def _preprocess(self, features):
"""Preprocesses features for multilingual translation."""
inputs = features["inputs"]
targets = features["targets"]
target_tags = features["target_tags"]
# Expand target tags to beam width, if necessary.
if self._hparams.mode == tf_estimator.ModeKeys.PREDICT:
# <float32> [batch_size * beam_width, 1, 1, emb_size].
beam_width = self._hparams.beam_width
target_tags = tf.tile(target_tags, [beam_width, 1, 1, 1])
# Add target tags to the input sequences.
# <float32> [batch_size, seq_len + 1, 1, emb_size].
inputs = tf.concat([target_tags, inputs], axis=1)
# Compute length of the input sequences.
inputs_length = common_layers.length_from_embedding(inputs)
inputs = common_layers.flatten4d3d(inputs)
# Preprocess targets.
targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right.
targets_length = common_layers.length_from_embedding(targets) + 1
targets = common_layers.flatten4d3d(targets)
return inputs, inputs_length, targets, targets_length
def lstm_seq2seq_internal(inputs, targets, hparams, train):
"""The basic LSTM seq2seq model, main step used for training."""
with tf.variable_scope("lstm_seq2seq"):
if inputs is not None:
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
inputs = tf.reverse_sequence(inputs, inputs_length, seq_axis=1)
_, final_encoder_state = lstm(inputs, inputs_length, hparams,
train, "encoder")
else:
final_encoder_state = None
# LSTM decoder.
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(
shifted_targets) + 1
decoder_outputs, _ = lstm(common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams,
train,
"decoder",
initial_state=final_encoder_state)
return tf.expand_dims(decoder_outputs, axis=2)
def lstm_seq2seq_internal_bid_encoder(inputs, targets, hparams, train):
"""The basic LSTM seq2seq model with bidirectional encoder."""
with tf.variable_scope("lstm_seq2seq_bid_encoder"):
if inputs is not None:
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
_, final_encoder_state = lstm_bid_encoder(
inputs, inputs_length, hparams, train, "encoder")
else:
inputs_length = None
final_encoder_state = None
# LSTM decoder.
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(shifted_targets) + 1
hparams_decoder = copy.copy(hparams)
hparams_decoder.hidden_size = 2 * hparams.hidden_size
decoder_outputs, _ = lstm(
common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams_decoder,
train,
"decoder",
initial_state=final_encoder_state)
return tf.expand_dims(decoder_outputs, axis=2)
def lstm_seq2seq_internal_attention(inputs, targets, hparams, train):
"""LSTM seq2seq model with attention, main step used for training."""
with tf.variable_scope("lstm_seq2seq_attention"):
# This is a temporary fix for varying-length sequences within in a batch.
# A more complete fix should pass a length tensor from outside so that
# all the lstm variants can use it.
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
encoder_outputs, final_encoder_state = lstm(
inputs, hparams, train, "encoder", sequence_length=inputs_length)
# LSTM decoder with attention
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(
shifted_targets) + 1
decoder_outputs, _ = lstm_attention_decoder(
common_layers.flatten4d3d(shifted_targets),
hparams,
train,
"decoder",
final_encoder_state,
encoder_outputs,
encoder_output_length=inputs_length,
decoder_input_length=targets_length)
return tf.expand_dims(decoder_outputs, axis=2)
def _build_inputs_and_targets(self,
from_seqs=None,
from_tags=None,
to_seqs=None,
to_tags=None):
"""Given from and to sequences and tags, construct inputs and targets."""
del from_tags # Unused.
if from_seqs is not None:
inputs = from_seqs
inputs_length = common_layers.length_from_embedding(inputs)
if to_tags is not None:
# Add to-tags to the inputs and adjust lengths.
# <float32> [batch_size, seq_len + 1, 1, emb_size].
inputs = tf.concat([to_tags, inputs], axis=1)
inputs_length = inputs_length + 1
inputs = common_layers.flatten4d3d(inputs)
else:
inputs = None
inputs_length = None
if to_seqs is not None:
# Shift to-sequences to form targets.
# <float32> [batch_size, seq_len, 1, emb_size].
targets = common_layers.shift_right(to_seqs)
# Add 1 to account for the padding added to the left from shift_right.
targets_length = common_layers.length_from_embedding(targets) + 1
targets = common_layers.flatten4d3d(targets)
else:
targets = None
targets_length = None
return (inputs, inputs_length), (targets, targets_length)
def lstm_seq2seq_internal_bid_encoder(inputs, targets, hparams, train):
"""The basic LSTM seq2seq model with bidirectional encoder."""
