Question

我正在尝试使用带有TensorFlow后端的Keras创建一个非池化层。 this paper中描述了我要实现的解池操作。这是SegNet所使用的相同的分池操作。

解池：在convnet中，最大池化操作是不可逆的，但是我们可以获得一个近似值通过记录一组开关中每个池化区域内的最大值变量。在deconvnet中，分拆操作使用这些开关从将上面的图层放到适当的位置，保留刺激的结构。

我的大部分代码都是对this implementation的改编，来自旧版的Keras。

到目前为止，我已经编写了一个自定义层，它正在正确执行卸载操作。我遇到的问题是当Keras尝试在反向传播阶段计算梯度时，出现错误：

raise ValueError('An operation has `None` for gradient. '
ValueError: An operation has `None` for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.

据我了解，此错误是由于我必须使用Keras不知道其导数的操作引起的。

这是我的代码：

from keras import backend as K
from keras.engine.topology import Layer
from keras.layers import UpSampling2D
import numpy as np
import operator

class MyUnpooler(Layer):

    # Initialisations
    def __init__(self, pool_layer, pool_input, size=(2,2), **kwargs):
        self.pool_layer = pool_layer
        self.pool_input = pool_input
        self.size = size
        super(MyUnpooler, self).__init__(**kwargs)

    # This method would be used to create weights
    # I don't have any trainable weights but this must be implemented
    def build(self, input_shape):
        super(MyUnpooler, self).build(input_shape)

    # This method is for the layers' logic
    # x is always input to the layer
    def call(self, x):
        # Zeros for later
        zeros = K.zeros_like(self.pool_input)

        # Mask template
        upsampled = UpSampling2D(size=self.size)(self.pool_layer)
        upsampled_shape = upsampled.get_shape().as_list()[1:]
        input_shape = self.pool_input.get_shape().as_list()[1:]

        size_diff = map(operator.sub, input_shape, upsampled_shape)
        unfiltered_mask = K.spatial_2d_padding(upsampled, padding=((0,size_diff[1]),(0,size_diff[2])))

        # Create the mask
        self.mask = K.tf.equal(self.pool_input, unfiltered_mask)
        assert self.mask.get_shape().as_list() == self.pool_input.get_shape().as_list()

        self.unpooled = K.tf.where(self.mask, self.pool_input, zeros)
        return K.tf.to_float(self.unpooled)

    def compute_output_shape(self, input_shape):
        # input_shape is not actually the input shape we need...
        # We need to be able to UpSample the layer to calculate the dimensions
        upsampled = UpSampling2D(size=self.size)(self.pool_layer)
        upsampled_shape = upsampled.get_shape().as_list()[1:]
        inp_shape = self.pool_input.get_shape().as_list()[1:]

        size_diff = map(operator.sub, inp_shape, upsampled_shape)
        unf_mask = K.spatial_2d_padding(upsampled, padding=((0,size_diff[1]),(0,size_diff[2])))

        return tuple(unf_mask.get_shape().as_list())

如果有更好的方法来进行分池操作，那么我也可以完全忽略到目前为止的尝试。

是否可以重写此分拆操作，以便Keras可以计算导数？

0 个答案: