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):
x = self.state
out, state = stateful_lstm(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_lstm(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)