dm_resnet.py 文件源码

def _shortcut(input, residual, weight_decay=.0001, dropout=.0):
    # Expand channels of shortcut to match residual.
    # Stride appropriately to match residual (width, height)
    # Should be int if network architecture is correctly configured.
    # !!! The dropout argument is just a place holder. 
    # !!! It shall not be applied to identity mapping.
    stride_width = input._keras_shape[ROW_AXIS] // residual._keras_shape[ROW_AXIS]
    stride_height = input._keras_shape[COL_AXIS] // residual._keras_shape[COL_AXIS]
    equal_channels = residual._keras_shape[CHANNEL_AXIS] == input._keras_shape[CHANNEL_AXIS]

    shortcut = input
    # 1 X 1 conv if shape is different. Else identity.
    if stride_width > 1 or stride_height > 1 or not equal_channels:
        shortcut = Convolution2D(nb_filter=residual._keras_shape[CHANNEL_AXIS],
                                 nb_row=1, nb_col=1,
                                 subsample=(stride_width, stride_height),
                                 init="he_normal", border_mode="valid", 
                                 W_regularizer=l2(weight_decay))(input)

    return merge([shortcut, residual], mode="sum")


# Builds a residual block with repeating bottleneck blocks.