def Tok2Vec(width, embed_size, **kwargs):
pretrained_vectors = kwargs.get("pretrained_vectors", None)
cnn_maxout_pieces = kwargs.get("cnn_maxout_pieces", 3)
subword_features = kwargs.get("subword_features", True)
conv_depth = kwargs.get("conv_depth", 4)
bilstm_depth = kwargs.get("bilstm_depth", 0)
cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
with Model.define_operators(
{">>": chain, "|": concatenate, "**": clone, "+": add, "*": reapply}
):
norm = HashEmbed(width, embed_size, column=cols.index(NORM), name="embed_norm")
if subword_features:
prefix = HashEmbed(
width, embed_size // 2, column=cols.index(PREFIX), name="embed_prefix"
)
suffix = HashEmbed(
width, embed_size // 2, column=cols.index(SUFFIX), name="embed_suffix"
)
shape = HashEmbed(
width, embed_size // 2, column=cols.index(SHAPE), name="embed_shape"
)
else:
prefix, suffix, shape = (None, None, None)
if pretrained_vectors is not None:
glove = StaticVectors(pretrained_vectors, width, column=cols.index(ID))
if subword_features:
embed = uniqued(
(glove | norm | prefix | suffix | shape)
>> LN(Maxout(width, width * 5, pieces=3)),
column=cols.index(ORTH),
)
else:
embed = uniqued(
(glove | norm) >> LN(Maxout(width, width * 2, pieces=3)),
column=cols.index(ORTH),
)
elif subword_features:
embed = uniqued(
(norm | prefix | suffix | shape)
>> LN(Maxout(width, width * 4, pieces=3)),
column=cols.index(ORTH),
)
else:
embed = norm
convolution = Residual(
ExtractWindow(nW=1)
>> LN(Maxout(width, width * 3, pieces=cnn_maxout_pieces))
)
tok2vec = FeatureExtracter(cols) >> with_flatten(
embed >> convolution ** conv_depth, pad=conv_depth
)
if bilstm_depth >= 1:
tok2vec = tok2vec >> PyTorchBiLSTM(width, width, bilstm_depth)
# Work around thinc API limitations :(. TODO: Revise in Thinc 7
tok2vec.nO = width
tok2vec.embed = embed
return tok2vec
def Tok2Vec(width, embed_size, **kwargs):
pretrained_vectors = kwargs.get("pretrained_vectors", None)
cnn_maxout_pieces = kwargs.get("cnn_maxout_pieces", 3)
subword_features = kwargs.get("subword_features", True)
conv_depth = kwargs.get("conv_depth", 4)
bilstm_depth = kwargs.get("bilstm_depth", 0)
cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
with Model.define_operators(
{">>": chain, "|": concatenate, "**": clone, "+": add, "*": reapply}
):
norm = HashEmbed(width, embed_size, column=cols.index(NORM), name="embed_norm")
if subword_features:
prefix = HashEmbed(
width, embed_size // 2, column=cols.index(PREFIX), name="embed_prefix"
)
suffix = HashEmbed(
width, embed_size // 2, column=cols.index(SUFFIX), name="embed_suffix"
)
shape = HashEmbed(
width, embed_size // 2, column=cols.index(SHAPE), name="embed_shape"
)
else:
prefix, suffix, shape = (None, None, None)
if pretrained_vectors is not None:
glove = StaticVectors(pretrained_vectors, width, column=cols.index(ID))
if subword_features:
embed = uniqued(
(glove | norm | prefix | suffix | shape)
>> LN(Maxout(width, width * 5, pieces=3)),
column=cols.index(ORTH),
)
else:
embed = uniqued(