with tf.variable_scope("lstm_seq2seq_bid_encoder"):
if inputs is not None:
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
_, final_encoder_state = lstm_bid_encoder(inputs, inputs_length,
hparams, train,
"encoder")
else:
inputs_length = None
final_encoder_state = None
# LSTM decoder.
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(
shifted_targets) + 1
hparams_decoder = copy.copy(hparams)
hparams_decoder.hidden_size = 2 * hparams.hidden_size
decoder_outputs, _ = lstm(common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams_decoder,
train,
"decoder",
initial_state=final_encoder_state)
return tf.expand_dims(decoder_outputs, axis=2)
def lstm_seq2seq_internal(inputs, targets, hparams, train):
"""The basic LSTM seq2seq model, main step used for training."""
with tf.variable_scope("lstm_seq2seq"):
if inputs is not None:
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
inputs = tf.reverse_sequence(inputs, inputs_length, seq_axis=1)
_, final_encoder_state = lstm(inputs, inputs_length, hparams, train,
"encoder")
else:
final_encoder_state = None
# LSTM decoder.
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(shifted_targets) + 1
decoder_outputs, _ = lstm(
common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams,
train,
"decoder",
initial_state=final_encoder_state)
return tf.expand_dims(decoder_outputs, axis=2)
def nested_list_operations(input_seq, action_seq, hp, name=""):
batch_size, _, _, hidden_size = common_layers.shape_list(input_seq)
# hidden_size = hp.hidden_size
# if hp.concat_context:
# hidden_size *= 2
cell = NestedListOperationCell(hidden_size,
list_size=hp.list_size,
num_lists=hp.num_lists)
sequence_length = common_layers.length_from_embedding(input_seq)
sequence_length = tf.identity(sequence_length, "sequence_length")
# hidden_size = common_layers.shape_list(input_seq)[-1]
cell_input = tf.concat([action_seq, input_seq], axis=-1)
# batch_size = common_layers.shape_list(cell_input)[0]
cell_input = tf.squeeze(cell_input, axis=2)
cell_input = tf.identity(cell_input, "cell_input")
initial_state = tf.zeros(shape=[batch_size, cell.state_size],
dtype=tf.float32)
with tf.variable_scope(name):
history, final_states = tf.nn.dynamic_rnn(cell,
cell_input,
sequence_length,
initial_state=initial_state,
dtype=tf.float32,
time_major=False)
grid_structured_states = tf.reshape(
final_states, [-1, cell.num_lists, cell.list_size, hidden_size])
grid_structured_states = tf.identity(grid_structured_states,
"grid_structured_states")
return grid_structured_states
def bid_gru_encode(input_seq, hparams, target_space, features, name, sequence_length=None):
if sequence_length == None:
sequence_length = common_layers.length_from_embedding(input_seq)
input_seq = common_layers.flatten4d3d(input_seq)
with tf.variable_scope(name):
cell_fw = [tf.nn.rnn_cell.GRUCell(hparams.hidden_size) for _ in range(hparams.num_hidden_layers)]
cell_bw = [tf.nn.rnn_cell.GRUCell(hparams.hidden_size) for _ in range(hparams.num_hidden_layers)]
((encoder_fw_outputs, encoder_bw_outputs),
(encoder_fw_state, encoder_bw_state)) = tf.nn.bidirectional_dynamic_rnn(
tf.nn.rnn_cell.MultiRNNCell(cell_fw),
tf.nn.rnn_cell.MultiRNNCell(cell_bw),
input_seq,
sequence_length,
initial_state_fw=None,
initial_state_bw=None,
dtype=tf.float32,
time_major=False)
encoder_outputs = tf.concat((encoder_fw_outputs, encoder_bw_outputs), 2)
encoder_outputs = tf.expand_dims(encoder_outputs, axis=-2)
final_output = tf.concat((encoder_fw_state[-1], encoder_bw_state[-1]),-1)
final_output = tf.expand_dims(final_output, axis=-2)
final_output = tf.expand_dims(final_output, axis=-2)
return encoder_outputs, final_output
def create_model_encode_decode(
self, inputs, y_id): # inp[batch step 1 hid] yid[batch step 1 1]
hparams = self.hparams
train_flag = self.train_flag
vocab_size = self.vocabsz
embeddings_y = self.embeddings_y
with tf.variable_scope("foo", reuse=tf.AUTO_REUSE):
### y embed
y = tf.nn.embedding_lookup(embeddings_y, y_id)
y = tf.squeeze(y, axis=3) # [? ? 1 hid]
if len(inputs.shape) == 2: # [batch hid]
inputs = tf.expand_dims(tf.expand_dims(inputs, axis=1), axis=1)
inputs_length = common_layers.length_from_embedding(
inputs) # [batch step 1 hid]
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
inputs = tf.reverse_sequence(inputs, inputs_length, seq_axis=1)
