def cnn_tagger(width: int, vector_width: int, nr_classes: int = 17):
with Model.define_operators({">>": chain}):
model = strings2arrays() >> with_array(
HashEmbed(nO=width, nV=vector_width, column=0) >> expand_window(
window_size=1) >> Relu(nO=width, nI=width * 3) >> Relu(
nO=width, nI=width) >> Softmax(nO=nr_classes, nI=width))
return model
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]:
exclusive_classes = not linear_model.attrs["multi_label"]
with Model.define_operators({">>": chain, "|": concatenate}):
width = tok2vec.maybe_get_dim("nO")
attention_layer = ParametricAttention(
width) # TODO: benchmark performance difference of this layer
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:
model.set_dim("nO", 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
return model
def main(n_hidden: int = 256,
dropout: float = 0.2,
n_iter: int = 10,
batch_size: int = 128):
# Define the model
model: Model = chain(
Relu(nO=n_hidden, dropout=dropout),
Relu(nO=n_hidden, dropout=dropout),
Softmax(),
)
# Load the data
(train_X, train_Y), (dev_X, dev_Y) = ml_datasets.mnist()
# Set any missing shapes for the model.
model.initialize(X=train_X[:5], Y=train_Y[:5])
train_data = model.ops.multibatch(batch_size,
train_X,
train_Y,
shuffle=True)
dev_data = model.ops.multibatch(batch_size, dev_X, dev_Y)
# Create the optimizer.
optimizer = Adam(0.001)
for i in range(n_iter):
for X, Y in tqdm(train_data, leave=False):
Yh, backprop = model.begin_update(X)
backprop(Yh - Y)
model.finish_update(optimizer)
# Evaluate and print progress
correct = 0
total = 0
for X, Y in dev_data:
Yh = model.predict(X)
correct += (Yh.argmax(axis=1) == Y.argmax(axis=1)).sum()
total += Yh.shape[0]
score = correct / total
msg.row((i, f"{score:.3f}"), widths=(3, 5))
def test_model_gpu():
prefer_gpu()
n_hidden = 32
dropout = 0.2
(train_X, train_Y), (dev_X, dev_Y) = ml_datasets.mnist()
model = chain(
Relu(nO=n_hidden, dropout=dropout),
Relu(nO=n_hidden, dropout=dropout),
Softmax(),
)
# making sure the data is on the right device
train_X = model.ops.asarray(train_X)
train_Y = model.ops.asarray(train_Y)
dev_X = model.ops.asarray(dev_X)
dev_Y = model.ops.asarray(dev_Y)
model.initialize(X=train_X[:5], Y=train_Y[:5])
optimizer = Adam(0.001)
batch_size = 128
for i in range(2):
batches = model.ops.multibatch(batch_size, train_X, train_Y, shuffle=True)
for X, Y in batches:
Yh, backprop = model.begin_update(X)
backprop(Yh - Y)
model.finish_update(optimizer)
# Evaluate and print progress
correct = 0
total = 0
for X, Y in model.ops.multibatch(batch_size, dev_X, dev_Y):
Yh = model.predict(X)
correct += (Yh.argmax(axis=1) == Y.argmax(axis=1)).sum()
total += Yh.shape[0]
def TransformersTagger(
starter: str, n_tags: int = 17
) -> Model[List[List[str]], List[Floats2d]]:
return chain(
TransformersTokenizer(starter),
Transformer(starter),
with_array(Softmax(nO=n_tags)),
)
def test_validation():
model = chain(Relu(10), Relu(10), with_ragged(reduce_max()), Softmax())
with pytest.raises(DataValidationError):
model.initialize(X=model.ops.alloc2f(1, 10), Y=model.ops.alloc2f(1, 10))
with pytest.raises(DataValidationError):
model.initialize(X=model.ops.alloc3f(1, 10, 1), Y=model.ops.alloc2f(1, 10))
with pytest.raises(DataValidationError):
model.initialize(X=[model.ops.alloc2f(1, 10)], Y=model.ops.alloc2f(1, 10))
def build_tagger_model(
tok2vec: Model[List[Doc], List[Floats2d]],
nO: Optional[int] = None) -> Model[List[Doc], List[Floats2d]]:
"""Build a tagger model, using a provided token-to-vector component. The tagger
model simply adds a linear layer with softmax activation to predict scores
given the token vectors.
tok2vec (Model[List[Doc], List[Floats2d]]): The token-to-vector subnetwork.
nO (int or None): The number of tags to output. Inferred from the data if None.
"""
# TODO: glorot_uniform_init seems to work a bit better than zero_init here?!
t2v_width = tok2vec.get_dim("nO") if tok2vec.has_dim("nO") else None
output_layer = Softmax(nO, t2v_width, init_W=zero_init)
softmax = with_array(output_layer) # type: ignore
model = chain(tok2vec, softmax)
model.set_ref("tok2vec", tok2vec)
model.set_ref("softmax", output_layer)
model.set_ref("output_layer", output_layer)
return model
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].
"""
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)
model.set_dim(
"nO", nO
) # type: ignore # TODO: remove type ignore once Thinc has been updated
model.attrs["multi_label"] = not exclusive_classes
return model
def build_multi_task_model(
tok2vec: Model,
maxout_pieces: int,
token_vector_width: int,
nO: Optional[int] = None,
) -> Model:
softmax = Softmax(nO=nO, nI=token_vector_width * 2)
model = chain(
tok2vec,
Maxout(
nO=token_vector_width * 2,
nI=token_vector_width,
nP=maxout_pieces,
dropout=0.0,
),
LayerNorm(token_vector_width * 2),
softmax,
)
model.set_ref("tok2vec", tok2vec)
model.set_ref("output_layer", softmax)
return model
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].
