Random state argument for consistent random augmentations

[Dale-Black]

I would like to suggest an argument for RandomResizeCrop and other stochastic transformations. Something like RandomResizeCrop(seed=...) would be really convenient for augmenting supervised data with the same random seed. This isn't necessary obviously as you have already given me a helpful workaround, but in case this is easy to implement I wanted to add this suggestion

[lorenzoh]

So you want to fix the randomness of a stochastic transformation so it always gives the same result? What would seed do exactly?

[Dale-Black]

I was thinking something like this. But I’m not sure if this is much cleaner than the route you have already proposed actually

[lorenzoh]

Hm, I don't really see an advantage in this. One thing to do could be to implement a wrapper transform Deterministic that wraps a transform and always passes the same random state through. That way you wouldn't need to change existing transforms.

[Dale-Black]

Fair enough. I think I agree with you the more I think about it