如何在 PyTorch 中合并 2D 卷积?

How do I merge 2D Convolutions in PyTorch?

从线性代数中我们知道线性算子是结合的。

在深度学习领域,这个概念被用来证明在 NN 层之间引入非线性是合理的,以防止俗称 linear lasagna, (reference) 的现象。

在信号处理中,这也会导致一个众所周知的技巧来优化内存 and/or 运行时要求 (reference)。

所以合并卷积从不同的角度来看都是一个非常有用的工具。如何用PyTorch实现?

如果我们有 y = x * a * b(其中 * 表示卷积,a, b 是你的内核),我们可以定义 c = a * b 这样 y = x * c = x * a * b 如下:

import torch

def merge_conv_kernels(k1, k2):
    """
    :input k1: A tensor of shape ``(out1, in1, s1, s1)``
    :input k1: A tensor of shape ``(out2, in2, s2, s2)``
    :returns: A tensor of shape ``(out2, in1, s1+s2-1, s1+s2-1)``
      so that convolving with it equals convolving with k1 and
      then with k2.
    """
    padding = k2.shape[-1] - 1
    # Flip because this is actually correlation, and permute to adapt to BHCW
    k3 = torch.conv2d(k1.permute(1, 0, 2, 3), k2.flip(-1, -2),
                      padding=padding).permute(1, 0, 2, 3)
    return k3

为了说明等价性,这个例子将两个分别具有900和5000个参数的内核组合成一个等效 28个参数的内核:

# Create 2 conv. kernels
out1, in1, s1 = (100, 1, 3)
out2, in2, s2 = (2, 100, 5)
kernel1 = torch.rand(out1, in1, s1, s1, dtype=torch.float64)
kernel2 = torch.rand(out2, in2, s2, s2, dtype=torch.float64)

# propagate a random tensor through them. Note that padding
# corresponds to the "full" mathematical operation (s-1)
b, c, h, w = 1, 1, 6, 6
x = torch.rand(b, c, h, w, dtype=torch.float64) * 10
c1 = torch.conv2d(x, kernel1, padding=s1 - 1)
c2 = torch.conv2d(c1, kernel2, padding=s2 - 1)

# check that the collapsed conv2d is same as c2:
kernel3 = merge_conv_kernels(kernel1, kernel2)
c3 = torch.conv2d(x, kernel3, padding=kernel3.shape[-1] - 1)
print(kernel3.shape)
print((c2 - c3).abs().sum() < 1e-5)

注意:等价假设我们有无限的数值分辨率。我认为有关于堆叠许多低分辨率浮点线性运算并表明网络从数值误差中获利的研究,但我找不到它。感谢任何参考!