For conv2D can you specify convolution or correlation .
From source:
Many machine learning libraries implement cross-correlation but call it convolution.
Is this the case with PyTorch? Does it matter. I haven’t found corelate2D class.
Comments…
PyTorch uses cross-correlation.
It shouldn’t make any difference, since the kernels will just be flipped for a convolution.
However, if you are trying to apply some manually defined kernels for image processing (e.g. edge detectors), you should flip them to get the desired results.
That being said, it should not make a difference in your training, if you just use randomly initialized kernels.