def call(self, inputs, state):
# 注意这个wrapper改变了attention操作顺序，注意与原版的区别
if not isinstance(state, seq2seq.AttentionWrapperState):
raise TypeError(
"Expected state to be instance of AttentionWrapperState. "
"Received type %s instead." % type(state))
if self._is_multi:
previous_alignments = state.alignments
previous_alignment_history = state.alignment_history
else:
previous_alignments = [state.alignments]
previous_alignment_history = [state.alignment_history]
all_alignments = []
all_attentions = []
all_histories = []
all_attentions_state = []
for i, attention_mechanism in enumerate(self._attention_mechanisms):
if isinstance(self._cell, rnn.LSTMCell):
rnn_cell_state = state.cell_state.h
else:
rnn_cell_state = state.cell_state
# Some of code are changed based on https://github.com/bgshih/aster/issues/2
attention, alignments, attention_state = _compute_attention(
attention_mechanism, rnn_cell_state, previous_alignments[i],
self._attention_layers[i] if self._attention_layers else None)
alignment_history = previous_alignment_history[i].write(
state.time, alignments) if self._alignment_history else ()
all_alignments.append(alignments)
all_histories.append(alignment_history)
all_attentions.append(attention)
all_attentions_state.append(attention_state)
attention = array_ops.concat(all_attentions, 1)
cell_inputs = self._cell_input_fn(inputs, attention)
cell_output, next_cell_state = self._cell(cell_inputs,
state.cell_state)
next_state = seq2seq.AttentionWrapperState(
time=state.time + 1,
cell_state=next_cell_state,
attention=attention,
alignments=self._item_or_tuple(all_alignments),
alignment_history=self._item_or_tuple(all_histories),
attention_state=self._item_or_tuple(all_attentions_state))
if self._output_attention:
return attention, next_state
else:
return cell_output, next_state
def get_zero_state(self):
batch_size, dtype = tf.shape(self.inputs)[0], self.inputs.dtype
decoder_state0 = self.decoder.zero_state(batch_size, dtype)
alignments = self.bahdanau(
decoder_state0,
self.bahdanau.initial_alignments(batch_size, dtype))
expanded_alignments = tf.expand_dims(alignments, 1)
attention_mechanism_values = self.bahdanau.values
context = expanded_alignments @ attention_mechanism_values
context1 = tf.squeeze(context, [1])
t0 = tf.zeros([], dtype=tf.int32)
return seq2seq.AttentionWrapperState(cell_state=decoder_state0,
time=t0,
alignments=alignments,
alignment_history=(),
attention=context1)
def decoding_layer(dec_embed_input, embeddings, enc_output, enc_state,
vocab_size, text_length, summary_length, max_summary_length,
rnn_size, vocab_to_int, keep_prob, batch_size, num_layers):
for layer in range(num_layers):
with tf.variable_scope('decoder_{}'.format(layer)):
lstm = tf.nn.rnn_cell.LSTMCell(
rnn_size,
initializer=tf.random_uniform_initializer(-0.1, 0.1, seed=2))
dec_cell = tf.nn.rnn_cell.DropoutWrapper(lstm,
input_keep_prob=keep_prob)
#全连接层
output_layer = Dense(vocab_size,
kernel_initializer=tf.truncated_normal_initializer(
mean=0.0, stddev=0.1))
attn_mech = seq.BahdanauAttention(rnn_size,
enc_output,
text_length,
normalize=False,
name='BahdanauAttention')
dec_cell = seq.AttentionWrapper(cell=dec_cell,
attention_mechanism=attn_mech,
attention_layer_size=rnn_size)
# 引入注意力机制
initial_state = seq.AttentionWrapperState(
enc_state[0], _zero_state_tensors(rnn_size, batch_size, tf.float32))
with tf.variable_scope("decode"):
training_logits = training_decoding_layer(dec_embed_input,
summary_length, dec_cell,
initial_state, output_layer,
vocab_size,
max_summary_length)
with tf.variable_scope("decode", reuse=True):
inference_logits = inference_decoding_layer(
embeddings, vocab_to_int['<GO>'], vocab_to_int['<EOS>'], dec_cell,
initial_state, output_layer, max_summary_length, batch_size)
return training_logits, inference_logits
def call(self, inputs, state):
"""Perform a step of attention-wrapped RNN.
