Add image classification script, no trainer

[NielsRogge]

What does this PR do?

This PR adds an example script for image classification that leverages Accelerate instead of the HuggingFace Trainer.

To do:

• verify local train_dir and validation_dir
• update README
• add log fixes (Tensorboard)

Both can be updated after #16585 is merged.

[HuggingFaceDocBuilderDev]

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

[NielsRogge]

@sgugger not sure why, but the test for the script fails:

WARNING  datasets.builder:builder.py:388 Using custom data configuration huggingface--image-classification-test-sample-b7448dc7ae37f2cf
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:388 ***** Running training *****
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:389   Num examples = 8
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:390   Num Epochs = 3
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:391   Instantaneous batch size per device = 2
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:392   Total train batch size (w. parallel, distributed & accumulation) = 2
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:393   Gradient Accumulation steps = 1
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:394   Total optimization steps = 12
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:471 epoch 0: {'accuracy': 0.0}
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:471 epoch 1: {'accuracy': 0.0}
INFO     run_image_classification_no_trainer:run_image_classification_no_trainer.py:471 epoch 2: {'accuracy': 0.0}

Weirdly, it passes locally for me.

[sgugger]

Nice new addition!
Tested locally and had the test pass one time out of 4, so maybe it's a seed issue? Setting a seed made it pass several times in a row. The learning rate seems extremely high for a Transformers model (but maybe ViT exepects that) so I'd say the test is a bit unstable without setting the seed.

[NielsRogge]

I'm getting issues when only passing id2label and label2id to the config, but not the num_labels:

if size_average is not None or reduce is not None:
            reduction = _Reduction.legacy_get_string(size_average, reduce)
>       return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)
E       IndexError: Target 6 is out of bounds.

[sgugger]

Oh ok, shouldn't be the case. Let's put back the num_labels for now and I'll have a look later at why it failed to update properly.