Change no trainer image_classification test

[muellerzr]

What does this PR do?

I noticed that this test has been failing for the last few months, and it's due to the fact that on a single GPU or CPU the tests pass (we hit 50% accuracy), but on multi GPU it's a hit or miss whether or not it does (each epoch/batch is either 50% or 0%, and we happen to miss that coin flip each time).

This PR modifies the no_trainer test to mimic the equivalent pytorch example tests, and I can confirm it passes repeatedly on a single GPU, CPU, and multi GPU: https://github.com/huggingface/transformers/blob/main/examples/pytorch/test_pytorch_examples.py#L390-L417

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@sgugger @NielsRogge

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks for fixing (and making the test stronger!)

However I'm noticing it now takes 68s for the test (the longest time) so can we lower the threshold and the number of steps somehow?

[muellerzr]

@sgugger I've addressed the speed concern in the test, but you won't actually see it show up due to a bug in the scheduler step recalculation. I'll open up another PR for that fix and then merge this one after that's done