- Step 1: Mix the `inputs` and previous step's `attention` output via
`cell_input_fn`.
- Step 2: Call the wrapped `cell` with this input and its previous state.
- Step 3: Score the cell's output with `attention_mechanism`.
- Step 4: Calculate the alignments by passing the score through the
`normalizer`.
- Step 5: Calculate the context vector as the inner product between the
alignments and the attention_mechanism's values (memory).
- Step 6: Calculate the attention output by concatenating the cell output
and context through the attention layer (a linear layer with
`attention_layer_size` outputs).
Args:
inputs: (Possibly nested tuple of) Tensor, the input at this time step.
state: An instance of `AttentionWrapperState` containing
tensors from the previous time step.
Returns:
A tuple `(attention_or_cell_output, next_state)`, where:
- `attention_or_cell_output` depending on `output_attention`.
- `next_state` is an instance of `AttentionWrapperState`
containing the state calculated at this time step.
Raises:
TypeError: If `state` is not an instance of `AttentionWrapperState`.
"""
if not isinstance(state, tf.contrib.seq2seq.AttentionWrapperState):
raise TypeError(
"Expected state to be instance of AttentionWrapperState. "
"Received type %s instead." % type(state))
# Step 1: Calculate the true inputs to the cell based on the
# previous attention value.
cell_inputs = self._cell_input_fn(inputs, state.attention)
cell_state = state.cell_state
cell_output, next_cell_state = self._cell(cell_inputs, cell_state)
cell_batch_size = (cell_output.shape[0].value
or tf.shape(cell_output)[0])
error_message = (
"When applying AttentionWrapper %s: " % self.name +
"Non-matching batch sizes between the memory "
"(encoder output) and the query (decoder output). Are you using "
"the BeamSearchDecoder? You may need to tile your memory input via "
"the tf.contrib.seq2seq.tile_batch function with argument "
"multiple=beam_width.")
with tf.control_dependencies(
self._batch_size_checks(cell_batch_size, error_message)):
cell_output = tf.identity(cell_output, name="checked_cell_output")
if self._is_multi:
previous_alignments = state.alignments
previous_alignment_history = state.alignment_history
else:
previous_alignments = [state.alignments]
previous_alignment_history = [state.alignment_history]
all_alignments = []
all_attentions = []
all_histories = []
for i, attention_mechanism in enumerate(self._attention_mechanisms):
if self.coverage:
# if we use coverage mode, previous alignments is coverage vector
# alignment history stack has shape: decoder time * batch * atten_len
# convert it to coverage vector
previous_alignments[i] = tf.cond(
previous_alignment_history[i].size() > 0,
lambda: tf.reduce_sum(tf.transpose(
previous_alignment_history[i].stack(), [1, 2, 0]),
axis=2),
lambda: tf.zeros_like(previous_alignments[i]))
# debug
# previous_alignments[i] = tf.Print(previous_alignments[i],[previous_alignment_history[i].size(), tf.shape(previous_alignments[i]),previous_alignments[i]],message="atten wrapper:")
attention, alignments, _ = _compute_attention(
attention_mechanism, cell_output, previous_alignments[i],
self._attention_layers[i] if self._attention_layers else None)
alignment_history = previous_alignment_history[i].write(
state.time, alignments) if self._alignment_history else ()
all_alignments.append(alignments)
all_histories.append(alignment_history)
all_attentions.append(attention)
attention = tf.concat(all_attentions, 1)
next_state = seq2seq.AttentionWrapperState(
time=state.time + 1,
cell_state=next_cell_state,
attention=attention,
alignments=self._item_or_tuple(all_alignments),
alignment_history=self._item_or_tuple(all_histories),
attention_state=state.attention_state)
if self._output_attention:
return attention, next_state
else:
return cell_output, next_state
def call(self, inputs, state):
"""Perform a step of attention-wrapped RNN.
