def add_logits_op_target(self):
if self.cnn_format == "NHWC":
x = tf.transpose(self.state_target, [0, 2, 3, 1])
else:
x = self.state_target
w, b, out, _ = conv2d_layer(x, 32, [8, 8], [4, 4], scope_name="conv1_target", summary_tag=None,
activation=tf.nn.relu, data_format=self.cnn_format)
self.w_target["wc1"] = w
self.w_target["bc1"] = b
w, b, out, _ = conv2d_layer(out, 64, [4, 4], [2, 2], scope_name="conv2_target", summary_tag=None,
activation=tf.nn.relu, data_format=self.cnn_format)
self.w_target["wc2"] = w
self.w_target["bc2"] = b
w, b, out, _ = conv2d_layer(out, 64, [3, 3], [1, 1], scope_name="conv3_target", summary_tag=None,
activation=tf.nn.relu, data_format=self.cnn_format)
self.w_target["wc3"] = w
self.w_target["bc3"] = b
shape = out.get_shape().as_list()
out_flat = tf.reshape(out, [-1, reduce(lambda x, y: x * y, shape[1:])])
w, b, out = fully_connected_layer(out_flat, 512, scope_name="fully1_target")
self.w_target["wf1"] = w
self.w_target["bf1"] = b
w, b, out = fully_connected_layer(out, self.n_actions, scope_name="out_target", activation=None)
self.w_target["wout"] = w
self.w_target["bout"] = b
self.q_target_out = out
self.q_target_action = tf.argmax(self.q_target_out, axis=1)
def add_logits_op_train(self):
if self.cnn_format == "NHWC":
x = tf.transpose(self.state, [0, 2, 3, 1])
else:
x = self.state
self.image_summary = []
w, b, out, summary = conv2d_layer(x,
32, [8, 8], [4, 4],
scope_name="conv1_train",
summary_tag="conv1_out",
activation=tf.nn.relu,
data_format=self.cnn_format)
self.w["wc1"] = w
self.w["bc1"] = b
self.image_summary.append(summary)
w, b, out, summary = conv2d_layer(out,
64, [4, 4], [2, 2],
scope_name="conv2_train",
summary_tag="conv2_out",
activation=tf.nn.relu,
data_format=self.cnn_format)
self.w["wc2"] = w
self.w["bc2"] = b
self.image_summary.append(summary)
w, b, out, summary = conv2d_layer(out,
64, [3, 3], [1, 1],
scope_name="conv3_train",
summary_tag="conv3_out",
activation=tf.nn.relu,
data_format=self.cnn_format)
self.w["wc3"] = w
self.w["bc3"] = b
self.image_summary.append(summary)
shape = out.get_shape().as_list()
out_flat = tf.reshape(
out, [tf.shape(out)[0], 1, shape[1] * shape[2] * shape[3]])
out, state = stateful_lstm(out_flat,
self.num_lstm_layers,
self.lstm_size,
tuple([self.lstm_state_train]),
scope_name="lstm_train")
self.state_output_c = state[0][0]
self.state_output_h = state[0][1]
shape = out.get_shape().as_list()
out = tf.reshape(out, [tf.shape(out)[0], shape[2]])
w, b, out = fully_connected_layer(out,
self.n_actions,
scope_name="out_train",
activation=None)
self.w["wout"] = w
self.w["bout"] = b
self.q_out = out
self.q_action = tf.argmax(self.q_out, axis=1)
def add_logits_op_train(self):
# self.colour_summary = []
# self.colour_summary.append('')
# if self.cnn_format == "NHWC":
# x = tf.transpose(self.state, [0, 2, 3, 1])
# else:
# x = self.state
# w, b, out, summary = conv2d_layer(x, 32, [8, 8], [4, 4], scope_name="conv1_train", summary_tag="conv1_out",
# activation=tf.nn.relu, data_format=self.cnn_format)
# self.w["wc1"] = w
# self.w["bc1"] = b
# self.image_summary.append(summary)
# w, b, out, summary = conv2d_layer(out, 64, [4, 4], [2, 2], scope_name="conv2_train", summary_tag="conv2_out",
# activation=tf.nn.relu, data_format=self.cnn_format)
# self.w["wc2"] = w
# self.w["bc2"] = b
# self.image_summary.append(summary)
# w, b, out, summary = conv2d_layer(out, 64, [3, 3], [1, 1], scope_name="conv3_train", summary_tag="conv3_out",
# activation=tf.nn.relu, data_format=self.cnn_format)
# self.w["wc3"] = w
# self.w["bc3"] = b
# self.image_summary.append(summary)
# shape = out.get_shape().as_list()
# out_flat = tf.reshape(out, [-1, reduce(lambda x, y: x * y, shape[1:])])
x = self.state