_, final_encoder_state = lstm_yr(
inputs, inputs_length, hparams, train_flag,
"encoder") # finale_encode_state must be lstmStateTuple
##
# LSTM decoder.
shifted_targets = common_layers.shift_right(
y) # [46,23,78]->[0,46,23] | [batch step 1 hid]
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(
shifted_targets) + 1
decoder_outputs, _ = lstm_yr(
common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams,
train_flag,
"decoder",
initial_state=final_encoder_state)
# decode output [batch step hid]
decoder_outputs = tf.layers.dense(inputs=decoder_outputs,
units=vocab_size)
# ->[batch step vocabsz]
decoder_outputs = self.tensor3dto4d(decoder_outputs)
return decoder_outputs
def target_reversing_and_padding(target_seq, list_size):
targets_length = common_layers.length_from_embedding(target_seq)
flipped_target_seq = tf.reverse_sequence(target_seq,
targets_length,
seq_axis=1)
flipped_target_seq = tf.pad(flipped_target_seq,
[[0, 0], [0, list_size], [0, 0], [0, 0]])
flipped_target_seq = flipped_target_seq[:, :list_size, :, :]
flipped_target_seq = tf.identity(flipped_target_seq, "flipped_target_seq")
return flipped_target_seq
def body(self, features):
inputs = features["inputs"]
train = self._hparams.mode == tf.estimator.ModeKeys.TRAIN
encoder_outputs, final_encoder_state, encoder_decoder_attention_bias, inputs_length = \
self.encode(inputs, self._hparams)
if "targets_actions" in features:
targets = features["targets_actions"]
else:
tf.logging.warn(
"CopySeq2Seq must be used with a SemanticParsing problem with a ShiftReduceGrammar; bad things will happen otherwise"
)
targets = features["targets"]
# LSTM decoder with attention
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(
shifted_targets) + 1
shifted_targets = common_layers.flatten4d3d(shifted_targets)
hparams_decoder = copy.copy(self._hparams)
hparams_decoder.hidden_size = 2 * self._hparams.hidden_size
decoder_output = lstm_attention_decoder(shifted_targets,
hparams_decoder, train,
"decoder", final_encoder_state,
encoder_outputs, inputs_length,
targets_length)
decoder_output = tf.expand_dims(decoder_output, axis=2)
body_output = dict()
target_modality = self._problem_hparams.target_modality \
if self._problem_hparams else {"targets": None}
assert self._hparams.pointer_layer in ("attentive",
"decaying_attentive")
for key, modality in target_modality.items():
if isinstance(modality, CopyModality):
with tf.variable_scope("copy_layer/" + key):
if self._hparams.pointer_layer == "decaying_attentive":
output_layer = DecayingAttentivePointerLayer(
encoder_outputs)
else:
output_layer = AttentivePointerLayer(encoder_outputs)
scores = output_layer(decoder_output)
scores += encoder_decoder_attention_bias
body_output[key] = scores
else:
body_output[key] = decoder_output
return body_output
def lstm_seq2seq_internal_attention_bid_encoder(inputs, targets, hparams,
train):
"""LSTM seq2seq model with attention, main step used for training."""