- Step 1: Mix the `inputs` and previous step's `attention` output via
`cell_input_fn`.
- Step 2: Call the wrapped `cell` with this input and its previous state.
- Step 3: Score the cell's output with `attention_mechanism`.
- Step 4: Calculate the alignments by passing the score through the
`normalizer`.
- Step 5: Calculate the context vector as the inner product between the
alignments and the attention_mechanism's values (memory).
- Step 6: Calculate the attention output by POOLING the cell output
and context through the attention layer (a linear layer with
`attention_layer_size` outputs).
"""
if not isinstance(state, seq2seq.AttentionWrapperState):
raise TypeError("Expected state to be instance of "
"AttentionWrapperState. "
"Received type %s instead." % type(state))
# Step 1: Calculate the true inputs to the cell based on the
# previous attention value.
cell_inputs = self._cell_input_fn(inputs, state.attention)
cell_state = state.cell_state
cell_output, next_cell_state = self._cell(cell_inputs, cell_state)
cell_batch_size = (cell_output.shape[0].value
or array_ops.shape(cell_output)[0])
error_message = (
"When applying AttentionWrapper %s: " % self.name +
"Non-matching batch sizes between the memory "
"(encoder output) and the query (decoder output). Are you using "
"the BeamSearchDecoder? You may need to tile your memory input via "
"the tf.contrib.seq2seq.tile_batch function with argument "
"multiple=beam_width.")
with ops.control_dependencies(
self._batch_size_checks(cell_batch_size, error_message)):
cell_output = array_ops.identity(cell_output,
name="checked_cell_output")
if self._is_multi:
previous_alignments = state.alignments
previous_alignment_history = state.alignment_history
else:
previous_alignments = [state.alignments]
previous_alignment_history = [state.alignment_history]
all_alignments = []
all_attentions = []
all_histories = []
for i, attention_mechanism in enumerate(self._attention_mechanisms):
attention, alignments = _compute_attention(
attention_mechanism,
cell_output,
previous_alignments[i],
self._attention_layers[i] if self._attention_layers else None,
reuse=i > 0)
alignment_history = previous_alignment_history[i].write(
state.time, alignments) if self._alignment_history else ()
all_alignments.append(alignments)
all_histories.append(alignment_history)
all_attentions.append(attention)
if self.pooling == 'avgpool':
attention = tf.reduce_mean(tf.stack(all_attentions, axis=1),
axis=1)
else:
raise ValueError('Unknown pooling method')
next_state = seq2seq.AttentionWrapperState(
time=state.time + 1,
cell_state=next_cell_state,
attention=attention,
alignments=self._item_or_tuple(all_alignments),
alignment_history=self._item_or_tuple(all_histories))
if self._output_attention:
return attention, next_state
else:
return cell_output, next_state
def _init_encoder(self):
with tf.variable_scope("Encoder") as scope:
encoder_inputs = self._maybe_add_dense_layers()
cell_type = self._hparams.cell_type[0 if self._data_type ==
'video' else 1]
batch_size = self._hparams.batch_size[0 if self._mode ==
'train' else 1]
if self._hparams.encoder_type == 'unidirectional':
self._encoder_cells, initial_state = build_rnn_layers(
cell_type=cell_type,
num_units_per_layer=self._num_units_per_layer,
use_dropout=self._hparams.use_dropout,
dropout_probability=self._dropout_probability,
mode=self._mode,
as_list=True,
batch_size=batch_size,
dtype=self._hparams.dtype)
#self._encoder_cells, initial_state = create_model(model=cell_type,
