def predictor(enc_flat,
action,
lstm_states,
pred_depth,
reuse=False,
scope_prefix='',
hparams=None):
"""LSTM predictor network."""
with tf.variable_scope(scope_prefix + 'predict', reuse=reuse):
enc_final_size = enc_flat.get_shape().as_list()[1]
action_size = action.get_shape().as_list()[1]
initial_size = (enc_final_size + action_size)
batch_size = tf.shape(enc_flat)[0]
init_stddev = 1e-2
pre_pred = tf.concat([enc_flat, action], 1)
pre_pred = tf.layers.dense(
pre_pred,
initial_size,
kernel_initializer=tf.truncated_normal_initializer(
stddev=init_stddev))
# This is only needed or the GAN version.
if hparams.pred_noise_std > 0:
# Add the noise like this so a pretrained model can be used.
pred_noise = tf.random_normal(shape=[batch_size, 100],
stddev=hparams.pred_noise_std)
pre_pred += tf.layers.dense(
pred_noise,
initial_size,
kernel_initializer=tf.truncated_normal_initializer(
stddev=init_stddev),
name='noise_dense')
pre_pred = tf.nn.relu(pre_pred)
if lstm_states[pred_depth - 2] is None:
back_connect = tf.tile(
tf.get_variable('back_connect_init',
shape=[1, initial_size * 2],
initializer=tf.truncated_normal_initializer(
stddev=init_stddev)), (batch_size, 1))
else:
back_connect = lstm_states[pred_depth - 2]
lstm_init_stddev = 1e-4
part_pred, lstm_states[0] = common_video.lstm_cell(
tf.concat([pre_pred, back_connect], 1),
lstm_states[0],
initial_size,
use_peepholes=True,
initializer=tf.truncated_normal_initializer(
stddev=lstm_init_stddev),
num_proj=initial_size)
part_pred = contrib.layers().layer_norm(part_pred)
pred = part_pred
for pred_layer_num in range(1, pred_depth, 2):
part_pred, lstm_states[pred_layer_num] = common_video.lstm_cell(
pred,
lstm_states[pred_layer_num],
initial_size,
use_peepholes=True,
initializer=tf.truncated_normal_initializer(
stddev=lstm_init_stddev),
num_proj=initial_size)
pred += part_pred
part_pred, lstm_states[
pred_layer_num + 1] = common_video.lstm_cell(
tf.concat([pred, pre_pred], 1),
lstm_states[pred_layer_num + 1],
initial_size,
use_peepholes=True,
initializer=tf.truncated_normal_initializer(
stddev=lstm_init_stddev),
num_proj=initial_size)
part_pred = contrib.layers().layer_norm(part_pred)
pred += part_pred
pred = tf.layers.dense(
pred,
enc_final_size,
kernel_initializer=tf.truncated_normal_initializer(
stddev=init_stddev))
if hparams.enc_pred_use_l2norm:
pred = tf.nn.l2_normalize(pred, 1)
return pred
def predictor(enc_flat,
action,
lstm_states,
pred_depth,
reuse=False,
scope_prefix='',
hparams=None):
"""LSTM predictor network."""
with tf.variable_scope(scope_prefix + 'predict', reuse=reuse):
enc_final_size = enc_flat.get_shape().as_list()[1]
action_size = action.get_shape().as_list()[1]
initial_size = (enc_final_size + action_size)
batch_size = tf.shape(enc_flat)[0]
init_stddev = 1e-2
pre_pred = tf.concat([enc_flat, action], 1)
pre_pred = tf.layers.dense(
pre_pred,
initial_size,
kernel_initializer=tf.truncated_normal_initializer(stddev=init_stddev))
# This is only needed or the GAN version.
if hparams.pred_noise_std > 0:
# Add the noise like this so a pretrained model can be used.
pred_noise = tf.random_normal(
shape=[batch_size, 100], stddev=hparams.pred_noise_std)
pre_pred += tf.layers.dense(
pred_noise,
initial_size,
kernel_initializer=tf.truncated_normal_initializer(
stddev=init_stddev),
name='noise_dense')
pre_pred = tf.nn.relu(pre_pred)
if lstm_states[pred_depth - 2] is None:
back_connect = tf.tile(
tf.get_variable(
'back_connect_init',
shape=[1, initial_size * 2],
initializer=tf.truncated_normal_initializer(stddev=init_stddev))
, (batch_size, 1))
else:
back_connect = lstm_states[pred_depth - 2]
lstm_init_stddev = 1e-4
part_pred, lstm_states[0] = common_video.lstm_cell(
tf.concat([pre_pred, back_connect], 1),
lstm_states[0],
initial_size,
use_peepholes=True,
initializer=tf.truncated_normal_initializer(stddev=lstm_init_stddev),
num_proj=initial_size)
part_pred = tf.contrib.layers.layer_norm(part_pred)
pred = part_pred
for pred_layer_num in range(1, pred_depth, 2):
part_pred, lstm_states[pred_layer_num] = common_video.lstm_cell(
pred,
lstm_states[pred_layer_num],
initial_size,
use_peepholes=True,
initializer=tf.truncated_normal_initializer(stddev=lstm_init_stddev),
num_proj=initial_size)
pred += part_pred
part_pred, lstm_states[pred_layer_num + 1] = common_video.lstm_cell(
tf.concat([pred, pre_pred], 1),
lstm_states[pred_layer_num + 1],
initial_size,
use_peepholes=True,
initializer=tf.truncated_normal_initializer(stddev=lstm_init_stddev),
num_proj=initial_size)
part_pred = tf.contrib.layers.layer_norm(part_pred)
pred += part_pred
pred = tf.layers.dense(
pred,
enc_final_size,
kernel_initializer=tf.truncated_normal_initializer(stddev=init_stddev))
if hparams.enc_pred_use_l2norm:
pred = tf.nn.l2_normalize(pred, 1)
return pred