(glove | norm) >> LN(Maxout(width, width * 2, pieces=3)),
column=cols.index(ORTH),
)
elif subword_features:
embed = uniqued(
(norm | prefix | suffix | shape)
>> LN(Maxout(width, width * 4, pieces=3)),
column=cols.index(ORTH),
)
else:
embed = norm
convolution = Residual(
ExtractWindow(nW=1)
>> LN(Maxout(width, width * 3, pieces=cnn_maxout_pieces))
)
tok2vec = FeatureExtracter(cols) >> with_flatten(
embed >> convolution ** conv_depth, pad=conv_depth
)
if bilstm_depth >= 1:
tok2vec = tok2vec >> PyTorchBiLSTM(width, width, bilstm_depth)
# Work around thinc API limitations :(. TODO: Revise in Thinc 7
tok2vec.nO = width
tok2vec.embed = embed
return tok2vec
def my_tok_to_vec(width, embed_size, pretrained_vectors, **kwargs):
# Circular imports :(
from spacy._ml import PyTorchBiLSTM
cnn_maxout_pieces = kwargs.get("cnn_maxout_pieces", 3)
conv_depth = kwargs.get("conv_depth", 4)
bilstm_depth = kwargs.get("bilstm_depth", 0)
cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
storage = []
with Model.define_operators({">>": chain, "|": concatenate, "**": clone}):
# norm = HashEmbed(width, embed_size, column=cols.index(NORM), name="embed_norm")
# prefix = HashEmbed(
# width, embed_size // 2, column=cols.index(PREFIX), name="embed_prefix"
# )
# suffix = HashEmbed(
# width, embed_size // 2, column=cols.index(SUFFIX), name="embed_suffix"
# )
shape = HashEmbed(
width, embed_size // 2, column=cols.index(SHAPE), name="embed_shape"
)
glove = Vectors(storage, pretrained_vectors, width, column=cols.index(NORM), )
vec_width = glove.nV
embed = uniqued(
(glove | shape)
>> LN(Maxout(width, width + vec_width, pieces=3)),
column=cols.index(ORTH),
)
convolution = Residual(
ExtractWindow(nW=1)
>> LN(Maxout(width, width * 3, pieces=cnn_maxout_pieces))
)
tok2vec = SaveDoc(storage) >> FeatureExtracter(cols) >> with_flatten(
embed >> convolution ** conv_depth, pad=conv_depth
)
if bilstm_depth >= 1:
tok2vec = tok2vec >> PyTorchBiLSTM(width, width, bilstm_depth)
# Work around thinc API limitations :(. TODO: Revise in Thinc 7
tok2vec.nO = width
tok2vec.embed = embed
return tok2vec
def Doc2Feats(config):
columns = config["columns"]
return FeatureExtracter(columns)
def build_text_classifier(nr_class, width=64, **cfg):
depth = cfg.get("depth", 2)
nr_vector = cfg.get("nr_vector", 5000)
pretrained_dims = cfg.get("pretrained_dims", 0)
with Model.define_operators({
">>": chain,
"+": add,
"|": concatenate,
"**": clone
}):
if cfg.get("low_data") and pretrained_dims:
model = (SpacyVectors >> flatten_add_lengths >> with_getitem(
0, Affine(width, pretrained_dims)) >>
ParametricAttention(width) >> Pooling(sum_pool) >>
Residual(ReLu(width, width))**2 >> zero_init(
Affine(nr_class, width, drop_factor=0.0)) >> logistic)
return model
lower = HashEmbed(width, nr_vector, column=1)
prefix = HashEmbed(width // 2, nr_vector, column=2)
suffix = HashEmbed(width // 2, nr_vector, column=3)
shape = HashEmbed(width // 2, nr_vector, column=4)
trained_vectors = FeatureExtracter(
