spaCy/spacy/ml/models/textcat.py
Daniël de Kok e2a3952de5
Add spacy.TextCatParametricAttention.v1 (#13201)
* Add spacy.TextCatParametricAttention.v1

This layer provides is a simplification of the ensemble classifier that
only uses paramteric attention. We have found empirically that with a
sufficient amount of training data, using the ensemble classifier with
BoW does not provide significant improvement in classifier accuracy.
However, plugging in a BoW classifier does reduce GPU training and
inference performance substantially, since it uses a GPU-only kernel.

* Fix merge fallout
2024-01-02 10:03:06 +01:00

359 lines
12 KiB
Python

from functools import partial
from typing import List, Optional, Tuple, cast
from thinc.api import (
Dropout,
Gelu,
LayerNorm,
Linear,
Logistic,
Maxout,
Model,
ParametricAttention,
ParametricAttention_v2,
Relu,
Softmax,
SparseLinear,
SparseLinear_v2,
chain,
clone,
concatenate,
list2ragged,
reduce_first,
reduce_last,
reduce_max,
reduce_mean,
reduce_sum,
residual,
resizable,
softmax_activation,
with_cpu,
)
from thinc.layers.chain import init as init_chain
from thinc.layers.resizable import resize_linear_weighted, resize_model
from thinc.types import ArrayXd, Floats2d
from ...attrs import ORTH
from ...errors import Errors
from ...tokens import Doc
from ...util import registry
from ..extract_ngrams import extract_ngrams
from ..staticvectors import StaticVectors
from .tok2vec import get_tok2vec_width
NEG_VALUE = -5000
@registry.architectures("spacy.TextCatCNN.v2")
def build_simple_cnn_text_classifier(
tok2vec: Model, exclusive_classes: bool, nO: Optional[int] = None
) -> Model[List[Doc], Floats2d]:
"""
Build a simple CNN text classifier, given a token-to-vector model as inputs.
If exclusive_classes=True, a softmax non-linearity is applied, so that the
outputs sum to 1. If exclusive_classes=False, a logistic non-linearity
is applied instead, so that outputs are in the range [0, 1].
"""
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):
resizable_layer = resize_model(resizable_layer, new_nO)
model.set_ref("output_layer", resizable_layer.layers[0])
model.set_dim("nO", new_nO, force=True)
return model
@registry.architectures("spacy.TextCatBOW.v2")
def build_bow_text_classifier(
exclusive_classes: bool,
ngram_size: int,
no_output_layer: bool,
nO: Optional[int] = None,
) -> Model[List[Doc], Floats2d]:
return _build_bow_text_classifier(
exclusive_classes=exclusive_classes,
ngram_size=ngram_size,
no_output_layer=no_output_layer,
nO=nO,
sparse_linear=SparseLinear(nO=nO),
)
@registry.architectures("spacy.TextCatBOW.v3")
def build_bow_text_classifier_v3(
exclusive_classes: bool,
ngram_size: int,
no_output_layer: bool,
length: int = 262144,
nO: Optional[int] = None,
) -> Model[List[Doc], Floats2d]:
if length < 1:
raise ValueError(Errors.E1056.format(length=length))
# Find k such that 2**(k-1) < length <= 2**k.
length = 2 ** (length - 1).bit_length()
return _build_bow_text_classifier(
exclusive_classes=exclusive_classes,
ngram_size=ngram_size,
no_output_layer=no_output_layer,
nO=nO,
sparse_linear=SparseLinear_v2(nO=nO, length=length),
)
def _build_bow_text_classifier(
exclusive_classes: bool,
ngram_size: int,
no_output_layer: bool,
sparse_linear: Model[Tuple[ArrayXd, ArrayXd, ArrayXd], ArrayXd],
nO: Optional[int] = None,
) -> Model[List[Doc], Floats2d]:
fill_defaults = {"b": 0, "W": 0}
with Model.define_operators({">>": chain}):
output_layer = None
if not no_output_layer:
fill_defaults["b"] = NEG_VALUE
output_layer = softmax_activation() if exclusive_classes else Logistic()
resizable_layer: Model[Floats2d, Floats2d] = resizable(
sparse_linear,
resize_layer=partial(resize_linear_weighted, fill_defaults=fill_defaults),
)
model = extract_ngrams(ngram_size, attr=ORTH) >> resizable_layer
model = with_cpu(model, model.ops)
if output_layer:
model = model >> with_cpu(output_layer, output_layer.ops)
if nO is not None:
model.set_dim("nO", cast(int, nO))
model.set_ref("output_layer", sparse_linear)
model.attrs["multi_label"] = not exclusive_classes
model.attrs["resize_output"] = partial(
resize_and_set_ref, resizable_layer=resizable_layer
)
return model
@registry.architectures("spacy.TextCatEnsemble.v2")
def build_text_classifier_v2(
tok2vec: Model[List[Doc], List[Floats2d]],
linear_model: Model[List[Doc], Floats2d],
nO: Optional[int] = None,
) -> Model[List[Doc], Floats2d]:
