def accuracy_op(self):
clear_lazyprop_on_lazyprop_cleared(self, 'accuracy_op',
self.input_layer)
return tf.reduce_mean(
tf.cast(
tf.nn.in_top_k(self._pre_softmax_activation_predict,
tf.argmax(self.target_placeholder, 1), 1),
tf.float32))
def _pre_softmax_activation_train(self):
clear_lazyprop_on_lazyprop_cleared(self,
'_pre_softmax_activation_train',
self.input_layer,
'activation_train')
with self.name_scope(is_train=True):
input_activation = self._process_input_activation_train(
self.input_layer.activation_train)
return self._layer_activation(input_activation, True)
def _pre_softmax_activation_predict(self):
clear_lazyprop_on_lazyprop_cleared(self,
'_pre_softmax_activation_predict',
self.input_layer,
'activation_predict')
with self.name_scope(is_predict=True):
input_activation = self._process_input_activation_predict(
self.input_layer.activation_predict)
return self._layer_activation(input_activation, False)
def log_probability_of_targets_op(self):
clear_lazyprop_on_lazyprop_cleared(self,
'log_probability_of_targets_op',
self.input_layer)
return tf.reduce_sum(
tf.log(
tf.reduce_sum(
tf.nn.softmax(self.activation_predict) *
self.target_placeholder, 1)))
def bactivation_train(self):
"""The activation used for training this layer, this will often be the same as prediction except with dropout or
random noise applied.
Returns:
tensorflow.Tensor
"""
clear_lazyprop_on_lazyprop_cleared(self, 'bactivation_train', self, 'activation_train')
return self._layer_bactivation(self.activation_train, True)
def bactivation_predict(self):
"""The activation used for predictions from this layer, this will often be the same as training except without
dropout or random noise applied.
Returns:
tensorflow.Tensor
"""
clear_lazyprop_on_lazyprop_cleared(self, 'bactivation_predict', self, 'activation_predict')
return self._layer_bactivation(self.activation_predict, False)
def gradients_with_respect_to_error_op(self):
clear_lazyprop_on_lazyprop_cleared(self, "gradients_with_respect_to_error_op",
self.last_layer, "target_loss_op_predict")
gradients_ops = []
for variable in self.resizable_variables:
gradients_ops.append(tf.gradients(self.last_layer.target_loss_op_predict, variable)[0])
return gradients_ops
def hessien_with_respect_to_error_op(self):
clear_lazyprop_on_lazyprop_cleared(self, "hessien_with_respect_to_error_op",
self, "gradients_with_respect_to_error_op")
hessian_ops = []
for variable, gradients in zip(self.resizable_variables, self.gradients_with_respect_to_error_op):
hessian_ops.append(tf.gradients(gradients, variable)[0])
# TODO: use tf.hessian in tensorflow 1. also use tf.diag_part
return hessian_ops
def activation_predict(self):
"""The activation used for predictions from this layer, this will often be the same as training except without
dropout or random noise applied.
Returns:
tensorflow.Tensor
"""
clear_lazyprop_on_lazyprop_cleared(self, 'activation_predict', self.input_layer)
input_tensor = self.input_layer.activation_predict
with self.name_scope(is_predict=True):
input_tensor = self._process_input_activation_predict(input_tensor)
return self._layer_activation(input_tensor, False)
def activation_train(self):
"""The activation used for training this layer, this will often be the same as prediction except with dropout or
random noise applied.
Returns:
tensorflow.Tensor
"""
clear_lazyprop_on_lazyprop_cleared(self, 'activation_train', self.input_layer)
input_tensor = self.input_layer.activation_train
with self.name_scope(is_train=True):
input_tensor = self._process_input_activation_train(input_tensor)
return self._layer_activation(input_tensor, True)
def test_clear_lazyprop_on_lazyprop_cleared(self):
prop_class_1 = _PropClass()
prop_class_2 = _PropClass()
clear_lazyprop_on_lazyprop_cleared(prop_class_2, 'lazyprop',
prop_class_1, 'lazyprop')
prop_class_1.STATIC_VAL = 1
prop_class_2.STATIC_VAL = 2
self.assertEqual(prop_class_1.lazyprop, 1)
self.assertEqual(prop_class_2.lazyprop, 2)
prop_class_1.STATIC_VAL = 3
prop_class_2.STATIC_VAL = 4
clear_lazyprop(prop_class_1, 'lazyprop')
self.assertEqual(prop_class_1.lazyprop, 3)
self.assertEqual(prop_class_2.lazyprop, 4)
def target_loss_op_predict(self):
clear_lazyprop_on_lazyprop_cleared(self, 'target_loss_op_predict',
self.input_layer)
with self.name_scope(is_predict=True):
return self._target_loss_op(self._pre_softmax_activation_predict)
def activation_train(self):
clear_lazyprop_on_lazyprop_cleared(self, 'activation_train',
self.input_layer,
'activation_train')
with self.name_scope(is_train=True):
return tf.nn.softmax(self._pre_softmax_activation_train)