optimizer wrapper improve docs

[tmabraham]

There is often confusion about how to use OptimWrapper (so we can use PyTorch optimizers). Even @muellerzr has experienced some confusion in this regard (see #3225). This PR is to improve some of the documentation to clarify this. I don't know if this is the best approach so I'm open to moving the text I wrote to another location, or adding some text somewhere else, etc...

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[jph00]

Do we know what's changed that's caused the compose()/inline approach to not work any more? cc @muellerzr

[muellerzr]

@jph00 I sadly don't particularly know what exactly is causing the issue. (I also don't know what version fixed this, as I didn't really use this functionality much).

For some reason it looks like it's trying to do something like AdamW()() (IE call the class instance) rather then .step I believe? But I also don't know what the old behavior exactly did.

@tmabraham used it much more, so he may have an idea?

[tmabraham]

There are two problems when it comes to the compose approach. This is what I am trying:

opt_func = compose(partial(torch.optim.Adam, eps=1e-6), OptimWrapper)

(I had also tried something similar earlier with AdaBelief)

If I try passing this opt_func into Learner like this:

learn_ = Learner(dls, resnet34(), opt_func=opt_func, metrics=error_rate)
learn_.fit(1)

I get this error:

TypeError: __init__() got an unexpected keyword argument 'lr'

I believe this is a limitation of fastcore's compose function. compose passes the same kwargs into all of the composed functions, but OptimWrapper doesn't expect lr and other related arguments, hence the error.

If I try with cnn_learner like this:

path = untar_data(URLs.PETS)/'images'

def is_cat(x): return x[0].isupper()
dls = ImageDataLoaders.from_name_func(
    path, get_image_files(path), valid_pct=0.2, seed=42,
    label_func=is_cat, item_tfms=Resize(224))

learn = cnn_learner(dls, resnet34, opt_func=opt_func, metrics=error_rate)
learn.fit(1)

I get the following error:

TypeError: optimizer can only optimize Tensors, but one of the params is list

I am not sure why I get this error, but it may have something to do with the param_lists/the splitting of the model.

For these reasons, the compose approach does not currently work...

[tyoc213]

It will be nice to know the pytorch commit that caused this, I guess pytorch is 1.7.0 but has been since some time?. So I guess the different behaviour was introduced from 1.6 to current 1.7 isn't? or we can't be sure about that?

As a side note from what I know from the TPU thing, there are some tiny diffs very internal to optimizers from fastai and pytorch so perhaps could be a good time to test in deep compatibility between the two?

[tmabraham]

closing and bringing proposed changes to #3241