[ORTH, LOWER, PREFIX, SUFFIX, SHAPE, ID]) >> with_flatten(
uniqued(
(lower | prefix | suffix | shape) >> LN(
Maxout(width, width + (width // 2) * 3)),
column=0,
))
if pretrained_dims:
static_vectors = SpacyVectors >> with_flatten(
Affine(width, pretrained_dims))
# TODO Make concatenate support lists
vectors = concatenate_lists(trained_vectors, static_vectors)
vectors_width = width * 2
else:
vectors = trained_vectors
vectors_width = width
static_vectors = None
tok2vec = vectors >> with_flatten(
LN(Maxout(width, vectors_width)) >> Residual(
(ExtractWindow(nW=1) >> LN(Maxout(width, width * 3))))**depth,
pad=depth,
)
cnn_model = (
tok2vec >> flatten_add_lengths >> ParametricAttention(width) >>
Pooling(sum_pool) >> Residual(zero_init(Maxout(width, width))) >>
zero_init(Affine(nr_class, width, drop_factor=0.0)))
linear_model = build_bow_text_classifier(nr_class,
ngram_size=cfg.get(
"ngram_size", 1),
exclusive_classes=False)
if cfg.get("exclusive_classes"):
output_layer = Softmax(nr_class, nr_class * 2)
else:
output_layer = (zero_init(
Affine(nr_class, nr_class * 2, drop_factor=0.0)) >> logistic)
model = (linear_model | cnn_model) >> output_layer
model.tok2vec = chain(tok2vec, flatten)
model.nO = nr_class
model.lsuv = False
return model
def main(nH=6,
dropout=0.1,
nS=6,
nB=64,
nE=20,
use_gpu=-1,
lim=1000000,
nM=300,
mL=100,
save=False,
nTGT=5000,
save_name="model.pkl"):
if use_gpu != -1:
spacy.require_gpu()
device = 'cuda'
else:
device = 'cpu'
''' Read dataset '''
nlp = spacy.load('en_core_web_sm')
print('English model loaded')
for control_token in ("<eos>", "<bos>", "<pad>", "<cls>", "<mask>"):
nlp.tokenizer.add_special_case(control_token, [{ORTH: control_token}])
train, dev, test = get_iwslt()
print('Dataset loaded')
train, _ = zip(*train)
dev, _ = zip(*dev)
test, _ = zip(*test)
train = train[:lim]
dev = dev[:lim]
test = test[:lim]
''' Tokenize '''
train = spacy_tokenize(nlp.tokenizer, train, mL=mL)
dev = spacy_tokenize(nlp.tokenizer, dev, mL=mL)
test = spacy_tokenize(nlp.tokenizer, test, mL=mL)
print('Tokenization finished')
''' Set rank based on all the docs '''
all_docs = train + dev + test
set_rank(nlp.vocab, all_docs, nTGT=nTGT)
train = set_numeric_ids(nlp.vocab, train)
dev = set_numeric_ids(nlp.vocab, dev)
test = set_numeric_ids(nlp.vocab, test)
print('Numeric ids set')
word2indx, indx2word = get_dicts(nlp.vocab)
print('Vocab dictionaries grabbed')
with Model.define_operators({">>": chain}):
embed_cols = [ORTH, SHAPE, PREFIX, SUFFIX]
extractor = FeatureExtracter(attrs=embed_cols)
position_encode = PositionEncode(mL, nM)
model = (FeatureExtracter(attrs=embed_cols) >> with_flatten(
FancyEmbed(nM, nTGT, cols=embed_cols)) >> Residual(position_encode)
>> create_model_input() >> Encoder(
nM=nM, nS=nS, nH=nH, device=device) >> with_reshape(
Softmax(nO=nTGT, nI=nM)))
''' Progress tracking '''
losses = [0.]
train_accuracies = [0.]
train_totals = [0.]
dev_accuracies = [0.]
dev_loss = [0.]
def track_progress():
correct = 0.
total = 0.
