def build_model():
x = nn.Variable((batch_size, sentence_length_source))
mask = get_mask(x)
y = nn.Variable((batch_size, sentence_length_target))
enc_input = time_distributed(PF.embed)(
x, vocab_size_source, embedding_size, name='enc_embeddings') * mask
# -> (batch_size, sentence_length_source, embedding_size)
dec_input = F.concatenate(F.constant(w2i_target['<bos>'],
shape=(batch_size, 1)),
y[:, :sentence_length_target - 1],
axis=1)
dec_input = time_distributed(PF.embed)(dec_input,
vocab_size_target,
embedding_size,
name='dec_embeddings')
# -> (batch_size, sentence_length_target, embedding_size)
# encoder
with nn.parameter_scope('encoder'):
enc_output, c, h = lstm(enc_input,
hidden,
mask=mask,
return_sequences=True,
return_state=True)
# -> (batch_size, sentence_length_source, hidden), (batch_size, hidden), (batch_size, hidden)
# decoder
with nn.parameter_scope('decoder'):
dec_output = lstm(dec_input,
hidden,
initial_state=(c, h),
return_sequences=True)
# -> (batch_size, sentence_length_target, hidden)
attention_output = global_attention(dec_output,
enc_output,
mask=mask,
score='dot')
# -> (batch_size, sentence_length_target, hidden)
output = F.concatenate(dec_output, attention_output, axis=2)
output = time_distributed(PF.affine)(output,
vocab_size_target,
name='output')
# -> (batch_size, sentence_length_target, vocab_size_target)
t = F.reshape(y, (batch_size, sentence_length_target, 1))
entropy = time_distributed_softmax_cross_entropy(output, t)
mask = F.sum(F.sign(t), axis=2) # do not predict 'pad'.
count = F.sum(mask, axis=1)
entropy *= mask
loss = F.mean(F.sum(entropy, axis=1) / count)
return x, y, loss
def build_self_attention_model(train=True):
x = nn.Variable((batch_size, max_len))
t = nn.Variable((batch_size, 1))
mask = get_mask(x)
attention_mask = (F.constant(1, shape=mask.shape) - mask) * F.constant(
np.finfo(np.float32).min, shape=mask.shape)
with nn.parameter_scope('embedding'):
h = time_distributed(PF.embed)(x, vocab_size, embedding_size) * mask
with nn.parameter_scope('forward'):
h_f = lstm(h,
hidden_size,
mask=mask,
return_sequences=True,
return_state=False)
with nn.parameter_scope('backward'):
h_b = lstm(h[:, ::-1, ],
hidden_size,
mask=mask,
return_sequences=True,
return_state=False)[:, ::-1, ]
h = F.concatenate(h_f, h_b, axis=2)
if train:
h = F.dropout(h, p=dropout_ratio)
with nn.parameter_scope('da'):
a = F.tanh(time_distributed(PF.affine)(h, da))
if train:
a = F.dropout(a, p=dropout_ratio)
with nn.parameter_scope('r'):
a = time_distributed(PF.affine)(a, r)
if train:
a = F.dropout(a, p=dropout_ratio)
a = F.softmax(a + attention_mask, axis=1)
m = F.batch_matmul(a, h, transpose_a=True)
with nn.parameter_scope('output_mlp'):
output = F.relu(PF.affine(m, output_mlp_size))
if train:
output = F.dropout(output, p=dropout_ratio)
with nn.parameter_scope('output'):
y = F.sigmoid(PF.affine(output, 1))
accuracy = F.mean(F.equal(F.round(y), t))
loss = F.mean(F.binary_cross_entropy(
y, t)) + attention_penalty_coef * frobenius(
F.batch_matmul(a, a, transpose_a=True) - batch_eye(batch_size, r))
return x, t, accuracy, loss
def predict(x):
with nn.auto_forward():
x = x.reshape((1, sentence_length_source))
enc_input = nn.Variable.from_numpy_array(x)
mask = get_mask(enc_input)
enc_input = time_distributed(PF.embed)(enc_input,
vocab_size_source,
embedding_size,
name='enc_embeddings') * mask
# encoder
with nn.parameter_scope('encoder'):
enc_output, c, h = lstm(enc_input,
hidden,
mask=mask,
return_sequences=True,
return_state=True)
# decode
pad = nn.Variable.from_numpy_array(np.array([w2i_target['<bos>']]))
x = PF.embed(pad,
vocab_size_target,
embedding_size,
name='dec_embeddings')
_cell, _hidden = c, h
word_index = 0
ret = []
i = 0
while i2w_target[word_index] != '。' and i < 20:
with nn.parameter_scope('decoder'):
with nn.parameter_scope('lstm'):
_cell, _hidden = lstm_cell(x, _cell, _hidden)
q = F.reshape(_hidden, (1, 1, hidden))
attention_output = global_attention(q,
enc_output,
mask=mask,
score='dot')
attention_output = F.reshape(attention_output, (1, hidden))
output = F.concatenate(_hidden, attention_output, axis=1)
output = PF.affine(output, vocab_size_target, name='output')
word_index = np.argmax(output.d[0])
ret.append(word_index)
x = nn.Variable.from_numpy_array(
np.array([word_index], dtype=np.int32))
