def __init__(self, x, ac_space, trainable):
self.logstd = tf.get_variable(name='logstd',
shape=[1] + list(ac_space.shape),
initializer=tf.zeros_initializer(),
trainable=trainable)
self.mean = dense_layer(x, ac_space.shape[0], None, True, trainable,
'pol_mean', (True and trainable), False, False)
self.std = tf.exp(self.logstd)
if trainable:
variable_summaries(self.logstd, 'pol_logstd')
def dense_layer(input_layer,
nn_num,
act,
use_bias,
trainable,
scope,
monitor_weight=False,
monitor_bias=False,
monitor_output=False):
size = input_layer.shape[-1]
with tf.variable_scope(scope, reuse=False):
w = weight_variable(size, nn_num, trainable)
if monitor_weight:
variable_summaries(w, 'weights')
mul = tf.matmul(input_layer, w)
if use_bias:
b = bias_variable(nn_num, trainable)
if monitor_bias:
variable_summaries(b, 'bias')
with tf.name_scope('pre_active'):
preactive = tf.add(mul, b)
else:
with tf.name_scope('pre_active'):
preactive = mul
with tf.name_scope('activation'):
if (act != None):
activation = act(preactive)
else:
activation = preactive
if monitor_output:
variable_summaries(activation, 'activated')
return activation
def fc(input_layer,
nn_num,
use_bias,
trainable,
scope,
monitor_weight=False,
monitor_bias=False,
monitor_output=False):
size = input_layer.shape[-1]
with tf.variable_scope(scope, reuse=False):
w = weight_variable(size, nn_num, trainable)
if monitor_weight:
variable_summaries(w, 'weights')
mul = tf.matmul(input_layer, w)
if use_bias:
b = bias_variable(nn_num, trainable)
if monitor_bias:
variable_summaries(b, 'bias')
with tf.name_scope('pre_active'):
preactive = tf.add(mul, b)
else:
with tf.name_scope('pre_active'):
preactive = mul
if monitor_output:
variable_summaries(preactive, 'preactive')
return preactive