Tensorflow directly provides conv2d_backprop_filter and conv2d_backprop_input functions, but they are not available in pytorch as operations that can be called directly from python.
e.g.: https://www.tensorflow.org/api_docs/python/tf/nn/conv2d_backprop_filter. These are useful when implementing alternate forms of backprop through convolution (discretized, layerwise relevance backprop, etc).
More discussion here https://discuss.pytorch.org/t/how-to-do-convolution-backward-manually/1430
Although pytorch obviously can do these operations during backprop using a C implementation, it doesn't currently expose the functions directly in python.
Tensorflow directly provides conv2d_backprop_filter and conv2d_backprop_input functions, but they are not available in pytorch as operations that can be called directly from python.
e.g.: https://www.tensorflow.org/api_docs/python/tf/nn/conv2d_backprop_filter. These are useful when implementing alternate forms of backprop through convolution (discretized, layerwise relevance backprop, etc).
More discussion here https://discuss.pytorch.org/t/how-to-do-convolution-backward-manually/1430
Although pytorch obviously can do these operations during backprop using a C implementation, it doesn't currently expose the functions directly in python.