shape = x.get_shape().as_list()
x_flat = tf.reshape(x, [-1, reduce(lambda x,y: x*y, shape[1:])])
w, b, out = fully_connected_layer(x_flat, 512, scope_name="fully1_train")
self.w["wf1"] = w
self.w["bf1"] = b
w, b, out = fully_connected_layer(out, self.n_actions, scope_name="out_train", activation=None)
self.w["wout"] = w
self.w["bout"] = b
self.q_out = out
self.q_action = tf.argmax(self.q_out, axis=1)
def add_logits_op_target(self):
if self.cnn_format == "NHWC":
x = tf.transpose(self.state_target, [0, 2, 3, 1])
else:
x = self.state_target
w, b, out, _ = conv2d_layer(x,
32, [8, 8], [4, 4],
scope_name="conv1_target",
summary_tag=None,
activation=tf.nn.relu,
data_format=self.cnn_format)
self.w_target["wc1"] = w
self.w_target["bc1"] = b
w, b, out, _ = conv2d_layer(out,
64, [4, 4], [2, 2],
scope_name="conv2_target",
summary_tag=None,
activation=tf.nn.relu,
data_format=self.cnn_format)
self.w_target["wc2"] = w
self.w_target["bc2"] = b
w, b, out, _ = conv2d_layer(out,
64, [3, 3], [1, 1],
scope_name="conv3_target",
summary_tag=None,
activation=tf.nn.relu,
data_format=self.cnn_format)
self.w_target["wc3"] = w
self.w_target["bc3"] = b
shape = out.get_shape().as_list()
out_flat = tf.reshape(
out, [tf.shape(out)[0], 1, shape[1] * shape[2] * shape[3]])
out, state = stateful_gru(out_flat,
self.num_lstm_layers,
self.gru_size,
scope_name="gru_target")
self.state_output_target = state[0]
# self.state_output_target_h = state[0][1]
shape = out.get_shape().as_list()
out = tf.reshape(out, [tf.shape(out)[0], shape[2]])
w, b, out = fully_connected_layer(out,
self.n_actions,
scope_name="out_target",
activation=None)
self.w_target["wout"] = w
self.w_target["bout"] = b
self.q_target_out = out
self.q_target_action = tf.argmax(self.q_target_out, axis=1)
def add_logits_op_train(self):
self.image_summary = []
if self.cnn_format == "NHWC": #Se ajusta la entrada x de acuerdo lal hardware
x = tf.transpose(self.state, [0, 2, 3, 1])
else:
x = self.state
#Se lleva acabo la primera convolución sobre la enrada
w, b, out, summary = conv2d_layer(x, 32, [8, 8], [4, 4], scope_name="conv1_train", summary_tag="conv1_out",
activation=tf.nn.relu, data_format=self.cnn_format)
self.w["wc1"] = w #se almacena los valores de peso y bias y summary de la primera conv
self.w["bc1"] = b
self.image_summary.append(summary)
#Se lleva a cabo la segunda convolución con la salida de la primera
w, b, out, summary = conv2d_layer(out, 64, [4, 4], [2, 2], scope_name="conv2_train", summary_tag="conv2_out",
activation=tf.nn.relu, data_format=self.cnn_format)
self.w["wc2"] = w #Se almacena la valores de peso, bias y summary de la segunda conv
self.w["bc2"] = b
self.image_summary.append(summary)
#Se lleva a cabo la tercerca convolución con la salida de la segunda.