with tf.variable_scope("lstm_seq2seq_attention_bid_encoder"):
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
encoder_outputs, final_encoder_state = lstm_bid_encoder(
inputs, inputs_length, hparams, train, "encoder")
# LSTM decoder with attention
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(shifted_targets) + 1
hparams_decoder = copy.copy(hparams)
hparams_decoder.hidden_size = 2 * hparams.hidden_size
decoder_outputs = lstm_attention_decoder(
common_layers.flatten4d3d(shifted_targets), hparams_decoder, train,
"decoder", final_encoder_state, encoder_outputs,
inputs_length, targets_length)
return tf.expand_dims(decoder_outputs, axis=2)
def body(self, features):
if self._hparams.initializer == "orthogonal":
raise ValueError("LSTM models fail with orthogonal initializer.")
train = self._hparams.mode == tf.estimator.ModeKeys.TRAIN
inputs = features.get("inputs")
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
inputs = tf.reverse_sequence(inputs, inputs_length, seq_axis=1)
encoder_output, _ = lstm(inputs, inputs_length, self._hparams, train,
"encoder")
return tf.expand_dims(encoder_output, axis=2)
def body(self, features):
if self._hparams.initializer == "orthogonal":
raise ValueError("LSTM models fail with orthogonal initializer.")
train = self._hparams.mode == tf.estimator.ModeKeys.TRAIN
inputs = features.get("inputs")
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
inputs = tf.reverse_sequence(inputs, inputs_length, seq_axis=1)
encoder_output, _ = lstm(inputs, inputs_length, self._hparams, train,
"encoder")
return tf.expand_dims(encoder_output, axis=2)
def _build_lm_inputs(self, features):
"""Builds inputs and targets for LM training."""
targets = features["targets"]
target_tags = features["target_tags"]
if self._hparams.mode == tf.estimator.ModeKeys.PREDICT:
target_tags = tf.tile(target_tags, [self._hparams.beam_width, 1, 1, 1])
# Construct LM inputs.
inputs = common_layers.shift_right(targets, pad_value=target_tags)
inputs_length = common_layers.length_from_embedding(targets) + 1
inputs = common_layers.flatten4d3d(inputs)
return inputs, inputs_length
def lstm_seq2seq_internal_attention(inputs, targets, hparams, train):
"""LSTM seq2seq model with attention, main step used for training."""
with tf.variable_scope("lstm_seq2seq_attention"):
# This is a temporary fix for varying-length sequences within in a batch.
# A more complete fix should pass a length tensor from outside so that
# all the lstm variants can use it.
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
# LSTM encoder.
inputs = tf.reverse_sequence(inputs, inputs_length, seq_axis=1)
encoder_outputs, final_encoder_state = lstm(
inputs, inputs_length, hparams, train, "encoder")
# LSTM decoder with attention.
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(shifted_targets) + 1
decoder_outputs = lstm_attention_decoder(
common_layers.flatten4d3d(shifted_targets), hparams, train, "decoder",
final_encoder_state, encoder_outputs, inputs_length, targets_length)
return tf.expand_dims(decoder_outputs, axis=2)
def encode(self, inputs, hparams, features=None):
train = hparams.mode == tf.estimator.ModeKeys.TRAIN
inputs_length = common_layers.length_from_embedding(inputs)
# Flatten inputs.
inputs = common_layers.flatten4d3d(inputs)
encoder_padding = common_attention.embedding_to_padding(inputs)
encoder_decoder_attention_bias = common_attention.attention_bias_ignore_padding(
encoder_padding)