# TODO: build the model with _build_parametric_attention_with_residual_nonlinear
# in spaCy v4. We don't do this in spaCy v3 to preserve model
# compatibility.
exclusive_classes = not linear_model.attrs["multi_label"]
with Model.define_operators({">>": chain, "|": concatenate}):
width = tok2vec.maybe_get_dim("nO")
attention_layer = ParametricAttention(width)
maxout_layer = Maxout(nO=width, nI=width)
norm_layer = LayerNorm(nI=width)
cnn_model = (
tok2vec
>> list2ragged()
>> attention_layer
>> reduce_sum()
>> residual(maxout_layer >> norm_layer >> Dropout(0.0))
)
nO_double = nO * 2 if nO else None
if exclusive_classes:
output_layer = Softmax(nO=nO, nI=nO_double)
else:
output_layer = Linear(nO=nO, nI=nO_double) >> Logistic()
model = (linear_model | cnn_model) >> output_layer
model.set_ref("tok2vec", tok2vec)
if model.has_dim("nO") is not False and nO is not None:
model.set_dim("nO", cast(int, nO))
model.set_ref("output_layer", linear_model.get_ref("output_layer"))
model.set_ref("attention_layer", attention_layer)
model.set_ref("maxout_layer", maxout_layer)
model.set_ref("norm_layer", norm_layer)
model.attrs["multi_label"] = not exclusive_classes
model.init = init_ensemble_textcat # type: ignore[assignment]
return model
def init_ensemble_textcat(model, X, Y) -> Model:
tok2vec_width = get_tok2vec_width(model)
model.get_ref("attention_layer").set_dim("nO", tok2vec_width)
model.get_ref("maxout_layer").set_dim("nO", tok2vec_width)
model.get_ref("maxout_layer").set_dim("nI", tok2vec_width)
model.get_ref("norm_layer").set_dim("nI", tok2vec_width)
model.get_ref("norm_layer").set_dim("nO", tok2vec_width)
init_chain(model, X, Y)
return model
@registry.architectures("spacy.TextCatLowData.v1")
def build_text_classifier_lowdata(
width: int, dropout: Optional[float], nO: Optional[int] = None
) -> Model[List[Doc], Floats2d]:
# Don't document this yet, I'm not sure it's right.
# Note, before v.3, this was the default if setting "low_data" and "pretrained_dims"
with Model.define_operators({">>": chain, "**": clone}):
model = (
StaticVectors(width)
>> list2ragged()
>> ParametricAttention(width)
>> reduce_sum()
>> residual(Relu(width, width)) ** 2
>> Linear(nO, width)
)
if dropout:
model = model >> Dropout(dropout)
model = model >> Logistic()
return model
@registry.architectures("spacy.TextCatParametricAttention.v1")
def build_textcat_parametric_attention_v1(
tok2vec: Model[List[Doc], List[Floats2d]],
exclusive_classes: bool,
nO: Optional[int] = None,
) -> Model[List[Doc], Floats2d]:
width = tok2vec.maybe_get_dim("nO")
parametric_attention = _build_parametric_attention_with_residual_nonlinear(
tok2vec=tok2vec,
nonlinear_layer=Maxout(nI=width, nO=width),
key_transform=Gelu(nI=width, nO=width),
)
with Model.define_operators({">>": chain}):
if exclusive_classes:
output_layer = Softmax(nO=nO)
else:
output_layer = Linear(nO=nO) >> Logistic()
model = parametric_attention >> output_layer
if model.has_dim("nO") is not False and nO is not None:
model.set_dim("nO", cast(int, nO))
model.set_ref("output_layer", output_layer)
model.attrs["multi_label"] = not exclusive_classes
return model
def _build_parametric_attention_with_residual_nonlinear(
*,
tok2vec: Model[List[Doc], List[Floats2d]],
nonlinear_layer: Model[Floats2d, Floats2d],
key_transform: Optional[Model[Floats2d, Floats2d]] = None,
) -> Model[List[Doc], Floats2d]:
with Model.define_operators({">>": chain, "|": concatenate}):
width = tok2vec.maybe_get_dim("nO")
attention_layer = ParametricAttention_v2(nO=width, key_transform=key_transform)
norm_layer = LayerNorm(nI=width)
parametric_attention = (
tok2vec
>> list2ragged()
>> attention_layer
>> reduce_sum()
>> residual(nonlinear_layer >> norm_layer >> Dropout(0.0))
)
parametric_attention.init = _init_parametric_attention_with_residual_nonlinear
parametric_attention.set_ref("tok2vec", tok2vec)
parametric_attention.set_ref("attention_layer", attention_layer)
parametric_attention.set_ref("nonlinear_layer", nonlinear_layer)
parametric_attention.set_ref("norm_layer", norm_layer)
return parametric_attention
def _init_parametric_attention_with_residual_nonlinear(model, X, Y) -> Model:
tok2vec_width = get_tok2vec_width(model)
model.get_ref("attention_layer").set_dim("nO", tok2vec_width)
model.get_ref("nonlinear_layer").set_dim("nO", tok2vec_width)
model.get_ref("nonlinear_layer").set_dim("nI", tok2vec_width)
model.get_ref("norm_layer").set_dim("nI", tok2vec_width)
model.get_ref("norm_layer").set_dim("nO", tok2vec_width)
init_chain(model, X, Y)
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