keras_training.py 文件源码

def basic_block(filters, init_strides=(1, 1), is_first_block_of_first_layer=False):
    """Basic 3 X 3 convolution blocks for use on resnets with layers <= 34.
    Follows improved proposed scheme in http://arxiv.org/pdf/1603.05027v2.pdf
    """
    def f(input):

        if is_first_block_of_first_layer:
            # don't repeat bn->relu since we just did bn->relu->maxpool
            conv1 = Conv2D(filters=filters, kernel_size=(3, 3),
                           strides=init_strides,
                           padding="same",
                           kernel_initializer="he_normal",
                           kernel_regularizer=l2(1e-4))(input)
        else:
            conv1 = _bn_relu_conv(filters=filters, kernel_size=(3, 3),
                                  strides=init_strides)(input)

        residual = _bn_relu_conv(filters=filters, kernel_size=(3, 3))(conv1)
        return _shortcut(input, residual)

    return f