Seeding in Transform, make the same transform all the time

[romainVala]

🐛Bug
I notice it at least for RandomAffine RandomElastic and RandomNoise

To reproduce

t = Compose([RandomNoise(seed=10)])
dataset = ImagesDataset(suj, transform=t); 

sample=dataset[0]
for i in range(0, 10):
    sample = dataset[0]
    h=sample.history
    print(h[0][1]['image']['std'])

Now change seed=10 to seed=None and you get a diferent number each time

I thing it is due to this line

torchio/torchio/transforms/augmentation/random_transform.py

Line 32 in 0e8172c

self.check_seed()

it should not be called here, juste once in the init()

[romainVala]

may be it was intended so, and this is my understanding of seeding which is different ...

I thought it would produce a different number each call, (but If I run it again, I would get the same number)

[fepegar]

Hi @romainVala,

This point of setting the seed is to get the same results every time. Maybe a better name for the argument would be random_state: https://scikit-learn.org/stable/glossary.html#term-random-state.

[fepegar]

Here's some more information about reproducibility in PyTorch: https://pytorch.org/docs/stable/notes/randomness.html

[romainVala]

hmm, but if you get the same results for each call, it is no more a random transform ...
my understanding of reproducibility is to get the same sequence of random transform non ?

[fepegar]

AFAIK random numbers in the libraries we use are generated using pseudorandom number generators. I suppose they're actually pseudorandom transforms :)

In [1]: import torch

In [3]: torch.rand(1)
Out[3]: tensor([0.0833])

In [4]: torch.rand(1)
Out[4]: tensor([0.7058])

In [5]: torch.rand(1)
Out[5]: tensor([0.8763])

In [6]: torch.manual_seed(42)
Out[6]: <torch._C.Generator at 0x7f01b881f3d0>

In [7]: torch.rand(1)
Out[7]: tensor([0.8823])

In [8]: torch.rand(1)
Out[8]: tensor([0.9150])

In [9]: torch.rand(1)
Out[9]: tensor([0.3829])

In [10]: torch.manual_seed(42)
Out[10]: <torch._C.Generator at 0x7f01b881f3d0>

In [11]: torch.rand(1)
Out[11]: tensor([0.8823])

In [12]: torch.rand(1)
Out[12]: tensor([0.9150])

In [13]: torch.rand(1)
Out[13]: tensor([0.3829])

You can set a global (torch) seed once at the beginning of your training to make your runs reproducible. If you want, for some reason, to generate the same transform with the same parameters twice, you can use the seed kwarg of the transform.

In this answer there is some more info.

[romainVala]

ok, this was then my understanding of seed keyword in the transform that confuse me (you should specify more this behavior in the doc )

I do not really see any case where this may be useful ...
(I had a use case where I want a specific transform like a rotation of 10, then I use RandomAffine with degree=(10,10) for that ... )

Anyway as you want:
I just need to remove seed argument, for all my transform (and redo the training from this WE ... )

thanks