如何在Keras中使用自定义2D卷积内核进行实验? [英] How to experiment with custom 2d-convolution kernels in Keras?

问题描述

在具有kernel_size=3的默认Conv2D层中，其中一个过滤器的切片的权重可以这样命名:

In a default Conv2D layer with kernel_size=3 the weights of a slice of one of the filters could be named like this:

A B C
D E F
G H I

使用kernel_size=5这样:

A B C D E
F G H I J
K L M N O
P Q R S T
U V W X Y

现在，我想基于带有conv层的conv层构建(和训练/测试)一个模型，

Now I'd like to build (and train/test) a model based on conv layers with kernels like that:

A A B C C
A A B C C
D D E F F
G G H I I
G G H I I

如何实现这样的自定义层喜欢吗?

How could the implementation of such a custom layer look like?

推荐答案

也许像这样吗?

class CustomConv2D(Layer):
    def __init__(self, filters, **kwargs):
        self.filters = filters
        self.kernel_size = (3, 3)
        super(CustomConv2D, self).__init__(**kwargs)

    def build(self, input_shape):
        # only have a 3x3 kernel
        shape = self.kernel_size + (input_shape[-1], self.filters)
        self.kernel = self.add_weight(name='kernel', shape=shape,
                                      initializer='glorot_uniform')
        super(CustomConv2D, self).build(input_shape)

    def call(self, x):
        # duplicate rows 0 and 2
        dup_rows = K.stack([self.kernel[0]]*2 + [self.kernel[1]] + [self.kernel[2]]*2, axis=0)
        # duplicate cols 0 and 2
        dup_cols = K.stack([dup_rows[:,0]]*2 + [dup_rows[:,1]] + [dup_rows[:,2]]*2, axis=1)
        # having a 5x5 kernel now
        return K.conv2d(x, dup_cols)

    def compute_output_shape(self, input_shape):
        return input_shape[:-1] + (self.filters,)

诀窍是在3x3内核中(硬编码，您可能要泛化它)仅将每个过滤器仅存储9个权重，并复制第一行和最后一行和最后一列，以使其成为所需的5x5内核.然后，将此内核传递给K.conv2d()，就像在原始的Conv2d实现.

The trick is to simply store only 9 weights per filter in a 3x3 kernel (hard coded, you may want to generalize it) and to duplicate the first and last rows and columns to make it a 5x5 kernel the way you want it. Then this kernel is passed to K.conv2d() just like in the original Conv2d implementation.

我对其进行了测试，它似乎可以正常工作.您可能需要添加其他参数，例如padding，bias等.

I tested it and it seems to be working. You may want to add other parameters like padding, bias, etc.

这篇关于如何在Keras中使用自定义2D卷积内核进行实验?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！