''' Get dev stats '''
for X0 in minibatch(dev, size=nB):
X1, loss_mask = random_mask(X0, nlp, indx2word, nlp.vocab, mL)
Xh = model(X1)
L, C, t = get_loss(Xh, X0, X1, loss_mask)
correct += C
total += t
dev_loss[-1] += (L**2).sum()
dev_accuracies[-1] = correct / total
print(len(losses), losses[-1],
train_accuracies[-1] / train_totals[-1], dev_loss[-1],
dev_accuracies[-1])
dev_loss.append(0.)
losses.append(0.)
train_accuracies.append(0.)
dev_accuracies.append(0.)
train_totals.append(0.)
if save:
model.to_disk('.models/' + save_name)
''' Model training '''
with model.begin_training(batch_size=nB,
nb_epoch=nE) as (trainer, optimizer):
trainer.dropout = dropout
trainer.dropout_decay = 1e-4
optimizer.alpha = 0.001
optimizer.L2 = 1e-6
optimizer.max_grad_norm = 1.0
trainer.each_epoch.append(track_progress)
optimizer.alpha = 0.001
optimizer.L2 = 1e-6
optimizer.max_grad_norm = 1.0
for X0, _ in trainer.iterate(train, train):
X1, loss_mask = random_mask(X0, nlp, indx2word, nlp.vocab, mL)
Xh, backprop = model.begin_update(X1, drop=dropout)
dXh, C, total = get_loss(Xh, X0, X1, loss_mask)
backprop(dXh, sgd=optimizer)
losses[-1] += (dXh**2).sum()
train_accuracies[-1] += C
train_totals[-1] += total
def main(nH=6, dropout=0.1, nS=6, nB=15, nE=20, use_gpu=-1, lim=2000):
if use_gpu != -1:
# TODO: Make specific to different devices, e.g. 1 vs 0
spacy.require_gpu()
train, dev, test = get_iwslt()
train_X, train_Y = zip(*train)
dev_X, dev_Y = zip(*dev)
test_X, test_Y = zip(*test)
''' Read dataset '''
nlp_en = spacy.load('en_core_web_sm')
nlp_de = spacy.load('de_core_news_sm')
print('Models loaded')
for control_token in ("<eos>", "<bos>", "<pad>"):
nlp_en.tokenizer.add_special_case(control_token, [{
ORTH: control_token
}])
nlp_de.tokenizer.add_special_case(control_token, [{
ORTH: control_token
}])
train_X, train_Y = spacy_tokenize(nlp_en.tokenizer, nlp_de.tokenizer,
train_X[-lim:], train_Y[-lim:],
MAX_LENGTH)
dev_X, dev_Y = spacy_tokenize(nlp_en.tokenizer, nlp_de.tokenizer,
dev_X[-lim:], dev_Y[-lim:], MAX_LENGTH)
test_X, test_Y = spacy_tokenize(nlp_en.tokenizer, nlp_de.tokenizer,
test_X[-lim:], test_Y[-lim:], MAX_LENGTH)
train_X = set_numeric_ids(nlp_en.vocab, train_X, vocab_size=VOCAB_SIZE)
train_Y = set_numeric_ids(nlp_de.vocab, train_Y, vocab_size=VOCAB_SIZE)
nTGT = VOCAB_SIZE
with Model.define_operators({">>": chain}):
embed_cols = [ORTH, SHAPE, PREFIX, SUFFIX]
extractor = FeatureExtracter(attrs=embed_cols)
position_encode = PositionEncode(MAX_LENGTH, MODEL_SIZE)
model = (apply_layers(extractor, extractor) >> apply_layers(
with_flatten(FancyEmbed(MODEL_SIZE, 5000, cols=embed_cols)),
with_flatten(FancyEmbed(MODEL_SIZE, 5000, cols=embed_cols)),
) >> apply_layers(Residual(position_encode), Residual(position_encode))
>> create_batch() >> EncoderDecoder(nS=nS, nH=nH, nTGT=nTGT))
losses = [0.]
train_accuracies = [0.]
train_totals = [0.]
dev_accuracies = [0.]
dev_loss = [0.]
def track_progress():
correct = 0.
total = 0.
