Skip to content

Fix #23 #48

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Fix #23 #48

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
75d426a
specify epsilon of batch normalization
Mar 28, 2019
7554ca2
change addition target in identity block
Mar 28, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 12 additions & 9 deletions keras_vggface/models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -339,6 +339,7 @@ def senet_conv_block(input_tensor, kernel_size, filters,
bn_axis = 3
else:
bn_axis = 1
bn_eps = 0.0001

conv1_reduce_name = 'conv' + str(stage) + "_" + str(block) + "_1x1_reduce"
conv1_increase_name = 'conv' + str(stage) + "_" + str(
Expand All @@ -348,22 +349,22 @@ def senet_conv_block(input_tensor, kernel_size, filters,

x = Conv2D(filters1, (1, 1), use_bias=bias, strides=strides,
name=conv1_reduce_name)(input_tensor)
x = BatchNormalization(axis=bn_axis, name=conv1_reduce_name + "/bn")(x)
x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_reduce_name + "/bn")(x)
x = Activation('relu')(x)

x = Conv2D(filters2, kernel_size, padding='same', use_bias=bias,
name=conv3_name)(x)
x = BatchNormalization(axis=bn_axis, name=conv3_name + "/bn")(x)
x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv3_name + "/bn")(x)
x = Activation('relu')(x)

x = Conv2D(filters3, (1, 1), name=conv1_increase_name, use_bias=bias)(x)
x = BatchNormalization(axis=bn_axis, name=conv1_increase_name + "/bn")(x)
x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_increase_name + "/bn")(x)

se = senet_se_block(x, stage=stage, block=block, bias=True)

shortcut = Conv2D(filters3, (1, 1), use_bias=bias, strides=strides,
name=conv1_proj_name)(input_tensor)
shortcut = BatchNormalization(axis=bn_axis,
shortcut = BatchNormalization(axis=bn_axis, epsilon=bn_eps,
name=conv1_proj_name + "/bn")(shortcut)

m = layers.add([se, shortcut])
Expand All @@ -378,6 +379,7 @@ def senet_identity_block(input_tensor, kernel_size,
bn_axis = 3
else:
bn_axis = 1
bn_eps = 0.0001

conv1_reduce_name = 'conv' + str(stage) + "_" + str(block) + "_1x1_reduce"
conv1_increase_name = 'conv' + str(stage) + "_" + str(
Expand All @@ -386,20 +388,20 @@ def senet_identity_block(input_tensor, kernel_size,

x = Conv2D(filters1, (1, 1), use_bias=bias,
name=conv1_reduce_name)(input_tensor)
x = BatchNormalization(axis=bn_axis, name=conv1_reduce_name + "/bn")(x)
x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_reduce_name + "/bn")(x)
x = Activation('relu')(x)

x = Conv2D(filters2, kernel_size, padding='same', use_bias=bias,
name=conv3_name)(x)
x = BatchNormalization(axis=bn_axis, name=conv3_name + "/bn")(x)
x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv3_name + "/bn")(x)
x = Activation('relu')(x)

x = Conv2D(filters3, (1, 1), name=conv1_increase_name, use_bias=bias)(x)
x = BatchNormalization(axis=bn_axis, name=conv1_increase_name + "/bn")(x)
x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name=conv1_increase_name + "/bn")(x)

se = senet_se_block(x, stage=stage, block=block, bias=True)

m = layers.add([x, se])
m = layers.add([se, input_tensor])
m = Activation('relu')(m)

return m
Expand Down Expand Up @@ -427,11 +429,12 @@ def SENET50(include_top=True, weights='vggface',
bn_axis = 3
else:
bn_axis = 1
bn_eps = 0.0001

x = Conv2D(
64, (7, 7), use_bias=False, strides=(2, 2), padding='same',
name='conv1/7x7_s2')(img_input)
x = BatchNormalization(axis=bn_axis, name='conv1/7x7_s2/bn')(x)
x = BatchNormalization(axis=bn_axis, epsilon=bn_eps, name='conv1/7x7_s2/bn')(x)
x = Activation('relu')(x)
x = MaxPooling2D((3, 3), strides=(2, 2))(x)

Expand Down