I wanted to check how many combinations the torch.combinations() function would return for different r values.
values = torch.arange(23, dtype=torch.uint8, device="cuda:0")
combinations = torch.combinations(values, r=23)
Even though I picked r=23 which should just return 1 combination, I got the above mentioned error: CUDA out of memory error. Tried to allocate 7869836414.81 GiB (...).
Why does this happen?
Hi,
This line generates a list [0, 1, ..., 22]. If you want just a single number, close the range on it [23, 24].
values = torch.arange(23, 24, dtype=torch.uint8, device="cuda:0")
Bests
Thanks for your reply.
I mean when I run:
values = torch.arange(23, dtype=torch.uint8, device="cuda:0")
it creates a list with 23 values like you said: [0, 1, ..., 22].
So I wanted to test how many different combinations torch.combinations() returns when I increase the r value, which determines how many values are in a combination.
I get an error for values of r > 7 I think, which makes sense because it finds a lot of different combinations, which take up a lot of memory. But when I tested the function with r=23, I was surprised to get a CUDA out of memory error, since the list is 23 elements long and torch.combinations() should therefore just return the original list again. I tried it for a smaller list as well:
values = torch.arange(5)
combinations = torch.combinations(values, r=5)
# len(combinations) = 1
So I was wondering why I got an error with my first example.
Ow, sorry about my first answer, I was totally somewhere else. Actually, it seems you are right, it should return only a list of 22 numbers. The strange part is if you run your code for i=1 to i=9 it just works fine but after that it does not work and tries to allocate 9GB!
I ran below code:
import time
for i in range(1, 10):
t = time.time()
values = torch.arange(i)
combinations = torch.combinations(values, r=i, )
print('for i=',i,' => ',time.time() - t)
print(combinations)
# here is the strange output (look at the time)
# for i= 1 => 0.0003619194030761719
# tensor([[0]])
# for i= 2 => 0.0005240440368652344
# tensor([[0, 1]])
# for i= 3 => 0.00019550323486328125
# tensor([[0, 1, 2]])
# for i= 4 => 0.000286102294921875
# tensor([[0, 1, 2, 3]])
# for i= 5 => 0.0006687641143798828
# tensor([[0, 1, 2, 3, 4]])
# for i= 6 => 0.001373291015625
# tensor([[0, 1, 2, 3, 4, 5]])
# for i= 7 => 0.018665552139282227
# tensor([[0, 1, 2, 3, 4, 5, 6]])
# for i= 8 => 0.36806702613830566
# tensor([[0, 1, 2, 3, 4, 5, 6, 7]])
# for i= 9 => 8.350317239761353
# tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8]])
For i=10 I could not get any answer!
Here is the code using itertools which is the base code of torch.combinations:
import itertools
for i in range(1, 11):
t = time.time()
values = torch.arange(i)
combinations = list(itertools.combinations(values, r=i))
combinations = torch.tensor(combinations)
print('for i=',i,' => ',time.time() - t)
print(combinations)
# output
# for i= 1 => 0.0003228187561035156
# tensor([[0]])
# for i= 2 => 0.0002307891845703125
# tensor([[0, 1]])
# for i= 3 => 7.605552673339844e-05
# tensor([[0, 1, 2]])
# for i= 4 => 0.0005385875701904297
# tensor([[0, 1, 2, 3]])
# for i= 5 => 0.00017261505126953125
# tensor([[0, 1, 2, 3, 4]])
# for i= 6 => 0.0003154277801513672
# tensor([[0, 1, 2, 3, 4, 5]])
# for i= 7 => 9.918212890625e-05
# tensor([[0, 1, 2, 3, 4, 5, 6]])
# for i= 8 => 0.00019240379333496094
# tensor([[0, 1, 2, 3, 4, 5, 6, 7]])
# for i= 9 => 9.393692016601562e-05
# tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8]])
# for i= 10 => 0.0001595020294189453
# tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])
And the expected part is that it creates all the answers within same amount of time.
Can you report it on Github account of PyTorch?
Edit: add link to issue
Thank you