# num_cells=self._num_units_per_layer,
# batch_size=batch_size,
# as_list=False if cell_type == 'multi_skip_lstm' else True,
# learn_initial_state=True if 'skip' in cell_type else False,
# use_dropout=self._hparams.use_dropout,
# dropout_probability=self._dropout_probability)
self._encoder_cells = maybe_list(self._encoder_cells)
print(self._num_units_per_layer)
print('encoder_cells', self._encoder_cells)
print('AttentiveEncoder_initial_state', initial_state)
#### here weird code
# 1. reverse mem
# self._attended_memory = tf.reverse(self._attended_memory, axis=[1])
# 2. append zeros
# randval1 = tf.random.uniform(shape=[], minval=25, maxval=100, dtype=tf.int32)
# randval2 = tf.random.uniform(shape=[], minval=25, maxval=100, dtype=tf.int32)
# zeros_slice1 = tf.zeros([1, randval1, 256], dtype=tf.float32) # assuming we use inference on a batch size of 1
# zeros_slice2 = tf.zeros([1, randval2, 256], dtype=tf.float32)
# self._attended_memory = tf.concat([zeros_slice1, self._attended_memory, zeros_slice2], axis=1)
# self._attended_memory_length += randval1 + randval2
# 3. blank mem
# self._attended_memory = 0* self._attended_memory
# 4. mix with noise
# noise = tf.random.truncated_normal(shape=tf.shape(self._attended_memory))
# noise = tf.random.uniform(shape=tf.shape(self._attended_memory))
# self._attended_memory = noise
#### here stop weird code
if cell_type == 'skip_lstm' and self._hparams.separate_skip_rnn:
skip_cell = self._encoder_cells[0]
self._encoder_cells = self._encoder_cells[1:]
skip_out = tf.nn.dynamic_rnn(
cell=skip_cell,
inputs=encoder_inputs,
sequence_length=self._inputs_len,
parallel_iterations=batch_size,
swap_memory=False,
dtype=self._hparams.dtype,
scope=scope,
initial_state=skip_cell.trainable_initial_state(
batch_size),
)
print('skip_out', skip_out)
skip_output, skip_final_state = skip_out
h, updated_states = skip_output
print('skip_encoder_updated_states', updated_states)
cost_per_sample = self._hparams.cost_per_sample[1]
budget_loss = tf.reduce_mean(
tf.reduce_sum(cost_per_sample * updated_states, 1), 0)
meanUpdates = tf.reduce_mean(
tf.reduce_sum(updated_states, 1), 0)
self.skip_infos = SkipInfoTuple(updated_states,
meanUpdates, budget_loss)
# Tried to remove the skipped states in the output, but this destroys the input shape
#updated_states_shape = tf.shape(updated_states)
#h_shape = tf.shape(h)
#updated_states = tf.reshape(updated_states, [batch_size, updated_states_shape[1]])
#new_h = tf.boolean_mask(h, tf.where(updated_states == 1.))
#new_h = tf.where(updated_states == 1., h, tf.zeros(shape=h_shape))
#print('new_h', new_h)
#new_h_shape = tf.shape(new_h)
#new_h = tf.reshape(new_h, [batch_size, new_h_shape[0] / batch_size, new_h_shape[1]])
#self._inputs_len = (batch_size, new_h_shape[0] / batch_size)
print('skip_encoder_layer_h', h)
print('attended_memory', self._attended_memory)
print(self._encoder_cells)
attention_cells, dummy_initial_state = add_attention(
cell_type='lstm'
if self._hparams.separate_skip_rnn else cell_type,
cells=self._encoder_cells[-1],
attention_types=self._hparams.attention_type[0],
num_units=self._num_units_per_layer[-1],
memory=self._attended_memory,
memory_len=self._attended_memory_length,
mode=self._mode,
dtype=self._hparams.dtype,
initial_state=initial_state[-1] if
(isinstance(initial_state, tuple)
and not isinstance(initial_state, SkipLSTMStateTuple))
else initial_state,
batch_size=tf.shape(self._inputs_len),
write_attention_alignment=self._hparams.
