Skip to content
This repository was archived by the owner on Jun 23, 2025. It is now read-only.

Commit 87e4d0f

Browse files
committed
minor documentary adjustments
1 parent 088698a commit 87e4d0f

File tree

1 file changed

+6
-6
lines changed

1 file changed

+6
-6
lines changed

keras_contrib/wrappers/cdropout.py

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -36,7 +36,7 @@ class ConcreteDropout(Wrapper):
3636
prob_init: Tuple[float, float].
3737
Probability lower / upper bounds of dropout rate initialization.
3838
temp: float. Temperature.
39-
Determines the speed of probability adjustments.
39+
Determines the speed of probability (i.e. dropout rate) adjustments.
4040
seed: Seed for random probability sampling.
4141
4242
# References
@@ -74,6 +74,7 @@ def _concrete_dropout(self, inputs, layer_type):
7474
# Returns
7575
A tensor with the same shape as inputs and dropout applied.
7676
"""
77+
assert layer_type in {'dense', 'conv2d'}
7778
eps = K.cast_to_floatx(K.epsilon())
7879

7980
noise_shape = K.shape(inputs)
@@ -93,6 +94,7 @@ def _concrete_dropout(self, inputs, layer_type):
9394
)
9495
drop_prob = K.sigmoid(drop_prob / self.temp)
9596

97+
# apply dropout
9698
random_tensor = 1. - drop_prob
9799
retain_prob = 1. - self.p
98100
inputs *= random_tensor
@@ -104,7 +106,7 @@ def build(self, input_shape=None):
104106
input_shape = to_tuple(input_shape)
105107
if len(input_shape) == 2: # Dense_layer
106108
input_dim = np.prod(input_shape[-1]) # we drop only last dim
107-
elif len(input_shape) == 4: # Conv_layer
109+
elif len(input_shape) == 4: # Conv2D_layer
108110
input_dim = (input_shape[1]
109111
if K.image_data_format() == 'channels_first'
110112
else input_shape[3]) # we drop only channels
@@ -129,7 +131,7 @@ def build(self, input_shape=None):
129131

130132
super(ConcreteDropout, self).build(input_shape)
131133

132-
# initialize regularizer / prior KL term
134+
# initialize regularizer / prior KL term and add to layer-loss
133135
weight = self.layer.kernel
134136
kernel_regularizer = (
135137
self.weight_regularizer
@@ -146,9 +148,7 @@ def build(self, input_shape=None):
146148
def call(self, inputs, training=None):
147149
def relaxed_dropped_inputs():
148150
return self.layer.call(self._concrete_dropout(inputs, (
149-
'dense'
150-
if len(K.int_shape(inputs)) == 2
151-
else 'conv2d'
151+
'dense' if len(K.int_shape(inputs)) == 2 else 'conv2d'
152152
)))
153153

154154
return K.in_train_phase(relaxed_dropped_inputs,

0 commit comments

Comments
 (0)