def _build_net_critic(self, net_name):
norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
input_tfs = [norm_s_tf]
if (self.has_goal()):
norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
input_tfs += [norm_g_tf]
#####################[
if Settings.mode() == Mode.CWT_CNN_v1 or Settings.mode(
) == Mode.CWT_CNN_v2:
cnn_network = NetBuilder.build_net(
MyCNN.NAME, self.wt_tf, self.my_memory_buffer.channel_count)
input_tfs += [cnn_network]
#####################]
h = NetBuilder.build_net(net_name, input_tfs)
norm_val_tf = tf.layers.dense(
inputs=h,
units=1,
activation=None,
kernel_initializer=TFUtil.xavier_initializer)
norm_val_tf = tf.reshape(norm_val_tf, [-1])
val_tf = self.val_norm.unnormalize_tf(norm_val_tf)
return val_tf
def _build_net_actor(self, net_name, init_output_scale):
norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
input_tfs = [norm_s_tf]
if (self.has_goal()):
norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
input_tfs += [norm_g_tf]
#####################[
if Settings.mode() == Mode.CWT_CNN_v1 or Settings.mode(
) == Mode.CWT_CNN_v2:
cnn_network = NetBuilder.build_net(
MyCNN.NAME, self.wt_tf, self.my_memory_buffer.channel_count)
input_tfs += [cnn_network]
#####################]
h = NetBuilder.build_net(net_name, input_tfs)
norm_a_tf = tf.layers.dense(
inputs=h,
units=self.get_action_size(),
activation=None,
kernel_initializer=tf.random_uniform_initializer(
minval=-init_output_scale, maxval=init_output_scale))
a_tf = self.a_norm.unnormalize_tf(norm_a_tf)
return a_tf
def _build_net_critic(self, net_name, input_tfs, reuse=False):
out_size = 1
with tf.variable_scope('critic', reuse=reuse):
h = NetBuilder.build_net(net_name, input_tfs, reuse)
val_tf = tf.layers.dense(inputs=h, units=out_size, activation=None,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
reuse=reuse)
val_tf = tf.squeeze(val_tf, axis=-1)
return val_tf
def _build_net_actor(self, net_name, init_output_scale):
norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
input_tfs = [norm_s_tf]
if (self.has_goal()):
norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
input_tfs += [norm_g_tf]
h = NetBuilder.build_net(net_name, input_tfs)
norm_a_tf = tf.layers.dense(inputs=h, units=self.get_action_size(), activation=None,
kernel_initializer=tf.random_uniform_initializer(minval=-init_output_scale, maxval=init_output_scale))
a_tf = self.a_norm.unnormalize_tf(norm_a_tf)
return a_tf
def _build_net_actor(self, net_name, init_output_scale):
norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
input_tfs = [norm_s_tf]
if (self.has_goal()):
norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
input_tfs += [norm_g_tf]
h = NetBuilder.build_net(net_name, input_tfs)
norm_a_tf = tf.layers.dense(inputs=h, units=self.get_action_size(), activation=None,
kernel_initializer=tf.random_uniform_initializer(minval=-init_output_scale, maxval=init_output_scale))
a_tf = self.a_norm.unnormalize_tf(norm_a_tf)
return a_tf
def _build_net_critic(self, net_name):
norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
input_tfs = [norm_s_tf]
if (self.has_goal()):
norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
input_tfs += [norm_g_tf]
h = NetBuilder.build_net(net_name, input_tfs)
norm_val_tf = tf.layers.dense(inputs=h, units=1, activation=None,
kernel_initializer=TFUtil.xavier_initializer);
norm_val_tf = tf.reshape(norm_val_tf, [-1])
val_tf = self.val_norm.unnormalize_tf(norm_val_tf)
return val_tf
def _build_net_critic(self, net_name):
norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
input_tfs = [norm_s_tf]
if (self.has_goal()):
norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
input_tfs += [norm_g_tf]
h = NetBuilder.build_net(net_name, input_tfs)
norm_val_tf = tf.layers.dense(inputs=h, units=1, activation=None,
kernel_initializer=TFUtil.xavier_initializer);
norm_val_tf = tf.reshape(norm_val_tf, [-1])
val_tf = self.val_norm.unnormalize_tf(norm_val_tf)
return val_tf
def _build_disc_net(self,
net_name,
input_tfs,
init_output_scale,
reuse=False):
out_size = 1
h = net_builder.build_net(net_name, input_tfs, reuse)
logits_tf = tf.layers.dense(
inputs=h,
units=out_size,
activation=None,
reuse=reuse,
kernel_initializer=tf.random_uniform_initializer(
minval=-init_output_scale, maxval=init_output_scale),
name=self.DISC_LOGIT_NAME)
return logits_tf
def _build_net_critic(self, net_name):
with tf.device('cpu:0'):
norm_s_tf = self.s_norm.normalize_tf(self.s_tf)
input_tfs = [norm_s_tf]
if (self.has_goal()):
norm_g_tf = self.g_norm.normalize_tf(self.g_tf)
input_tfs += [norm_g_tf]
h = NetBuilder.build_net(net_name, input_tfs)
norm_val_tf = tf.layers.dense(inputs=h,
units=1,
activation=None,
kernel_initializer='glorot_uniform')
norm_val_tf = tf.reshape(norm_val_tf, [-1])
val_tf = self.val_norm.unnormalize_tf(norm_val_tf)
return val_tf
def _build_net_actor(self, net_name, input_tfs, init_output_scale, reuse=False):
with tf.variable_scope('actor', reuse=reuse):
h = NetBuilder.build_net(net_name, input_tfs, reuse)
std_type = TFDistributionGaussianDiag.StdType.Default
a_size = self.get_action_size()
mean_kernel_init = tf.random_uniform_initializer(minval=-init_output_scale, maxval=init_output_scale)
mean_bias_init = tf.zeros_initializer()
logstd_kernel_init = tf.random_uniform_initializer(minval=-init_output_scale, maxval=init_output_scale)
logstd_bias_init = np.log(self.exp_params_curr.noise) * np.ones(a_size)
logstd_bias_init = logstd_bias_init.astype(np.float32)
norm_a_pd_tf = TFDistributionGaussianDiag(input=h, dim=a_size, std_type=std_type,
mean_kernel_init=mean_kernel_init, mean_bias_init=mean_bias_init,
logstd_kernel_init=logstd_kernel_init, logstd_bias_init=logstd_bias_init,
reuse=reuse)
return norm_a_pd_tf