def __call__(self, x_t, hidden_t_prev, cell_t_prev):
return layers.lstm_unit(x_t=x_t,
hidden_t_prev=hidden_t_prev,
cell_t_prev=cell_t_prev,
forget_bias=forget_bias,
param_attr=self.attr_holder.param_attr,
bias_attr=self.attr_holder.bias_attr)
def test_lstm_unit(self):
program = Program()
with program_guard(program):
x_t_data = layers.data(
name='x_t_data', shape=[10, 10], dtype='float32')
x_t = layers.fc(input=x_t_data, size=10)
prev_hidden_data = layers.data(
name='prev_hidden_data', shape=[10, 30], dtype='float32')
prev_hidden = layers.fc(input=prev_hidden_data, size=30)
prev_cell_data = layers.data(
name='prev_cell', shape=[10, 30], dtype='float32')
prev_cell = layers.fc(input=prev_cell_data, size=30)
self.assertIsNotNone(
layers.lstm_unit(
x_t=x_t, hidden_t_prev=prev_hidden, cell_t_prev=prev_cell))
print(str(program))
def test_lstm_unit(self):
program = Program()
with program_guard(program):
x_t_data = layers.data(
name='x_t_data', shape=[10, 10], dtype='float32')
x_t = layers.fc(input=x_t_data, size=10)
prev_hidden_data = layers.data(
name='prev_hidden_data', shape=[10, 30], dtype='float32')
prev_hidden = layers.fc(input=prev_hidden_data, size=30)
prev_cell_data = layers.data(
name='prev_cell', shape=[10, 30], dtype='float32')
prev_cell = layers.fc(input=prev_cell_data, size=30)
self.assertIsNotNone(
layers.lstm_unit(
x_t=x_t, hidden_t_prev=prev_hidden, cell_t_prev=prev_cell))
print(str(program))
def call(self, global_img_feat, p_img_feat, embedding_fn, words=None):
# 图片特征
img_feat = layers.fc(p_img_feat, self.hid_size, num_flatten_dims=2, act='tanh') # [batch, k, hid]
img_feat_emb = layers.fc(p_img_feat, self.hid_size, num_flatten_dims=2)
if self.mode == 'eval':
word = layers.fill_constant_batch_size_like(global_img_feat, [-1],
dtype='int64',
value=config.data['start_idx'])
else:
words = layers.transpose(words, [1, 0]) # [seq, batch]
words.stop_gradient = True
# lstm 初始化
hid, cell = create_zero_state(global_img_feat), create_zero_state(global_img_feat)
# While loop 参数初始化
mx = decoder_config['sentence_length'] - 1 if self.mode == 'train' else decoder_config['infer_max_length']
if self.mode == 'eval':
mx = decoder_config['infer_max_length']
while_op_output = layers.create_array('int64')
else:
while_op_output = layers.create_array('float32')
max_step = layers.fill_constant(shape=[1], dtype='int64', value=mx)
step = layers.fill_constant(shape=[1], dtype='int64', value=0)
cond = layers.less_than(step, max_step)
while_op = layers.While(cond)
with while_op.block():
if self.mode == 'train':
st = layers.cast(step, 'int32')
word = layers.slice(words, axes=[0], starts=st, ends=st + 1)
word = layers.squeeze(word, [0])
word.stop_gradient = True
word_emb = embedding_fn(word)
# 这里可能用+效果更好？
xt = layers.concat([word_emb, global_img_feat], axis=-1) # [batch, feat]
h, c = layers.lstm_unit(xt, hid, cell, param_attr=fluid.ParamAttr('lstm_w'),
bias_attr=fluid.ParamAttr('lstm_b'))
p_word_emb = layers.fc(xt, size=self.hid_size)
p_hidden = layers.fc(hid, size=self.hid_size)
sentinel_gate = layers.sigmoid(p_word_emb + p_hidden) # [batch, hidden]
sentinel = layers.elementwise_mul(sentinel_gate, layers.tanh(c)) # [batch, hidden]
layers.assign(h, hid)
layers.assign(c, cell)
k = layers.shape(p_img_feat)[1]
p_hid = layers.fc(h, self.hid_size, act='tanh')
# attention 部分
# alpha
hid_emb = layers.fc(p_hid, self.hid_size) # [batch, hidden]
exp_hid_emb = layers.expand(layers.unsqueeze(hid_emb, 1), [1, k + 1, 1]) # [batch, k+1, hidden]
sentinel_emb = layers.unsqueeze(layers.fc(sentinel, self.hid_size), axes=1) # [batch, 1, hidden]
feat_emb = layers.concat([img_feat_emb, sentinel_emb], axis=1) # [batch, k+1, hidden]
z = layers.tanh(feat_emb + exp_hid_emb) # [batch, k+1, 1]
alpha = layers.fc(z, size=1, num_flatten_dims=2, act='softmax') # [batch, k+1, 1]
# context vector
context = layers.concat([img_feat, layers.unsqueeze(sentinel, axes=1)], axis=1) # [batch, k+1, hidden]
context = layers.elementwise_mul(context, alpha, axis=0)
context = layers.reduce_mean(context, dim=1) # [batch, hidden]
out = layers.fc(context + p_hid, self.hid_size, act='tanh')
word_pred = weight_tying_fc(out) # [batch, vocab]
if self.mode == 'eval':
next_word = layers.argmax(word_pred, axis=-1)
layers.assign(next_word, word)
next_word = layers.cast(next_word, 'float32')
layers.array_write(next_word, step, array=while_op_output)
else:
layers.array_write(word_pred, step, array=while_op_output)
layers.increment(step)
layers.less_than(step, max_step, cond=cond)
if self.mode == 'train':
output_time_major, _ = layers.tensor_array_to_tensor(while_op_output, axis=0, use_stack=True)
output = layers.transpose(output_time_major, [1, 0, 2])
else:
output_time_major = layers.tensor_array_to_tensor(while_op_output, axis=0, use_stack=True)[0]
output = layers.transpose(output_time_major, [1, 0])
return output
def test_pre_cell_type():
error_pre_cell = fluid.data(name='error_pre_cell',
shape=[batch_size, hidden_dim],
dtype='int32')
lstm_unit(inputs, pre_hidden, error_pre_cell)
def test_input_type():
error_input = fluid.data(name='error_input',
shape=[batch_size, emb_dim],
dtype='int32')
lstm_unit(error_input, pre_hidden, pre_cell)
def test_pre_cell_Variable():
lstm_unit(inputs, pre_hidden, np_pre_cell)
def test_input_Variable():
lstm_unit(np_input, pre_hidden, pre_cell)