"""
with Model.define_operators({">>": chain}):
cnn = tok2vec >> list2ragged() >> reduce_mean()
if exclusive_classes:
output_layer = Softmax(nO=nO, nI=tok2vec.maybe_get_dim("nO"))
model = cnn >> output_layer
model.set_ref("output_layer", output_layer)
else:
linear_layer = Linear(nO=nO, nI=tok2vec.maybe_get_dim("nO"))
model = cnn >> linear_layer >> Logistic()
model.set_ref("output_layer", linear_layer)
model.set_ref("tok2vec", tok2vec)
model.set_dim("nO", nO)
model.attrs["multi_label"] = not exclusive_classes
return model
def create_relu_softmax(width, dropout, nI, nO):
return chain(clone(ReLu(nO=width, dropout=dropout), 2), Softmax(10, width))
def create_embed_relu_relu_softmax(depth, width, vector_length):
with Model.define_operators({">>": chain}):
model = strings2arrays() >> with_array(
HashEmbed(width, vector_length) >> expand_window(window_size=1) >>
ReLu(width, width * 3) >> ReLu(width, width) >> Softmax(17, width))
return model
from thinc.api import chain, ReLu, reduce_max, Softmax, add bad_model = chain(ReLu(10), reduce_max(), Softmax()) bad_model2 = add(ReLu(10), reduce_max(), Softmax())
from thinc.api import chain, Relu, reduce_max, Softmax, add good_model = chain(Relu(10), Relu(10), Softmax()) reveal_type(good_model) good_model2 = add(Relu(10), Relu(10), Softmax()) reveal_type(good_model2) bad_model_undetected = chain(Relu(10), Relu(10), reduce_max(), Softmax()) reveal_type(bad_model_undetected) bad_model_undetected2 = add(Relu(10), Relu(10), reduce_max(), Softmax()) reveal_type(bad_model_undetected2)
from typing import Any, TypeVar
from thinc.api import chain, ReLu, reduce_max, Softmax, add, Model
good_model = chain(ReLu(10), ReLu(10), Softmax())
reveal_type(good_model)
good_model2 = add(ReLu(10), ReLu(10), Softmax())
reveal_type(good_model2)
bad_model_undetected = chain(ReLu(10), ReLu(10), ReLu(10), ReLu(10), Softmax())
reveal_type(bad_model_undetected)
bad_model_undetected2 = add(ReLu(10), ReLu(10), ReLu(10), ReLu(10), Softmax())
reveal_type(bad_model_undetected2)
def forward() -> None:
pass
OtherType = TypeVar("OtherType")
def other_function(layer1: Model, layer2: Model,
*layers: Model) -> Model[Any, OtherType]:
return Model("some_model", forward)
non_combinator_model = other_function(Model("x", forward), Model("y", forward),
Model("z", forward))
def TextCatEnsemble_v1(
width: int,
embed_size: int,
pretrained_vectors: Optional[bool],
exclusive_classes: bool,
ngram_size: int,
window_size: int,
conv_depth: int,
dropout: Optional[float],
nO: Optional[int] = None,
) -> Model:
# Don't document this yet, I'm not sure it's right.
cols = [ORTH, LOWER, PREFIX, SUFFIX, SHAPE, ID]
with Model.define_operators({">>": chain, "|": concatenate, "**": clone}):
lower = HashEmbed(nO=width,
nV=embed_size,
column=cols.index(LOWER),
dropout=dropout,
seed=10)
prefix = HashEmbed(
nO=width // 2,
nV=embed_size,
column=cols.index(PREFIX),
dropout=dropout,
seed=11,
)
suffix = HashEmbed(
nO=width // 2,
nV=embed_size,
column=cols.index(SUFFIX),
dropout=dropout,
seed=12,
)
shape = HashEmbed(
nO=width // 2,
nV=embed_size,
column=cols.index(SHAPE),
dropout=dropout,
seed=13,
)
width_nI = sum(
layer.get_dim("nO") for layer in [lower, prefix, suffix, shape])
trained_vectors = FeatureExtractor(cols) >> with_array(
uniqued(
(lower | prefix | suffix | shape) >> Maxout(
nO=width, nI=width_nI, normalize=True),
column=cols.index(ORTH),
))
if pretrained_vectors:
static_vectors = StaticVectors(width)
vector_layer = trained_vectors | static_vectors
vectors_width = width * 2
else:
vector_layer = trained_vectors
vectors_width = width
tok2vec = vector_layer >> with_array(
Maxout(width, vectors_width, normalize=True) >>
residual((expand_window(window_size=window_size) >> Maxout(
nO=width, nI=width *
((window_size * 2) + 1), normalize=True)))**conv_depth,
pad=conv_depth,
)
cnn_model = (tok2vec >> list2ragged() >> ParametricAttention(width) >>
reduce_sum() >> residual(Maxout(nO=width, nI=width)) >>
Linear(nO=nO, nI=width) >> Dropout(0.0))
linear_model = build_bow_text_classifier(
nO=nO,
ngram_size=ngram_size,
exclusive_classes=exclusive_classes,
no_output_layer=False,
)
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) >> Dropout(0.0) >> Logistic()
model = (linear_model | cnn_model) >> output_layer
model.set_ref("tok2vec", tok2vec)
if model.has_dim("nO") is not False:
model.set_dim("nO", nO)
model.set_ref("output_layer", linear_model.get_ref("output_layer"))
model.attrs["multi_label"] = not exclusive_classes
return model