res_auto.py 文件源码

def _bn_relu_deconv(nb_filter, nb_row, nb_col, subsample, output_shape):
    def f(x):
        norm = BatchNormalization(mode=2, axis=3)(x)
        activation = Activation("relu")(norm)
        return Deconvolution2D(
            nb_filter, nb_row, nb_col, W_regularizer=l2(1e-4),
            subsample=subsample, output_shape=output_shape,
            init="he_normal", border_mode="same")(activation)
    return f


# Bottleneck architecture for > 34 layer resnet.
# Follows improved proposed scheme in http://arxiv.org/pdf/1603.05027v2.pdf
# Returns a final conv layer of nb_filters * 4
# def _bottleneck(nb_filters, init_subsample=(1, 1)):
#     def f(x):
#         conv_1_1 = _bn_relu_conv(nb_filters, 1, 1, subsample=init_subsample)(x)
#         conv_3_3 = _bn_relu_conv(nb_filters, 3, 3)(conv_1_1)
#         residual = _bn_relu_conv(nb_filters * 4, 1, 1)(conv_3_3)
#         return _shortcut(x, residual)
#     return f


# Basic 3 X 3 convolution blocks.
# Use for resnet with layers <= 34
# Follows improved proposed scheme in http://arxiv.org/pdf/1603.05027v2.pdf