 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
+        # REDUCE
-        ("textcat", {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True}),
+        ("textcat", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True, "use_reduce_first": True, "use_reduce_max": True, "use_reduce_mean": True}),
-        ("textcat_multilabel", {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False}),
+        ("textcat_multilabel", {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False, "use_reduce_first": True, "use_reduce_max": True, "use_reduce_mean": True}),
     ],
 )
 # fmt: on
@@ -701,9 +701,9 @@ 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
+        # REDUCE V1
-        ("textcat", TRAIN_DATA_SINGLE_LABEL, {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True}),
+        ("textcat", TRAIN_DATA_SINGLE_LABEL, {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": True, "use_reduce_first": True, "use_reduce_max": True, "use_reduce_mean": True}),
-        ("textcat_multilabel", TRAIN_DATA_MULTI_LABEL, {"@architectures": "spacy.TextCatCNN.v2", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False}),
+        ("textcat_multilabel", TRAIN_DATA_MULTI_LABEL, {"@architectures": "spacy.TextCatReduce.v1", "tok2vec": DEFAULT_TOK2VEC_MODEL, "exclusive_classes": False, "use_reduce_first": True, "use_reduce_max": True, "use_reduce_mean": True}),
     ],
 )
 # fmt: on

spacy/tests/test_models.py

@@ -26,6 +26,8 @@ from spacy.ml.models import (
     build_Tok2Vec_model,
 )
 from spacy.ml.staticvectors import StaticVectors
+from spacy.pipeline import tok2vec
+from spacy.util import registry
 
 
 def get_textcat_bow_kwargs():
@@ -284,3 +286,16 @@ 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_max=False,
+            use_reduce_mean=False,
+        )

website/docs/api/architectures.mdx

@@ -1043,6 +1043,9 @@ 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.
 
+This model is identical to [TexCatReduce.v1](#TextCatReduce) with
+`use_reduce_mean=true`.
+
 | Name                | Description                                                                                                                                                                                    |
 | ------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `exclusive_classes` | Whether or not categories are mutually exclusive. ~~bool~~                                                                                                                                     |
@@ -1096,6 +1099,44 @@ the others, but may not be as accurate, especially if texts are short.
 
 </Accordion>
 
+### spacy.TextCatReduce.v1 {id="TextCatReduce"}
+
+> #### Example Config
+>
+> ```ini
+> [model]
+> @architectures = "spacy.TextCatReduce.v1"
+> exclusive_classes = false
+> use_reduce_first = 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.
+
+| 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_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"}