for batch in minibatch(zip(dev_X, dev_Y), size=1024):
X, Y = zip(*batch)
Yh, Y_mask = model((X, Y))
L, C = get_loss(model.ops, Yh, Y, Y_mask)
correct += C
dev_loss[-1] += (L**2).sum()
total += len(Y)
dev_accuracies[-1] = correct / total
n_train = train_totals[-1]
print(len(losses), losses[-1], train_accuracies[-1] / n_train,
dev_loss[-1], dev_accuracies[-1])
dev_loss.append(0.)
losses.append(0.)
train_accuracies.append(0.)
dev_accuracies.append(0.)
train_totals.append(0.)
with model.begin_training(batch_size=nB,
nb_epoch=nE) as (trainer, optimizer):
trainer.dropout = dropout
trainer.dropout_decay = 1e-4
trainer.each_epoch.append(track_progress)
optimizer.alpha = 0.001
optimizer.L2 = 1e-6
optimizer.max_grad_norm = 1.0
for X, Y in trainer.iterate(train_X, train_Y):
(Yh, X_mask), backprop = model.begin_update((X, Y), drop=dropout)
dYh, C = get_loss(model.ops, Yh, Y, X_mask)
backprop(dYh, sgd=optimizer)
losses[-1] += (dYh**2).sum()
train_accuracies[-1] += C
train_totals[-1] += sum(len(y) for y in Y)
def main(use_gpu=False, nb_epoch=100):
if use_gpu:
Model.ops = CupyOps()
Model.Ops = CupyOps
train, test = datasets.imdb(limit=2000)
print("Load data")
train_X, train_y = zip(*train)
test_X, test_y = zip(*test)
train_y = Model.ops.asarray(to_categorical(train_y, nb_classes=2))
test_y = Model.ops.asarray(to_categorical(test_y, nb_classes=2))
nlp = spacy.load("en_vectors_web_lg")
nlp.add_pipe(nlp.create_pipe("sentencizer"), first=True)
preprocessor = FeatureExtracter([ORTH, LOWER, PREFIX, SUFFIX, SHAPE, ID])
train_X = [
preprocessor(list(doc.sents)) for doc in tqdm.tqdm(nlp.pipe(train_X))
]
test_X = [
preprocessor(list(doc.sents)) for doc in tqdm.tqdm(nlp.pipe(test_X))
]
dev_X = train_X[-1000:]
dev_y = train_y[-1000:]
train_X = train_X[:-1000]
train_y = train_y[:-1000]
print("Parse data")
n_sent = sum([len(list(sents)) for sents in train_X])
print("%d sentences" % n_sent)
model = build_model(2,
width=128,
conv_depth=2,
depth=2,
train_X=train_X,
train_y=train_y)
with model.begin_training(train_X[:100],
train_y[:100]) as (trainer, optimizer):
epoch_loss = [0.0]
def report_progress():
with model.use_params(optimizer.averages):
print(
epoch_loss[-1],
epoch_var[-1],
model.evaluate(dev_X, dev_y),
trainer.dropout,
)
epoch_loss.append(0.0)
epoch_var.append(0.0)
trainer.each_epoch.append(report_progress)
batch_sizes = compounding(64, 64, 1.01)
trainer.dropout = 0.3
trainer.batch_size = int(next(batch_sizes))
trainer.dropout_decay = 0.0
trainer.nb_epoch = nb_epoch
# optimizer.alpha = 0.1
# optimizer.max_grad_norm = 10.0
# optimizer.b1 = 0.0
# optimizer.b2 = 0.0
epoch_var = [0.0]
for X, y in trainer.iterate(train_X, train_y):
yh, backprop = model.begin_update(X, drop=trainer.dropout)
losses = ((yh - y)**2.0).sum(axis=1) / y.shape[0]
epoch_var[-1] += losses.var()
