def get_loss(self, xb1, xb2):
feed = {
tbn('xb1:0'): xb1,
tbn('xb2:0'): xb2,
tbn('is_training:0'): False
}
losses = self.sess.run(tf.get_collection('losses'), feed_dict=feed)
lstring = ' '.join(['{:.3f}'.format(loss) for loss in losses])
return lstring
def train(self, xb1, xb2):
"""Take a training step with batches from each domain."""
self.iteration += 1
feed = {
tbn('xb1:0'): xb1,
tbn('xb2:0'): xb2,
tbn('lr:0'): self.learning_rate,
tbn('is_training:0'): True
}
_ = self.sess.run([obn('train_op_G')], feed_dict=feed)
_ = self.sess.run([obn('train_op_D')], feed_dict=feed)
def train(self, x1=None, x2=None):
self.iteration += 1
feed = {
tbn('lr:0'): self.args.learning_rate,
tbn('is_training:0'): True
}
if x1 is not None:
feed[tbn('xb1:0')] = x1
feed[tbn('xb2:0')] = x2
self.sess.run([obn('train_op_D')], feed_dict=feed)
self.sess.run([obn('train_op_G')], feed_dict=feed)
def get_layer(self, xb1, xb2, name=None):
tensor_name = "{}:0".format(name)
tensor = tbn(tensor_name)
feed = {
tbn('xb1:0'): xb1,
tbn('xb2:0'): xb2,
tbn('is_training:0'): False
}
layer = self.sess.run(tensor, feed_dict=feed)
return layer
def get_loss(self, xb1, xb2):
"""Return all of the loss values for the given input."""
feed = {
tbn('xb1:0'): xb1,
tbn('xb2:0'): xb2,
tbn('is_training:0'): False
}
ls = [tns for tns in tf.get_collection('losses')]
losses = self.sess.run(ls, feed_dict=feed)
lstring = ' '.join(['{:.3f}'.format(loss) for loss in losses])
return lstring
def get_layer(self, xb1, xb2, name):
"""Get a layer of the network by name for the entire datasets given in xb1 and xb2."""
tensor_name = "{}:0".format(name)
tensor = tbn(tensor_name)
feed = {
tbn('xb1:0'): xb1,
tbn('xb2:0'): xb2,
tbn('is_training:0'): False
}
layer = self.sess.run(tensor, feed_dict=feed)
return layer
def _restore(self, restore_folder, limit_gpu_fraction, no_gpu=False):
tf.reset_default_graph()
self.init_session(limit_gpu_fraction, no_gpu)
ckpt = tf.train.get_checkpoint_state(restore_folder)
self.saver = tf.train.import_meta_graph('{}.meta'.format(
ckpt.model_checkpoint_path))
self.saver.restore(self.sess, ckpt.model_checkpoint_path)
if self.x1 and self.x2:
self.sess.run(obn('initializerx1'),
feed_dict={tbn('datasetx1ph:0'): self.x1})
self.sess.run(obn('initializerx2'),
feed_dict={tbn('datasetx2ph:0'): self.x2})
self.iteration = 0
print("Model restored from {}".format(restore_folder))
def get_layer(self, xb1, xb2, name=None):
tensor_name = "{}:0".format(name)
tensor = tbn(tensor_name)
feed = {
tbn('xb1:0'): xb1,
tbn('xb2:0'): xb2,
tbn('is_training:0'): False
}
print('tensor shape', tensor.shape)
layer = self.sess.run(tensor, feed_dict=feed)
print('layer from get_layer', layer)
return layer
def init_session(self, limit_gpu_fraction=.4, no_gpu=False):
if no_gpu:
config = tf.ConfigProto(device_count={'GPU': 0})
self.sess = tf.Session(config=config)
elif limit_gpu_fraction:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=limit_gpu_fraction)
config = tf.ConfigProto(gpu_options=gpu_options)
self.sess = tf.Session(config=config)
else:
self.sess = tf.Session()
if not self.args.restore_folder:
self.sess.run(obn('initializerx1'),
feed_dict={tbn('datasetx1ph:0'): self.x1})
self.sess.run(obn('initializerx2'),
feed_dict={tbn('datasetx2ph:0'): self.x2})