w, b, out, summary = conv2d_layer(out, 64, [3, 3], [1, 1], scope_name="conv3_train", summary_tag="conv3_out",
activation=tf.nn.relu, data_format=self.cnn_format)
self.w["wc3"] = w #Se almacena los valores de perso bias y summary de la tercera conv
self.w["bc3"] = b
self.image_summary.append(summary)
shape = out.get_shape().as_list() #Se modficaa el tamaño de la salida
out_flat = tf.reshape(out, [-1, reduce(lambda x, y: x * y, shape[1:])])
#Se pasa los valores a la cuarta cpa conla salida modificada de la tercerca cov
w, b, out = fully_connected_layer(out_flat, 512, scope_name="fully1_train")
self.w["wf1"] = w #Se almacena los valores de peso bias de la primera fullconected
self.w["bf1"] = b
#Se pasa los valores a la última capa
w, b, out = fully_connected_layer(out, self.n_actions, scope_name="out_train", activation=None)
self.w["wout"] = w #Se almacenan los valores de peso bias para la segunda fullconected
self.w["bout"] = b
self.q_out = out #Se asigna el valor al atributo salida q
self.q_action = tf.argmax(self.q_out, axis=1) #Se define la acción como el índice en que se encuentra el mayor valor en qout
def add_logits_op_target(self):
if self.cnn_format == "NHWC": #Se ajusta la entrada de acuerdo al hardware
x = tf.transpose(self.state_target, [0, 2, 3, 1])
else:
x = self.state_target
#Se lleva a cabo la primer convolución con la entrada x
w, b, out, _ = conv2d_layer(x, 32, [8, 8], [4, 4], scope_name="conv1_target", summary_tag=None,
activation=tf.nn.relu, data_format=self.cnn_format)
self.w_target["wc1"] = w #Se almacena los valores de peso y bias
self.w_target["bc1"] = b
#Se lleva a cabo la segunda convolución sobre la salida de la primera conv
w, b, out, _ = conv2d_layer(out, 64, [4, 4], [2, 2], scope_name="conv2_target", summary_tag=None,
activation=tf.nn.relu, data_format=self.cnn_format)
self.w_target["wc2"] = w#Se almacena los valores de peso y bias
self.w_target["bc2"] = b
#Se lleva a cabo la terce conv sobre la salida de la segunda conv
w, b, out, _ = conv2d_layer(out, 64, [3, 3], [1, 1], scope_name="conv3_target", summary_tag=None,
activation=tf.nn.relu, data_format=self.cnn_format)
self.w_target["wc3"] = w#Se almacena los valores de peso y bias
self.w_target["bc3"] = b
shape = out.get_shape().as_list() #Se modifica el tamaño de la salida de la tercer conv
out_flat = tf.reshape(out, [-1, reduce(lambda x, y: x * y, shape[1:])])
#Se pasa los valores a la cuarta capa
w, b, out = fully_connected_layer(out_flat, 512, scope_name="fully1_target")
self.w_target["wf1"] = w#Se almacena los valores de peso y bias
self.w_target["bf1"] = b
#Se pasan los valores a la última capa
w, b, out = fully_connected_layer(out, self.n_actions, scope_name="out_target", activation=None)
self.w_target["wout"] = w#Se almacena los valores de peso y bias
self.w_target["bout"] = b
self.q_target_out = out #See almacena la salida
self.q_target_action = tf.argmax(self.q_target_out, axis=1) #Se almacena el índice de la acción que ofrece mejor retorno
def add_logits_op_train(self):
x = self.state
out, state = stateful_goru(x=x,
num_layers=self.num_lstm_layers,
lstm_size=self.lstm_size,
state_input=tuple([self.lstm_state_train]),
scope_name="lstm_train")
self.state_output_c = state[0][0]
self.state_output_h = state[0][1]
shape = out.get_shape().as_list() #[None,1,512]
out = tf.reshape(out, [tf.shape(out)[0], shape[2]])
w, b, out = fully_connected_layer(out,
self.n_actions,
scope_name="out_train",
activation=None)
self.w["wout"] = w
self.w["bout"] = b
self.q_out = out
self.q_action = tf.argmax(self.q_out, axis=1)
def add_logits_op_target(self):
x = self.state_target #shape: [None, 1, num_colors]
out, state = stateful_goru(x,
self.num_lstm_layers,
self.lstm_size,
tuple([self.lstm_state_target]),
scope_name="lstm_target")
self.state_output_target_c = state[0][0]
self.state_output_target_h = state[0][1]
shape = out.get_shape().as_list()
out = tf.reshape(out, [tf.shape(out)[0], shape[2]])
w, b, out = fully_connected_layer(out,
self.n_actions,
scope_name="out_target",
activation=None)
self.w_target["wout"] = w
self.w_target["bout"] = b
self.q_target_out = out
self.q_target_action = tf.argmax(self.q_target_out, axis=1)