loss = losses.mean()
backprop((yh - y) / yh.shape[0], optimizer)
epoch_loss[-1] += loss
trainer.batch_size = int(next(batch_sizes))
with model.use_params(optimizer.averages):
print("Avg dev.: %.3f" % model.evaluate(dev_X, dev_y))
def main(nH=6,
dropout=0.0,
nS=6,
nB=32,
nE=20,
use_gpu=-1,
lim=2000,
nM=300,
mL=100,
save=False,
save_name="model.pkl"):
if use_gpu != -1:
# TODO: Make specific to different devices, e.g. 1 vs 0
spacy.require_gpu()
device = 'cuda'
else:
device = 'cpu'
''' Read dataset '''
nlp = spacy.load('en_core_web_sm')
for control_token in ("<eos>", "<bos>", "<pad>", "<cls>"):
nlp.tokenizer.add_special_case(control_token, [{ORTH: control_token}])
train, dev = imdb(limit=lim)
print('Loaded imdb dataset')
train = train[:lim]
dev = dev[:lim]
train_X, train_Y = zip(*train)
dev_X, dev_Y = zip(*dev)
train_X = spacy_tokenize(nlp.tokenizer, train_X, mL=mL)
dev_X = spacy_tokenize(nlp.tokenizer, dev_X, mL=mL)
print('Tokenized dataset')
train_X = set_numeric_ids(nlp.vocab, train_X)
dev_X = set_numeric_ids(nlp.vocab, dev_X)
print('Numeric ids ready')
with Model.define_operators({">>": chain}):
embed_cols = [ORTH, SHAPE, PREFIX, SUFFIX]
extractor = FeatureExtracter(attrs=embed_cols)
position_encode = PositionEncode(mL, nM)
model = (FeatureExtracter(attrs=embed_cols) >> with_flatten(
FancyEmbed(nM, 5000, cols=embed_cols)) >> Residual(position_encode)
>> create_model_input() >> Categorizer(
nM=nM, nS=nS, nH=nH, device=device))
losses = [0.]
train_accuracies = [0.]
train_totals = [0.]
dev_accuracies = [0.]
dev_loss = [0.]
def track_progress():
correct = 0.
total = 0.
for batch in minibatch(zip(dev_X, dev_Y), size=1024):
X, Y = zip(*batch)
Yh = model(X)
L, C = get_loss(Yh, Y)
correct += C
dev_loss[-1] += (L**2).sum()
total += len(X)
dev_accuracies[-1] = correct / total
n_train = train_totals[-1]
print(len(losses), losses[-1], train_accuracies[-1] / n_train,
dev_loss[-1], dev_accuracies[-1])
dev_loss.append(0.)
losses.append(0.)
train_accuracies.append(0.)
dev_accuracies.append(0.)
train_totals.append(0.)
with model.begin_training(batch_size=nB,
nb_epoch=nE) as (trainer, optimizer):
trainer.dropout = dropout
trainer.dropout_decay = 1e-4
optimizer.alpha = 0.001
optimizer.L2 = 1e-6
optimizer.max_grad_norm = 1.0
trainer.each_epoch.append(track_progress)
optimizer.alpha = 0.001
optimizer.L2 = 1e-6
optimizer.max_grad_norm = 1.0
other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "textcat"]
for X, Y in trainer.iterate(train_X, train_Y):
Yh, backprop = model.begin_update(X)
dYh, C = get_loss(Yh, Y)
backprop(dYh, sgd=optimizer)
losses[-1] += (dYh**2).sum()
train_accuracies[-1] += C
train_totals[-1] += len(Y)
if save:
model.to_disk(save_name)
def main(nH=6,
dropout=0.0,
nS=6,
nB=32,
nE=20,
use_gpu=-1,
lim=2000,
nM=300,
mL=20,
nTGT=3500,
save=False,
