def build(self, input_tensor, net):
with self.scope():
input_tensor = utils.ensure_tensor_dimensionality(input_tensor, 2)
self.labels_placeholder = tf.placeholder(
tf.float32, shape=(None, int(input_tensor.shape[1])), name='Y')
loss = tf.reduce_mean(
self.activation(labels=self.labels_placeholder, logits=input_tensor))
for layer in net.layers:
loss = layer.contribute_to_loss(loss)
output = tf.identity(loss, 'loss')
self.register_train_summary(tf.summary.scalar('train_loss', loss))
self.register_test_summary(tf.summary.scalar('test_loss', loss))
with self.scope('accuracy_layer'):
correct_prediction = tf.equal(
tf.argmax(self.labels_placeholder, 1), tf.argmax(input_tensor, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name='accuracy')
self.register_train_summary(
tf.summary.scalar('train_accuracy', accuracy))
self.register_test_summary(tf.summary.scalar('test_accuracy', accuracy))
return output
def verify_input_tensor_dimensionality(self, input_tensor):
if self.input_tensor_shape_list != input_tensor.shape.as_list():
raise ValueError(
"Tensors' shape during building and applying are different: build={}, "
"apply={}".format(self.input_tensor_shape_list,
input_tensor.shape.as_list()))
return utils.ensure_tensor_dimensionality(input_tensor,
self.n_dimensions)
def apply_reverse(self, input_tensor):
if self.weights is None or self.bias is None:
raise RuntimeError('Layer is not built')
with self.reverse_scope():
input_tensor = utils.ensure_tensor_dimensionality(input_tensor, 2)
output = tf.nn.bias_add(input_tensor, tf.negative(self.bias))
output = tf.matmul(output, self.weights, transpose_b=True)
if self.activation:
output = self.activation(output)
return self.restore_dimensionality_after_reverse(output)
def apply_reverse(self, input_tensor):
if self.weights is None or self.bias is None:
raise RuntimeError('Layer is not built')
with self.reverse_scope():
input_tensor = utils.ensure_tensor_dimensionality(input_tensor, 4)
output = tf.nn.bias_add(input_tensor, tf.negative(self.bias))
output_shape = tf.stack([tf.shape(input_tensor)[0]] + self.input_shape[1:])
output = tf.nn.conv2d_transpose(output, self.weights, output_shape=output_shape,
strides=self.strides, padding='SAME')
if self.activation:
output = self.activation(output)
return self.restore_dimensionality_after_reverse(output)
def build(self, input_tensor, net):
with self.scope():
input_tensor = utils.ensure_tensor_dimensionality(input_tensor, 2)
loss = tf.reduce_mean(
tf.squared_difference(net.input_layer.input_placeholder, input_tensor), name='loss')
for layer in net.layers:
loss = layer.contribute_to_loss(loss)
output = tf.identity(loss, 'loss')
self.register_train_summary(tf.summary.scalar('train_loss', loss))
self.register_test_summary(tf.summary.scalar('test_loss', loss))
# TODO: this should not be here, probably in the network
self.register_test_summary(
tf.summary.image(
'input', tf.reshape(input_tensor, [-1, 28, 28, 1])))
return output
def apply(self, input_tensor):
input_tensor = utils.ensure_tensor_dimensionality(input_tensor, 4)
return tf.nn.max_pool(input_tensor, self.size, strides=[1, 2, 2, 1], padding='SAME')