The input to a transpose convolution can also be unfolded. Then, the forward pass (with zero bias) is a matrix-matrix multiplication between unfolded input and matrix-view of the kernel.
[x] Add unfold_transposeNd as im2col functional for conv_transposeNd
[x] Needs to be documented, as there is no directly equivalent function in torch (which also means there is no competitor for the benchmark)
[x] Add UnfoldTransposeNd as im2col module for ConvTransposeNd
[x] Needs to be documented, as there is no directly equivalent function in torch (which also means there is no competitor for the benchmark)
[x] Test via the above equivalence of the forward pass
The input to a transpose convolution can also be unfolded. Then, the forward pass (with zero bias) is a matrix-matrix multiplication between unfolded input and matrix-view of the kernel.
unfold_transposeNdasim2colfunctional forconv_transposeNdtorch(which also means there is no competitor for the benchmark)UnfoldTransposeNdasim2colmodule forConvTransposeNdtorch(which also means there is no competitor for the benchmark)README