load=False,
save_name="model.pkl",
load_name="model.pkl"):
if use_gpu != -1:
# TODO: Make specific to different devices, e.g. 1 vs 0
spacy.require_gpu()
device = 'cuda'
else:
device = 'cpu'
train, dev, test = get_iwslt()
train_X, train_Y = zip(*train)
dev_X, dev_Y = zip(*dev)
test_X, test_Y = zip(*test)
''' Read dataset '''
nlp_en = spacy.load('en_core_web_sm')
nlp_de = spacy.load('de_core_news_sm')
print('Models loaded')
for control_token in ("<eos>", "<bos>", "<pad>"):
nlp_en.tokenizer.add_special_case(control_token, [{
ORTH: control_token
}])
nlp_de.tokenizer.add_special_case(control_token, [{
ORTH: control_token
}])
train_lim = min(lim, len(train_X))
dev_lim = min(lim, len(dev_X))
test_lim = min(lim, len(test_X))
train_X, train_Y = spacy_tokenize(nlp_en.tokenizer, nlp_de.tokenizer,
train_X[:train_lim], train_Y[:train_lim],
mL)
dev_X, dev_Y = spacy_tokenize(nlp_en.tokenizer, nlp_de.tokenizer,
dev_X[:dev_lim], dev_Y[:dev_lim], mL)
test_X, test_Y = spacy_tokenize(nlp_en.tokenizer, nlp_de.tokenizer,
test_X[:test_lim], test_Y[:test_lim], mL)
all_X_docs = train_X + dev_X + test_X
all_y_docs = train_Y + dev_Y + test_Y
set_rank(nlp_en.vocab, all_X_docs, nTGT=nTGT)
set_rank(nlp_de.vocab, all_y_docs, nTGT=nTGT)
train_X = set_numeric_ids(nlp_en.vocab, train_X)
dev_X = set_numeric_ids(nlp_en.vocab, dev_X)
test_X = set_numeric_ids(nlp_en.vocab, test_X)
train_Y = set_numeric_ids(nlp_de.vocab, train_Y)
dev_Y = set_numeric_ids(nlp_de.vocab, dev_Y)
test_Y = set_numeric_ids(nlp_de.vocab, test_Y)
en_word2indx, en_indx2word = get_dicts(nlp_en.vocab)
de_word2indx, de_indx2word = get_dicts(nlp_de.vocab)
nTGT += 1
if not load:
with Model.define_operators({">>": chain}):
embed_cols = [ORTH, SHAPE, PREFIX, SUFFIX]
extractor = FeatureExtracter(attrs=embed_cols)
position_encode = PositionEncode(mL, nM)
model = (apply_layers(extractor, extractor) >> apply_layers(
with_flatten(FancyEmbed(nM, 5000, cols=embed_cols)),
with_flatten(FancyEmbed(nM, 5000, cols=embed_cols)),
) >> apply_layers(Residual(position_encode),
Residual(position_encode)) >> create_batch() >>
EncoderDecoder(
nS=nS, nH=nH, nTGT=nTGT, nM=nM, device=device))
else:
model = Model.from_disk(load_name)
losses = [0.]
train_accuracies = [0.]
train_totals = [0.]
dev_accuracies = [0.]
dev_loss = [0.]
def track_progress():
correct = 0.
total = 0.
for batch in minibatch(zip(dev_X, dev_Y), size=1024):
X, Y = zip(*batch)
Yh, Y_mask = model((X, Y))
L, C, total = get_loss(model.ops, Yh, Y, Y_mask)
correct += C
dev_loss[-1] += (L**2).sum()
dev_accuracies[-1] = correct / total
n_train = train_totals[-1]
print(len(losses), losses[-1], train_accuracies[-1] / n_train,
dev_loss[-1], dev_accuracies[-1])
dev_loss.append(0.)
losses.append(0.)
train_accuracies.append(0.)
dev_accuracies.append(0.)
train_totals.append(0.)
