def exe_decoder(self, dec, enc_output, attention=None):
params = copy.deepcopy(self.e_params)
if "gru" in self.e_params["fw_cell"]:
params["fw_cell"] = "gru_block"
elif "rnn" in self.e_params["fw_cell"]:
params["fw_cell"] = "rnn"
else:
params["fw_cell"] = "lstm_block"
with tf.variable_scope("decoder",
initializer=self.initializer,
reuse=tf.AUTO_REUSE):
if self.dtype == "grid":
outputs, _ = rnn_utils.execute_decoder_cnn(
dec, enc_output, self.decoder_length, params, attention,
self.use_cnn, self.use_gen_cnn, self.mtype,
self.use_batch_norm, self.dropout)
outputs = tf.stack(outputs, axis=1)
else:
if "gru" in self.e_params["fw_cell"] or self.e_params[
"fw_cell"] == "rnn":
enc_output = tf.squeeze(enc_output[0], 0)
dec_data = tf.transpose(dec, [0, 2, 1, 3])
dec_data = tf.reduce_mean(dec_data, axis=1)
outputs = rnn_utils.execute_decoder(dec_data, enc_output,
self.decoder_length,
params, attention,
self.dropout_placeholder)
outputs = tf.reshape(tf.stack(outputs, axis=1),
shape=(pr.batch_size,
self.decoder_length))
return outputs
def exe_decoder(self, dec_hidden_vectors, fn_state):
with tf.variable_scope("decoder",
initializer=self.initializer,
reuse=tf.AUTO_REUSE):
params = deepcopy(self.e_params)
if "gru" in self.e_params["fw_cell"]:
params["fw_cell"] = "gru_block"
else:
params["fw_cell"] = "lstm_block"
# concatenate with z noise
dec_hidden_vectors = tf.concat([dec_hidden_vectors, self.z],
axis=1)
dec_hidden_vectors = tf.layers.dense(dec_hidden_vectors,
self.rnn_hidden_units,
name="generation_hidden_seed",
activation=tf.nn.tanh)
outputs = rnn_utils.execute_decoder_cnn(
None, fn_state, self.decoder_length, params,
dec_hidden_vectors, self.use_cnn, self.use_gen_cnn, self.mtype,
self.use_batch_norm, self.dropout)
outputs = tf.stack(outputs, axis=1)
outputs = tf.reshape(outputs, [
pr.batch_size, self.decoder_length, pr.grid_size * pr.grid_size
])
return outputs
def exe_decoder_critic(self, dec, enc_output, attention=None):
with tf.variable_scope("decoder",
initializer=self.initializer,
reuse=tf.AUTO_REUSE):
params = copy.deepcopy(self.e_params)
params["fw_cell"] = "gru_block"
outputs, _ = rnn_utils.execute_decoder_cnn(
dec, enc_output, self.decoder_length, params, attention,
self.use_cnn, self.use_gen_cnn, self.mtype,
self.use_batch_norm, self.dropout)
# batch_size x decoder_length x grid_size x grid_size
outputs = tf.stack(outputs, axis=1)
# batch_size x decoder_length
return outputs
def exe_decoder(self, dec_hidden_vectors, fn_state):
with tf.variable_scope("decoder", initializer=self.initializer, reuse=tf.AUTO_REUSE):
params = deepcopy(self.e_params)
if "gru" in self.e_params["fw_cell"]:
params["fw_cell"] = "gru_block"
else:
params["fw_cell"] = "lstm_block"
dctype = 5
if self.grid_size == 32:
dctype = 7
outputs = rnn_utils.execute_decoder_cnn(None, fn_state, self.decoder_length, params, dec_hidden_vectors, \
self.use_cnn, self.use_gen_cnn, self.mtype, self.use_batch_norm, \
self.dropout, dctype=dctype, offset=self.offset)
classes = None
outputs = tf.stack(outputs, axis=1)
outputs = tf.reshape(outputs, [self.batch_size, self.decoder_length - self.offset, self.grid_size * self.grid_size])
return outputs, classes
def exe_decoder(self, forecast_outputs, fn_state, weather_hidden_vector,
enc_inputs):
last_enc_timestep = enc_inputs[:, :, :, 0]
with tf.variable_scope("decoder",
initializer=self.initializer,
reuse=tf.AUTO_REUSE):
params = deepcopy(self.e_params)
if "gru" in self.e_params["fw_cell"]:
params["fw_cell"] = "gru_block"
else:
params["fw_cell"] = "lstm_block"
outputs, classes = rnn_utils.execute_decoder_cnn(forecast_outputs, fn_state, self.decoder_length, params, weather_hidden_vector, True, \
self.use_gen_cnn, self.mtype, False, self.dropout, dctype=5, classify=bool(self.num_class), \
init_input=last_enc_timestep, grid_input=False)
outputs = tf.stack(outputs, axis=1)
if classes:
classes = tf.stack(classes, axis=1)
return outputs, classes