x = PF.embed(x,
vocab_size_target,
embedding_size,
name='dec_embeddings')
i += 1
return ret
def build_model(train=True, get_embeddings=False):
x = nn.Variable((batch_size, sentence_length, ptb_dataset.word_length))
mask = expand_dims(F.sign(x), axis=-1)
t = nn.Variable((batch_size, sentence_length))
with nn.parameter_scope('char_embedding'):
h = PF.embed(x, char_vocab_size, char_embedding_dim) * mask
h = F.transpose(h, (0, 3, 1, 2))
output = []
for f, f_size in zip(filters, filster_sizes):
_h = PF.convolution(h, f, kernel=(1, f_size), pad=(0, f_size//2), name='conv_{}'.format(f_size))
_h = F.max_pooling(_h, kernel=(1, ptb_dataset.word_length))
output.append(_h)
h = F.concatenate(*output, axis=1)
h = F.transpose(h, (0, 2, 1, 3))
mask = get_mask(F.sum(x, axis=2))
embeddings = F.reshape(h, (batch_size, sentence_length, sum(filters))) * mask
if get_embeddings:
return x, embeddings
with nn.parameter_scope('highway1'):
h = time_distributed(highway)(embeddings)
with nn.parameter_scope('highway2'):
h = time_distributed(highway)(h)
with nn.parameter_scope('lstm1'):
h = lstm(h, lstm_size, mask=mask, return_sequences=True)
with nn.parameter_scope('lstm2'):
h = lstm(h, lstm_size, mask=mask, return_sequences=True)
with nn.parameter_scope('hidden'):
h = F.relu(time_distributed(PF.affine)(h, lstm_size))
if train:
h = F.dropout(h, p=dropout_ratio)
with nn.parameter_scope('output'):
y = time_distributed(PF.affine)(h, word_vocab_size)
mask = F.sign(t) # do not predict 'pad'.
entropy = time_distributed_softmax_cross_entropy(y, expand_dims(t, axis=-1)) * mask
count = F.sum(mask, axis=1)
loss = F.mean(F.div2(F.sum(entropy, axis=1), count))
return x, t, loss
batch_size,
shuffle=True,
with_file_cache=False)
def global_average_pooling_1d(x, mask):
count = F.sum(mask, axis=1)
global_average_pooled = F.sum(h, axis=1) / count
return global_average_pooled
x = nn.Variable((batch_size, max_len))
t = nn.Variable((batch_size, 1))
mask = get_mask(x)
with nn.parameter_scope('embedding'):
h = time_distributed(PF.embed)(x, vocab_size, embedding_size) * mask
h = global_average_pooling_1d(h, mask)
with nn.parameter_scope('output'):
y = F.sigmoid(PF.affine(h, 1))
accuracy = F.mean(F.equal(F.round(y), t))
loss = F.mean(F.binary_cross_entropy(y, t))
# Create solver.
solver = S.Adam()
solver.set_parameters(nn.get_parameters())
trainer = Trainer(inputs=[x, t],
loss=loss,
metrics={
'cross entropy': loss,
batch_size,
shuffle=True,
with_file_cache=False)
x = nn.Variable((batch_size, sentence_length))
mask = get_mask(x)
t = nn.Variable((batch_size, sentence_length))
with nn.parameter_scope('embedding'):
h = PF.embed(x, vocab_size, embedding_size) * mask
with nn.parameter_scope('lstm1'):
h = lstm(h, hidden_size, mask=mask, return_sequences=True)
with nn.parameter_scope('lstm2'):
h = lstm(h, hidden_size, mask=mask, return_sequences=True)
with nn.parameter_scope('output'):
y = time_distributed(PF.affine)(h, vocab_size)
mask = F.sum(mask, axis=2) # do not predict 'pad'.
entropy = time_distributed_softmax_cross_entropy(y, expand_dims(
t, axis=-1)) * mask
# count = F.sum(mask, axis=1)
# loss = F.mean(F.div2(F.sum(entropy, axis=1), count))
loss = F.sum(entropy) / F.sum(mask)
# Create solver.
solver = S.Momentum(1e-2, momentum=0.9)
solver.set_parameters(nn.get_parameters())
trainer = Trainer(inputs=[x, t],
loss=loss,
metrics={'PPL': np.e**loss},
train_data_iter = data_iterator_simple(load_train_func,
len(x_train),
batch_size,
shuffle=True,
with_file_cache=False)
valid_data_iter = data_iterator_simple(load_valid_func,
len(x_valid),
batch_size,
shuffle=True,
with_file_cache=False)
x = nn.Variable((batch_size, sentence_length))
mask = get_mask(x)
t = nn.Variable((batch_size, sentence_length))
with nn.parameter_scope('embedding'):
h = time_distributed(PF.embed)(x, vocab_size, embedding_size) * mask
with nn.parameter_scope('rnn'):
h = simple_rnn(h, hidden_size, mask=mask, return_sequences=True)
with nn.parameter_scope('output'):
y = time_distributed(PF.affine)(h, vocab_size)
mask = F.sum(mask, axis=2)
entropy = time_distributed_softmax_cross_entropy(y, expand_dims(
t, axis=-1)) * mask
count = F.sum(mask, axis=1)
loss = F.mean(F.div2(F.sum(entropy, axis=1), count))
# Create solver.
solver = S.Momentum(1e-2, momentum=0.9)
solver.set_parameters(nn.get_parameters())