with model.begin_training(batch_size=nB,
nb_epoch=nE) as (trainer, optimizer):
trainer.dropout = dropout
trainer.dropout_decay = 1e-4
optimizer.alpha = 0.001
optimizer.L2 = 1e-6
optimizer.max_grad_norm = 1.0
trainer.each_epoch.append(track_progress)
optimizer.alpha = 0.001
optimizer.L2 = 1e-6
optimizer.max_grad_norm = 1.0
for X, Y in trainer.iterate(train_X, train_Y):
(Yh, X_mask), backprop = model.begin_update((X, Y))
dYh, C, total = get_loss(model.ops, Yh, Y, X_mask)
backprop(dYh, sgd=optimizer)
losses[-1] += (dYh**2).sum()
train_accuracies[-1] += C
train_totals[-1] += total
if save:
model.to_disk(save_name)
def Tok2Vec(width, embed_size, **kwargs):
# Circular imports :(
from .._ml import CharacterEmbed
from .._ml import PyTorchBiLSTM
pretrained_vectors = kwargs.get("pretrained_vectors", None)
cnn_maxout_pieces = kwargs.get("cnn_maxout_pieces", 3)
subword_features = kwargs.get("subword_features", True)
char_embed = kwargs.get("char_embed", False)
if char_embed:
subword_features = False
conv_depth = kwargs.get("conv_depth", 4)
bilstm_depth = kwargs.get("bilstm_depth", 0)
cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
with Model.define_operators({">>": chain, "|": concatenate, "**": clone}):
norm = HashEmbed(width,
embed_size,
column=cols.index(NORM),
name="embed_norm",
seed=6)
if subword_features:
prefix = HashEmbed(width,
embed_size // 2,
column=cols.index(PREFIX),
name="embed_prefix",
seed=7)
suffix = HashEmbed(width,
embed_size // 2,
column=cols.index(SUFFIX),
name="embed_suffix",
seed=8)
shape = HashEmbed(width,
embed_size // 2,
column=cols.index(SHAPE),
name="embed_shape",
seed=9)
else:
prefix, suffix, shape = (None, None, None)
if pretrained_vectors is not None:
glove = StaticVectors(pretrained_vectors,
width,
column=cols.index(ID))
if subword_features:
embed = uniqued(
(glove | norm | prefix | suffix | shape) >> LN(
Maxout(width, width * 5, pieces=3)),
column=cols.index(ORTH),
)
elif char_embed:
embed = concatenate_lists(
CharacterEmbed(nM=64, nC=8),
FeatureExtracter(cols) >> with_flatten(glove),
)
reduce_dimensions = LN(
Maxout(width, 64 * 8 + width, pieces=cnn_maxout_pieces))
else:
embed = uniqued(
(glove | norm) >> LN(Maxout(width, width * 2, pieces=3)),
column=cols.index(ORTH),
)
elif subword_features:
embed = uniqued(
(norm | prefix | suffix | shape) >> LN(
Maxout(width, width * 4, pieces=3)),
column=cols.index(ORTH),
)
elif char_embed:
embed = concatenate_lists(
CharacterEmbed(nM=64, nC=8),
FeatureExtracter(cols) >> with_flatten(norm),
)
reduce_dimensions = LN(
Maxout(width, 64 * 8 + width, pieces=cnn_maxout_pieces))
else:
embed = norm
convolution = Residual(
ExtractWindow(
nW=1) >> LN(Maxout(width, width *
3, pieces=cnn_maxout_pieces)))
if char_embed:
tok2vec = embed >> with_flatten(
reduce_dimensions >> convolution**conv_depth, pad=conv_depth)
else:
tok2vec = FeatureExtracter(cols) >> with_flatten(
embed >> convolution**conv_depth, pad=conv_depth)
if bilstm_depth >= 1:
tok2vec = tok2vec >> PyTorchBiLSTM(width, width, bilstm_depth)
# Work around thinc API limitations :(. TODO: Revise in Thinc 7
tok2vec.nO = width
tok2vec.embed = embed
return tok2vec
