I have a tensor which represents overlapping chunks (of 2D audio coefficients):
>>> x = torch.rand((2, 113, 12, 244)) # 12 blocks of 244 = 2928
>>> x = x.reshape(2, 113, 12*244)
>>> print(x.shape)
torch.Size([2, 113, 2928])
Each of these 12 blocks of size 244 has a 50% overlap with the next block. I would like to overlap-add on the last dimension with blocks of size 244, hop size of 122. The total coefficients should be 2928/2 = 1464.
I have read different forum posts, stackoverflow questions, and the documentation of Fold/Unfold, and still have trouble discovering which parameters I need to get my desired output:
>>> desired_output_size = (2, 113, 1464)
>>> y = torch.nn.functional.fold(x, kernel_size=(?), stride=(?))
>>> print(y.shape)
torch.Size([2, 113, 1464])
Could anybody provide some guidance? Thanks.