Reshape selected axis (with torch.reshape)?

1

Hi! So kind of a stupid question but say if I have a tensor dimension of T x B x C x H x W and I flattened it to TB x C x H x W, would torch reshape using the original dimension values gave back the same order of data? Is there a way that I can reshape only the selected axis to expand the dimension? and would the order also be maintained?

2

Yes, as long as you don’t permute the dimensions, view will yield the original tensor:

T, B, C, H, W = 2, 3, 4, 5, 6
x = torch.randn(T, B, C, H, W)
y = x.view(-1, C, H, W)
z = y.view(T, B, C, H, W)
print((z == x).all())
> tensor(True)
3

Gotcha, thank you! I was able to verify that is the case here.

4

@ptrblck
My tensor shape is 4*H,W
if arrange this using loop to bigger H,W that is factor of above i get correct Image.
But when i simply do view(BigH,BigW) i get incorrect image ,there is repetition of patches

5

Could you post a code snippet showing these unexpected results, please?

6

Pred= model (x)
#pred shape is 4 * 4* 272,352
pred= pred.argmax(1).contiguous()
#pred shape is 4 272352
pred= pred.view(2,2,272,356).view(544,704) #gives incorrect image
#expected ordering of index is[ [0,1][2,3]]

When use for loop i get right image
for i in rows
for j in columns
img[i:i+1 *272,j:j+1*352]= pred[index]
Index=index+1

7

In your code you are interleaving dimensions, which I tried to mention in:

as long as you don’t permute the dimensions, view will yield the original tensor

I’m also unsure how the for loop approach would work and if you have a typo or of you want to replace the values in img. Should the indexing be [i:(i+1)*272, j:(j+1)*352]?
The reshape operation is also failing, as it seems you want to flatten it to (2, 2, 272, 352).

8

Yes it was typo… in for loop
What do you mean by interleaving?
So there is no straight way ?
I even flattened whole tensor , then tried to reshape but it gave same results .
@ptrblck
I thought pytorch does sequential reshaping

9

You are trying to flatten dim0 into two different dimensions, which won’t work without a permutation.
Since your for loop approach wasn’t updated, I assume this code snippet creates the desired output and also shows how to permute the original tensor before reshaping:

pred = torch.randn(4, 4, 272, 352)
pred= pred.argmax(1).contiguous()
pred_ref = pred.view(2, 2, 272, 352).permute(0, 2, 1, 3).reshape(544, 704)
pred_ = pred.view(-1, 272, 352)

img = torch.zeros(544, 704)
index = 0
for i in range(2):
    for j in range(2):
        img[i*272:(i+1)*272,j*352:(j+1)*352]= pred_[index]
        index = index+1

print((img == pred_ref).all())
> tensor(True)
1 Like