Policy gradient example: record stats for tensorboard (#577)

* add tf metrics

* comments

* fix network scopes

* add doc

* use format string

* fix trace level

* plot intermediate and final sgd stats

* add back a global step
This commit is contained in:
Eric Liang
2017-05-21 14:51:24 -07:00
committed by Philipp Moritz
parent c440010cbd
commit 06241daf61
7 changed files with 113 additions and 45 deletions
@@ -17,21 +17,22 @@ def normc_initializer(std=1.0):
def fc_net(inputs, num_classes=10, logstd=False):
fc1 = slim.fully_connected(inputs, 128,
weights_initializer=normc_initializer(1.0),
scope="fc1")
fc2 = slim.fully_connected(fc1, 128,
weights_initializer=normc_initializer(1.0),
scope="fc2")
fc3 = slim.fully_connected(fc2, 128,
weights_initializer=normc_initializer(1.0),
scope="fc3")
fc4 = slim.fully_connected(fc3, num_classes,
weights_initializer=normc_initializer(0.01),
activation_fn=None, scope="fc4")
if logstd:
logstd = tf.get_variable(name="logstd", shape=[num_classes],
initializer=tf.zeros_initializer)
return tf.concat(1, [fc4, logstd])
else:
return fc4
with tf.name_scope("fc_net") as net:
fc1 = slim.fully_connected(inputs, 128,
weights_initializer=normc_initializer(1.0),
scope=net + "fc1")
fc2 = slim.fully_connected(fc1, 128,
weights_initializer=normc_initializer(1.0),
scope=net + "fc2")
fc3 = slim.fully_connected(fc2, 128,
weights_initializer=normc_initializer(1.0),
scope=net + "fc3")
fc4 = slim.fully_connected(fc3, num_classes,
weights_initializer=normc_initializer(0.01),
activation_fn=None, scope=net + "fc4")
if logstd:
logstd = tf.get_variable(name="logstd", shape=[num_classes],
initializer=tf.zeros_initializer)
return tf.concat(1, [fc4, logstd])
else:
return fc4