pytorch 中的所有值都替换为 pool max

如何解决pytorch 中的所有值都替换为 pool max

给定 z 等于

tensor([[[[0.0908,0.1286,0.6942,0.5161],[0.4227,0.2154,0.5990,0.8666],[0.3009,0.2399,0.1818,0.7551],[0.2396,0.4485,0.4027,0.5303]],[[0.8251,0.7457,0.2091,0.7313],[0.3823,0.7351,0.3823,0.2072],[0.0863,0.5489,0.6515,0.3855],[0.5247,0.8685,0.6078,0.6181]]]])

我们有

torch.nn.MaxUnpool2d(2)(*torch.nn.MaxPool2d(2,return_indices=True)(z))
tensor([[[[0.0000,0.0000,0.0000],[0.0000,0.0000]],0.0000]]]])

我怎样才能惯用地计算一个类似的逆，而不是所有非最大值都为零，所有都等于最大值？在这种情况下，我想要：

tensor([[[[0.4227,0.4227,0.8666,[0.4485,0.7551,0.7551]],0.8251,0.7313,[0.8251,[0.8685,0.6515],0.6515]]]])

该解决方案应该适用于所有内核大小、填充和膨胀，并且速度很快。例如，在这个简单的案例中，我可以使用 torch.nn.MaxPool2d(2)(z) 轻松计算出我想要的 Kroeneker 乘积，但这并不能概括。