model_to_fcn.py 文件源码

def get_output_for(self, input, **kwargs):
        input_shape = input.shape
        if self.dilation[0] > 1:
            # pad such that the time axis length is divisible by the dilation factor
            pad_w = (self.dilation[0] - input_shape[2] % self.dilation[0]) % self.dilation[0]
            input = T.concatenate((input, T.zeros((input_shape[0], input_shape[1], pad_w, input_shape[3]), input.dtype)), axis=2)
            # rearrange data to fold the time axis into the minibatch dimension
            input = input.reshape((input_shape[0], input_shape[1], -1, self.dilation[0], input_shape[3]))
            input = input.transpose(0, 3, 1, 2, 4)
            input = input.reshape((-1,) + tuple(input.shape[2:]))
        output = super(TimeDilatedMaxPool2DLayer, self).get_output_for(input, **kwargs)
        if self.dilation[0] > 1:
            # restore the time axis from the minibatch dimension
            output = output.reshape((input_shape[0], self.dilation[0]) + tuple(output.shape[1:]))
            output = output.transpose(0, 2, 3, 1, 4)
            output = output.reshape((input_shape[0], output.shape[1], -1, output.shape[4]))
            # remove the padding
            output = output[:, :, :output.shape[2] - pad_w]
        return output