I'm trying to reverse the order of the rows in a tensor that I create. I have tried with tensorflow and pytorch. Only thing I have found is the torch.flip() method. This does not work as it reverses not only the order of the rows, but also all of the elements in each row. I want the elements to remain the same. Is there an array operation of this to index the integers? For instance:
tensor_a = [1, 2, 3] [4, 5, 6] [7, 8, 9] I want it to be returned as: [7, 8, 9] [4, 5, 6] [1, 2, 3] however, torch.flip(tensor_a) = [9, 8, 7] [6, 5, 4] [3, 2, 1] Anyone have any suggestions?
2 Answers
According to documentation torch.flip has argument dims, which control what axis to be flipped. In this case torch.flip(tensor_a, dims=(0,)) will return expected result. Also torch.flip(tensor_a) will reverse all tensor, and torch.flip(tensor_a, dims=(1,)) will reverse every row, like [1, 2, 3] --> [3, 2, 1].
I am not sure about the performance of the solution I have, but you can do something as follow:
import torch y = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) # Uncomment the next two lines so u can see it works on GPU as well # device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") # y.to(device) y = y[[2, 1, 0], :] # y = y[::-1, :] # this works in numpy but not in pytorch :( print(y) You can check numpy documentation on Slicing and Advanced Indexing for similar examples. Hope this helps.
