Update VisionClassifierTrainer

[ReiiNoki]

Hello Yanis.

Recently I am taking a Pytorch course in Coursera "Advanced PyTorch Techniques and Applications" and the lecturer used your library for teaching in Module 5 of the course.

When I wrote the code following him I found a fatal problem from your library since it is now currently a beta version.

The problem came from VisionClassifierTrainer.
In the lastest version of library transfomers from huggingface, the class TrainingArguments now changes the keyword "evaluation_strategy" to “eval_strategy” thus your original code didn't work.

I have to change the keyword following the instruction of huggingface/transformers#7974 (comment).

After the change, my training can run but another error came out.

KeyError: "The metric_for_best_model training argument is set to 'eval_accuracy', which is not found in the evaluation metrics. The available evaluation metrics are: ['eval_loss']. Consider changing the metric_for_best_model via the TrainingArguments."

I have to go back to read the code from transformers and found they now only support "loss" as the endword. Thus I changed "accuracy" to "loss" for eval_metrc.

Finally, I can run the training from the course.

Thank you very much for you to work on this easy-to-use library for us to learn Pytorch and deep learning.
I hope my PR could help you to update and maintain it.