In tensorflow, it passes this input tensor (128,32,32,16) into the tf.extract_image_patches and output a tensor with size (128,32,32,144) with kernel size (1,3,3,1) and stride (1,1,1,1) as the parameter for the tf.extract_image_patches. so i assume in the tf.extract_image_patches it do overlap sampling. And i tried tensor.unfold cannot perform this. Maybe it can and I dont know how it will output the same as tensorflow…
Using a nn.Conv2d is an idea, by setting the weights to 1? but it misses the sampling part…
Thank you again…