# LSTM encoder.
encoder_outputs, final_encoder_state = lstm_bid_encoder(
inputs, inputs_length, self._hparams, train, "encoder")
return encoder_outputs, final_encoder_state, encoder_decoder_attention_bias, inputs_length
def body(self, features):
inputs = features["inputs"]
hparams = self._hparams
train = self._hparams.mode == tf.estimator.ModeKeys.TRAIN
with tf.variable_scope("lstm"):
inputs_length = common_layers.length_from_embedding(inputs)
inputs = common_layers.flatten4d3d(inputs)
_, final_encoder_state = lstm.lstm(inputs,
inputs_length,
hparams,
train,
name="encoder")
final_output = final_encoder_state[-1]
c, h = final_output
final_hidden_output = tf.expand_dims(h, axis=-2)
final_hidden_output = tf.expand_dims(final_hidden_output, axis=-2)
return final_hidden_output
def lstm_encode(input_seq, hparams, target_space, features, name, sequence_length=None):
if sequence_length == None:
sequence_length = common_layers.length_from_embedding(input_seq)
input_seq = common_layers.flatten4d3d(input_seq)
layers = [tf.nn.rnn_cell.LSTMCell(hparams.hidden_size) for _ in range(hparams.num_hidden_layers)]
with tf.variable_scope(name):
# hidden_outputs (outputs of last layer) : [batch_size, seq_len, hidden_size]
# layer_final_output (layer-wise final outputs) : [num_hidden_layers, 2, batch_size, hidden_size]
hidden_outputs, layer_final_output = tf.nn.dynamic_rnn(
tf.nn.rnn_cell.MultiRNNCell(layers),
input_seq,
sequence_length,
initial_state=None,
dtype=tf.float32,
time_major=False)
hidden_outputs = tf.expand_dims(hidden_outputs, axis=-2)
c, h = layer_final_output[-1]
final_output = h
final_output = tf.expand_dims(final_output, axis=-2)
final_output = tf.expand_dims(final_output, axis=-2)
return hidden_outputs, final_output
def render2cmd_v3_internal(self, features, hparams, train):
# inputs and targets are both sequences with
# shape = [batch, seq_len, 1, hparams.problem.feature_dim]
targets = features['targets']
losses = {}
sampled_bottleneck = self.pretrained_visual_encoder(features, hparams)
if hparams.sg_bottleneck:
sampled_bottleneck = tf.stop_gradient(sampled_bottleneck)
with tf.variable_scope('render2cmd_v3_internal'):
# override bottleneck, or return it, if requested
if 'bottleneck' in features:
if common_layers.shape_list(features['bottleneck'])[0] == 0:
# return sampled_bottleneck,
# set losses['training'] = 0 so self.top() doesn't get called on it
return sampled_bottleneck, {'training': 0.0}
else:
# we want to use the given bottleneck
sampled_bottleneck = features['bottleneck']
# finalize bottleneck
unbottleneck_dim = hparams.hidden_size * 2 # twice because using LSTM
if hparams.twice_decoder:
unbottleneck_dim = unbottleneck_dim * 2
# unbottleneck back to LSTMStateTuple
dec_initial_state = []
for hi in range(hparams.num_hidden_layers):
unbottleneck = self.unbottleneck(sampled_bottleneck,
unbottleneck_dim,
name_append='_{}'.format(hi))
dec_initial_state.append(
rnn.LSTMStateTuple(
c=unbottleneck[:, :unbottleneck_dim // 2],
h=unbottleneck[:, unbottleneck_dim // 2:]))
dec_initial_state = tuple(dec_initial_state)
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(
shifted_targets) + 1
# LSTM decoder
hparams_decoder = copy.copy(hparams)
if hparams.twice_decoder:
hparams_decoder.hidden_size = 2 * hparams.hidden_size
if hparams.mode == tf.estimator.ModeKeys.PREDICT:
decoder_outputs, _ = self.lstm_decoder_infer(
common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams_decoder,
features['targets_cls'],
train,
initial_state=dec_initial_state,
bottleneck=sampled_bottleneck)
else:
decoder_outputs, _ = self.lstm_decoder(
common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams_decoder,
features['targets_cls'],
train,
initial_state=dec_initial_state,
bottleneck=sampled_bottleneck)
ret = tf.expand_dims(decoder_outputs, axis=2)
return ret, losses
def infer_valid_length_from_top_list(top_list, list_size):
argmaxed_top_list = get_argmaxed_top_list(top_list, list_size)
valid_length = common_layers.length_from_embedding(argmaxed_top_list)
valid_length = tf.identity(valid_length, "output_valid_length")
return valid_length
def render2cmd_v3_internal(self, features, hparams, train):
# inputs and targets are both sequences with
# shape = [batch, seq_len, 1, hparams.problem.feature_dim]
print(
"render2cmd_v3_internal render2cmd_v3_internalrender2cmd_v3_internalrender2cmd_v3_internalrender2cmd_v3_internal"
)
all_targets = features['targets']
all_targets_cls = features['targets_cls']
all_targets_font_cls = features['targets_fnt']
all_targets_psr = features['targets_psr']
all_batch_size = common_layers.shape_list(all_targets)[0]
batch_size = all_batch_size // 2
sources = all_targets[:batch_size, ...]