write_attention_alignment,
fusion_type='linear_fusion',
)
print('AttentiveEncoder_initial_state2', initial_state)
if isinstance(initial_state, tuple) and not isinstance(
initial_state, SkipLSTMStateTuple):
initial_state = list(initial_state)
initial_state[-1] = dummy_initial_state
initial_state = tuple(initial_state)
else:
initial_state = dummy_initial_state
self._encoder_cells[-1] = attention_cells
# initial_state = self._encoder_cells.get_initial_state(inputs=None, batch_size=batch_size, dtype=self._hparams.dtype)
#initial_state = []
#for i, cell in enumerate(self._encoder_cells):
# print(i, cell)
# if isinstance(cell, SkipLSTMCell):
# #pass
# #with tf.variable_scope(f'layer_{i}') as init_scope:
# initial_state.append(cell.zero_state(batch_size, dtype=self._hparams.dtype))
# else:
# with tf.variable_scope(f'layer_{i}'):
# initial_state.append(
# cell.get_initial_state(inputs=None, batch_size=batch_size, dtype=self._hparams.dtype))
#initial_state = tuple(initial_state)
print('AttentiveEncoder_encoder_cells', self._encoder_cells)
self._encoder_cells = maybe_multirnn(self._encoder_cells)
print('AttentiveEncoder_encoder_cells', self._encoder_cells)
print('AttentiveEncoder_encoder_inputs', encoder_inputs)
print('AttentiveEncoder_inputs_len', self._inputs_len)
#initial_state = self._encoder_cells.get_initial_state(batch_size=batch_size, dtype=self._hparams.dtype)
#print('AttentiveEncoder_initial_state_final', initial_state)
out = tf.nn.dynamic_rnn(
cell=self._encoder_cells,
inputs=encoder_inputs
if not self._hparams.separate_skip_rnn else h,
sequence_length=self._inputs_len,
parallel_iterations=self._hparams.
batch_size[0 if self._mode == 'train' else 1],
swap_memory=False,
dtype=self._hparams.dtype,
scope=scope,
initial_state=initial_state,
)
self._encoder_outputs, self._encoder_final_state = out
print("AttentiveEncoder_dynamic_rnn_out",
self._encoder_outputs)
print("AttentiveEncoder_dynamic_rnn_fs",
self._encoder_final_state)
if not self._hparams.separate_skip_rnn and 'skip' in cell_type:
self._encoder_outputs, updated_states = self._encoder_outputs
cost_per_sample = self._hparams.cost_per_sample[1]
budget_loss = tf.reduce_mean(
tf.reduce_sum(cost_per_sample * updated_states, 1), 0)
meanUpdates = tf.reduce_mean(
tf.reduce_sum(updated_states, 1), 0)
self.skip_infos = SkipInfoTuple(updated_states,
meanUpdates, budget_loss)
if isinstance(self._encoder_final_state,
tuple) and not isinstance(
self._encoder_final_state,
seq2seq.AttentionWrapperState):
self._encoder_final_state = self._encoder_final_state[
-1]
print('AttentiveEncoder_final_state_inBetween',
self._encoder_final_state)
if isinstance(self._encoder_final_state,
seq2seq.AttentionWrapperState):
cell_state = self._encoder_final_state.cell_state
try:
cell_state = [
LSTMStateTuple(cs.c, cs.h) for cs in cell_state
]
except:
cell_state = LSTMStateTuple(
cell_state.c, cell_state.h)
self._encoder_final_state = seq2seq.AttentionWrapperState(
cell_state, self._encoder_final_state.attention,
self._encoder_final_state.time,
self._encoder_final_state.alignments,
self._encoder_final_state.alignment_history,
self._encoder_final_state.attention_state)
print('AttentiveEncoder_final_state',
self._encoder_final_state)
if self._hparams.write_attention_alignment is True:
# self.weights_summary = self._encoder_final_state[-1].attention_weight_history.stack()
self.attention_summary, self.attention_alignment = self._create_attention_alignments_summary(
maybe_list(self._encoder_final_state)[-1])