def center_bias(x, img_rows, img_cols, downsampling_factor_net, downsampling_factor_product, out_name=None, elt_name=None):
rows_elt = math.ceil(img_rows / downsampling_factor_net) // downsampling_factor_product
cols_elt = math.ceil(img_cols / downsampling_factor_net) // downsampling_factor_product
eltprod = EltWiseProduct(init='zero',
trainable=True,
W_regularizer=l2(1/(rows_elt*cols_elt*N_CAT)),
name=elt_name)(x)
output_ml_net = Activation('relu', name=out_name)(eltprod)
return output_ml_net
def ml_net_model(img_rows=480,
img_cols=640,
downsampling_factor_net=8,
downsampling_factor_product=10):
f = h5py.File("vgg16_weights.h5")
input_ml_net = Input(shape=(3, img_rows, img_cols))
#########################################################
# FEATURE EXTRACTION NETWORK #
#########################################################
weights = get_weights_vgg16(f, 1)
conv1_1 = Convolution2D(64,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(input_ml_net)
weights = get_weights_vgg16(f, 3)
conv1_2 = Convolution2D(64,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv1_1)
conv1_pool = MaxPooling2D((2, 2), strides=(2, 2),
border_mode='same')(conv1_2)
weights = get_weights_vgg16(f, 6)
conv2_1 = Convolution2D(128,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv1_pool)
weights = get_weights_vgg16(f, 8)
conv2_2 = Convolution2D(128,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv2_1)
conv2_pool = MaxPooling2D((2, 2), strides=(2, 2),
border_mode='same')(conv2_2)
weights = get_weights_vgg16(f, 11)
conv3_1 = Convolution2D(256,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv2_pool)
weights = get_weights_vgg16(f, 13)
conv3_2 = Convolution2D(256,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv3_1)
weights = get_weights_vgg16(f, 15)
conv3_3 = Convolution2D(256,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv3_2)
conv3_pool = MaxPooling2D((2, 2), strides=(2, 2),
border_mode='same')(conv3_3)
weights = get_weights_vgg16(f, 18)
conv4_1 = Convolution2D(512,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv3_pool)
weights = get_weights_vgg16(f, 20)
conv4_2 = Convolution2D(512,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv4_1)
weights = get_weights_vgg16(f, 22)
conv4_3 = Convolution2D(512,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv4_2)
conv4_pool = MaxPooling2D((2, 2), strides=(1, 1),
border_mode='same')(conv4_3)
weights = get_weights_vgg16(f, 25)
conv5_1 = Convolution2D(512,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv4_pool)
weights = get_weights_vgg16(f, 27)
conv5_2 = Convolution2D(512,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv5_1)
weights = get_weights_vgg16(f, 29)
conv5_3 = Convolution2D(512,
3,
3,
weights=weights,
activation='relu',
border_mode='same')(conv5_2)
#########################################################
# ENCODING NETWORK #
#########################################################
concatenated = merge([conv3_pool, conv4_pool, conv5_3],
mode='concat',
concat_axis=1)
dropout = Dropout(0.5)(concatenated)
int_conv = Convolution2D(64,
3,
3,
init='glorot_normal',
activation='relu',
border_mode='same')(dropout)
pre_final_conv = Convolution2D(1,
1,
1,
init='glorot_normal',
activation='relu')(int_conv)
#########################################################
# PRIOR LEARNING #
#########################################################
rows_elt = math.ceil(
img_rows / downsampling_factor_net) // downsampling_factor_product
cols_elt = math.ceil(
img_cols / downsampling_factor_net) // downsampling_factor_product
eltprod = EltWiseProduct(init='zero',
W_regularizer=l2(
1 / (rows_elt * cols_elt)))(pre_final_conv)
output_ml_net = Activation('relu')(eltprod)
model = Model(input=[input_ml_net], output=[output_ml_net])
for layer in model.layers:
print(layer.input_shape, layer.output_shape)
return model
def build_model(img_rows=480,
img_cols=640,
downsampling_factor_net=8,
downsampling_factor_product=10,
use_pretrained=None,
intermediate=False):
input_ml_net = Input(shape=(3, img_rows, img_cols))
conv1_1 = VGGConvLayer(64, vgg_layer=1)(input_ml_net)
conv1_2 = VGGConvLayer(64, vgg_layer=3)(conv1_1)
conv1_pool = MaxPooling2D((2, 2), strides=(2, 2),
border_mode='same')(conv1_2)
conv2_1 = VGGConvLayer(128, vgg_layer=6)(conv1_pool)
conv2_2 = VGGConvLayer(128, vgg_layer=8)(conv2_1)
conv2_pool = MaxPooling2D((2, 2), strides=(2, 2),
border_mode='same')(conv2_2)
conv3_1 = VGGConvLayer(256, vgg_layer=11)(conv2_pool)
conv3_2 = VGGConvLayer(256, vgg_layer=13)(conv3_1)
conv3_3 = VGGConvLayer(256, vgg_layer=15)(conv3_2)
conv3_pool = MaxPooling2D((2, 2), strides=(2, 2),
border_mode='same')(conv3_3)
conv4_1 = VGGConvLayer(512, vgg_layer=18)(conv3_pool)
conv4_2 = VGGConvLayer(512, vgg_layer=20)(conv4_1)
conv4_3 = VGGConvLayer(512, vgg_layer=22)(conv4_2)
conv4_pool = MaxPooling2D((2, 2), strides=(1, 1),
border_mode='same')(conv4_3)
conv5_1 = VGGConvLayer(512, vgg_layer=25)(conv4_pool)
conv5_2 = VGGConvLayer(512, vgg_layer=27)(conv5_1)
conv5_3 = VGGConvLayer(512, vgg_layer=29)(conv5_2)
concatenated = merge([conv3_pool, conv4_pool, conv5_3],
mode='concat',
concat_axis=1)
dropout = Dropout(0.5)(concatenated)
int_conv = Convolution2D(64,
3,
3,
init='glorot_normal',
activation='relu',
border_mode='same',
trainable=True,
name='int_conv')(dropout)
pre_final_conv = Convolution2D(1,
1,
1,
init='glorot_normal',
activation='relu',
name='pre_final_conv')(int_conv)
rows_elt = math.ceil(
img_rows / downsampling_factor_net) // downsampling_factor_product
cols_elt = math.ceil(
img_cols / downsampling_factor_net) // downsampling_factor_product
eltprod = EltWiseProduct(init='zero',
trainable=True,
W_regularizer=l2(1 / (rows_elt * cols_elt)),
name='eltprod')(pre_final_conv)
output_ml_net = Activation('relu')(eltprod)
model = Model(input=[input_ml_net], output=[output_ml_net])
# load pretrained weights
if use_pretrained == 'mlnet':
model.load_weights('models/mlnet_salicon_weights.pkl')
elif use_pretrained == 'my':
model.load_weights('models/generalist_finetune.h5')
elif use_pretrained is None:
return model
else:
raise StandardError('INVALID use_pretrained')
if intermediate:
# still has correct weights
model = Model(input=[input_ml_net], output=[concatenated])
return model