fb.resnet like transforms

[eladhoffer]

No description provided.

[fmassa]

I was wondering if it would make sense to add a inplace option to the constructor, that modifies the input tensor inplace.
This would only be a tiny optimization to avoid a copy+clone, don't know if it's worth it though.

[Maratyszcza]

@eladhoffer Very nice! Do you have an estimate of how these transforms improve accuracy for ImageNet models?

[eladhoffer]

Actually, I didn't find any difference for most models...

[fmassa]

@soumith @colesbury Shall we merge this (after a rebase/review)? These transforms don't seem to help improve performance on classification tasks, but users might want to try it out anyway for their own tasks.
cc @alykhantejani

[Maratyszcza]

@fmassa I am struggling to train tiny darknet model to reported accuracy, and thus far I matched everything, but the extra pre-processing transformations. It seems likely that extra transformations matter and are be necessary to reproduce the result on this classifier.

[alykhantejani]

@Maratyszcza did you find these transforms necessary to reproduce tiny darknet in the end?

[alykhantejani]

Hi @eladhoffer sorry for the delayed response on this - are you still interested in adding these? If not, we can get round to adding them in the next few weeks.

I think Brightness, Staturation and Contrast are valuable and bring us up to speed with what tensorflow have (see #271 for the list).

[eladhoffer]

Sure, how can I help?

[alykhantejani]

@eladhoffer Awesome. I think the following changes are needed:

1. The transforms at present appear to be based on Tensor inputs, but we seemed to have converged to use PIL/accimage Image inputs everywhere
2. We now have a functional interface and the objects are just a thin wrapper around functions, this will eventually allow us to do things like RandomOrder from this PR in one function as a transform (randomly calling other functions)
3. it would be good if you could add some tests for these transforms.

It'd be great if you have the time to make these changes to this PR, if not @chsasank has started implementing some other color transforms #275 - in any case it's probably worth you guys coordinating as to not duplicate work.

[chsasank]

Hey @eladhoffer, I am almost done implementing all the colour transforms with PIL in #275. So I'll just continue the PR and complete it. This is fine?

[eladhoffer]

Sure, let me know if you need any help.

[alykhantejani]

Closing this then as this will be implemented in #275