sources_cls = all_targets_cls[:batch_size, ...]
sources_fnt = all_targets_font_cls[:batch_size, ...]
sources_psr = all_targets_psr[:batch_size, ...]
targets = all_targets[batch_size:, ...]
targets_cls = all_targets_cls[batch_size:, ...]
targets_fnt = all_targets_font_cls[batch_size:, ...]
targets_psr = all_targets_psr[batch_size:, ...]
losses = {}
# sampled_bottleneck = self.pretrained_visual_encoder(features, hparams)
# if hparams.sg_bottleneck:
# sampled_bottleneck = tf.stop_gradient(sampled_bottleneck)
# embd = self.cls_embedding(sources_cls, sources_fnt, targets_cls, targets_fnt)
vis_embd = self.vis_encoder(sources_psr, targets_psr, targets_cls)
# print("embd embd embd embd embd embd embd ", embd.shape)
print("vis embd vis embd vis embd vis embd vis", vis_embd.shape)
sampled_bottleneck = vis_embd
with tf.variable_scope('render2cmd_v3_internal'):
# override bottleneck, or return it, if requested
# if 'bottleneck' in features:
# if common_layers.shape_list(features['bottleneck'])[0] == 0:
# # return sampled_bottleneck,
# # set losses['training'] = 0 so self.top() doesn't get called on it
# print("RETURNRETURNRETURNRETURNRETURNRETURNRETURNRETURNRETURNRETURNRETURN")
# return sampled_bottleneck, {'training': 0.0}
# else:
# # we want to use the given bottleneck
# sampled_bottleneck = features['bottleneck']
# finalize bottleneck
unbottleneck_dim = hparams.hidden_size * 2 # twice because using LSTM
if hparams.twice_decoder:
unbottleneck_dim = unbottleneck_dim * 2
dec_initial_state = []
# LSTM encoder
_, encoder_output_states = self.lstm_encoder(
common_layers.flatten4d3d(sources), hparams)
print(
"targets shape targets shape targets shape targets shape targets shape ",
targets.shape)
print('run stacking...')
print(
"sample bottleneck shape sample bottleneck shape sample bottleneck shape ",
sampled_bottleneck.shape)
print(
"sources shape sources shape sources shape sources shape sources shape",
sources.shape)
# input()
for hi in range(hparams.num_hidden_layers):
unbottleneck = self.unbottleneck(sampled_bottleneck,
unbottleneck_dim,
name_append='_{}'.format(hi))
c, h = encoder_output_states[hi]
# print(unbottleneck.shape)
# print(c.shape, h.shape)
# first_dim = common_layers.shape_list(unbottleneck)[0]
# print(first_dim)
# c = tf.tile(c,[first_dim,1])
# h = tf.tile(h,[first_dim,1])
# input()
dec_initial_state.append(
tf.nn.rnn_cell.LSTMStateTuple(
c=tf.concat(
[unbottleneck[:, :unbottleneck_dim // 2], c], 1),
h=tf.concat(
[unbottleneck[:, unbottleneck_dim // 2:], h], 1)))
dec_initial_state = tuple(dec_initial_state)
# print('checkshape dec_initial_state')
# print(dec_initial_state)
# input()
shifted_targets = common_layers.shift_right(targets)
# Add 1 to account for the padding added to the left from shift_right
targets_length = common_layers.length_from_embedding(
shifted_targets) + 1
# LSTM decoder
hparams_decoder = copy.copy(hparams)
if hparams.twice_decoder:
hparams_decoder.hidden_size = 2 * hparams.hidden_size
if hparams.mode == tf.estimator.ModeKeys.PREDICT:
decoder_outputs, _ = self.lstm_decoder_infer(
common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams_decoder,
targets_cls,
train,
initial_state=dec_initial_state,
bottleneck=sampled_bottleneck)
else:
decoder_outputs, _ = self.lstm_decoder(
common_layers.flatten4d3d(shifted_targets),
targets_length,
hparams_decoder,
targets_cls,
train,
initial_state=dec_initial_state,
bottleneck=sampled_bottleneck)
ret = tf.expand_dims(decoder_outputs, axis=2)
return ret, losses
