wassname/keras-contrib
1
0
Fork 0
mirror of https://github.com/wassname/keras-contrib.git synced 2026-08-17 11:20:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

Refactored conv_blocks

Browse Source
This commit is contained in:
Somshubra Majumdar
2017-02-15 13:56:00 -06:00
parent fe7c6df487
commit c425cdf3d6
1 changed files with 13 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files
keras_contrib/applications/ror.py
+13 -45
View File
@@ -156,12 +156,7 @@ def ResidualOfResidual(depth=40, width=2, dropout_rate=0.0,
return model
def __initial_conv(input):
x = Convolution2D(16, 3, 3, border_mode='same', init='he_normal')(input)
return x
def __conv1_block(input, k=1, dropout=0.0, initial=False):
def __initial_conv_block(input, k=1, dropout=0.0, initial=False):
init = input
channel_axis = 1 if K.image_dim_ordering() == "th" else -1
@@ -188,57 +183,29 @@ def __conv1_block(input, k=1, dropout=0.0, initial=False):
return m
def __conv2_block(input, k=1, dropout=0.0):
def __conv_block(input, nb_filters=32, k=1, dropout=0.0):
init = input
channel_axis = 1 if K.image_dim_ordering() == "th" else -1
# Check if input number of filters is same as 32 * k, else create convolution2d for this input
if K.image_dim_ordering() == "th":
if init._keras_shape[1] != 32 * k:
init = Convolution2D(32 * k, 1, 1, init='he_normal', border_mode='same')(init)
if init._keras_shape[1] != nb_filters * k:
init = Convolution2D(nb_filters * k, 1, 1, init='he_normal', border_mode='same')(init)
else:
if init._keras_shape[-1] != 32 * k:
init = Convolution2D(32 * k, 1, 1, init='he_normal', border_mode='same')(init)
if init._keras_shape[-1] != nb_filters * k:
init = Convolution2D(nb_filters * k, 1, 1, init='he_normal', border_mode='same')(init)
x = BatchNormalization(axis=channel_axis)(input)
x = Activation('relu')(x)
x = Convolution2D(32 * k, 3, 3, border_mode='same', init='he_normal')(x)
x = Convolution2D(nb_filters * k, 3, 3, border_mode='same', init='he_normal')(x)
if dropout > 0.0:
x = Dropout(dropout)(x)
x = BatchNormalization(axis=channel_axis)(x)
x = Activation('relu')(x)
x = Convolution2D(32 * k, 3, 3, border_mode='same', init='he_normal')(x)
m = merge([init, x], mode='sum')
return m
def __conv3_block(input, k=1, dropout=0.0):
init = input
channel_axis = 1 if K.image_dim_ordering() == "th" else -1
# Check if input number of filters is same as 64 * k, else create convolution2d for this input
if K.image_dim_ordering() == "th":
if init._keras_shape[1] != 64 * k:
init = Convolution2D(64 * k, 1, 1, init='he_normal', border_mode='same')(init)
else:
if init._keras_shape[-1] != 64 * k:
init = Convolution2D(64 * k, 1, 1, init='he_normal', border_mode='same')(init)
x = BatchNormalization(axis=channel_axis)(input)
x = Activation('relu')(x)
x = Convolution2D(64 * k, 3, 3, border_mode='same', init='he_normal')(x)
if dropout > 0.0:
x = Dropout(dropout)(x)
x = BatchNormalization(axis=channel_axis)(x)
x = Activation('relu')(x)
x = Convolution2D(64 * k, 3, 3, border_mode='same', init='he_normal')(x)
x = Convolution2D(nb_filters * k, 3, 3, border_mode='same', init='he_normal')(x)
m = merge([init, x], mode='sum')
return m
@@ -270,7 +237,8 @@ def __create_pre_residual_of_residual(nb_classes, img_input, include_top, depth=
channel_axis = 1 if K.image_dim_ordering() == "th" else -1
x = __initial_conv(img_input)
# Initial convolution layer
x = Convolution2D(16, 3, 3, border_mode='same', init='he_normal')(img_input)
nb_conv = 4 # Dont count 4 long residual connections in WRN models
conv0_level1_shortcut = Convolution2D(64 * width, 1, 1, border_mode='same', subsample=(4, 4),
@@ -280,7 +248,7 @@ def __create_pre_residual_of_residual(nb_classes, img_input, include_top, depth=
name='conv1_level2_shortcut')(x)
for i in range(N):
initial = (i == 0)
x = __conv1_block(x, k=width, dropout=dropout, initial=initial)
x = __initial_conv_block(x, k=width, dropout=dropout, initial=initial)
nb_conv += 2
# Add Level 2 shortcut
@@ -291,7 +259,7 @@ def __create_pre_residual_of_residual(nb_classes, img_input, include_top, depth=
conv2_level2_shortcut = Convolution2D(32 * width, 1, 1, border_mode='same',
name='conv2_level2_shortcut')(x)
for i in range(N):
x = __conv2_block(x, k=width, dropout=dropout)
x = __conv_block(x, k=width, dropout=dropout)
nb_conv += 2
# Add Level 2 shortcut
@@ -302,7 +270,7 @@ def __create_pre_residual_of_residual(nb_classes, img_input, include_top, depth=
conv3_level2_shortcut = Convolution2D(64 * width, 1, 1, border_mode='same',
name='conv3_level2_shortcut')(x)
for i in range(N):
x = __conv3_block(x, k=width, dropout=dropout)
x = __conv_block(x, nb_filters=64, k=width, dropout=dropout)
nb_conv += 2
# Add Level 2 shortcut