Implicitly setting a random seed?

[miriaford]

🐛 Bug

To Reproduce

Steps to reproduce the behavior:

import pytorch_lightning
import random
print(random.randint(0, 100000))

If you run this script repeatedly, you will get the same 'random' integer. PyTorch-lightning seems to be setting a random seed behind the hood. This is very dangerous because I want to show the variance of my algorithm, and PTL actually makes (part of) it deterministic.

Expected behavior

PTL shouldn't set the seed without the user's permission.

[williamFalcon]

pl doesn’t set the seed anywhere. we have a function you can use to do it. search the codebase for the word seed

[williamFalcon]

https://github.com/PyTorchLightning/pytorch-lightning/search?q=seed&type=

[miriaford]

Yes, but why do I get deterministic random int every time I run the script? I didn't set any seed explicitly. It's only 3 lines, you can try it yourself.

[miriaford]

Also if I simply remove the first line import pytorch_lightning, the script prints a different int every time (which is the expected behavior)

[williamFalcon]

(env) ➜  Desktop python a.py
319
(env) ➜  Desktop python a.py
350
(env) ➜  Desktop python a.py
548
(env) ➜  Desktop cat a.py
import pytorch_lightning
import random
print(random.randint(0, 1000))

[miriaford]

I don't know why I'm getting the deterministic behavior - I'm on PL 0.7.5, both OS X and Ubuntu 16.04 give me the same issue

[williamFalcon]

Not sure... we're very aware of correct research practices. Double check your environment or how you're running this...

Here's a discussion showing how hard we thought about adding the seed_everything function... we don't do new feature decisions lightly.

Check this PR
#1572

[miriaford]

thanks!