Auto modelcard

[sgugger]

What does this PR do?

This PR adds functionality in the Trainer to auto-generate model cards and some utilities to do the same without the Trainer if people are not using it. In passing, the old ModelCard class is deprecated (to be removed in v5).

As an example here is a repo that is generated by the run_glue script with this new functionality, using the following command on a machine with 2 GPUs:

accelerate launch examples/pytorch/text-classification/run_glue.py \
    --model_name_or_path bert-base-cased \
    --task_name mrpc \
    --do_train \
    --do_eval \
    --learning_rate 2e-5 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 16 \
    --evaluation_strategy epoch \
    --logging_strategy epoch \
    --weight_decay 1e-2 \
    --output_dir ~/tmp/test-glue-mrpc \
    --overwrite_output_dir \
    --push_to_hub

I've only adjusted the glue example for now, will do the others once we have settled on an API.

[patrickvonplaten]

Do you think we could directly define the finetuned_from model from the Trainer? Or is it not a good idea because some models could just have been pre-trained and not fine-tuned?

[sgugger]

The Trainer gets a model, it has no idea what the checkpoint used was.

[patrickvonplaten]

Think the subsections in the .md file need to be shifted one to the left:

\n ## -> \n##

[LysandreJik]

For reviewers: the deprecated ModelCard was set to be deprecated in February 2020, and hasn't been used since.

The result looks super cool! Love the model card resulting from the training.

As said offline, would be really cool to have the metadata be populated as well, as this would allow programmatic handling of checkpoints and would open the door to a myriad of features, paving the way for model evaluation.

I understand this may require some changes in the datasets lib, pinging @lhoestq as discussed offline.

Otherwise, LGTM!

[LysandreJik]

Great, I think this is a fantastic addition. To make it easier on reviewers, here's an example of the modelcard once uploaded to the hub:

https://huggingface.co/sgugger/tst-glue-mrpc/blob/main/README.md

An example of the metadata generated is visible here:

---
tags:
- text-classification
datasets:
- glue
metrics:
- accuracy
- f1

model-index:
- name: tst-glue-mrpc
  results:
  - task:
      name: Text Classification
      type: text-classification
    dataset:
      name: GLUE MRPC
      type: glue
    metrics:
      - name: Accuracy
        type: accuracy
        value: 0.8529411764705882
      - name: F1
        type: f1
        value: 0.8969072164948454
---

This follows the format defined in huggingface/huggingface_hub#39.

Might be of interest to @lewtun, @lhoestq

Let's go ahead and implement the remaining tasks! 🎉

[LysandreJik]

LGTM! Great job @sgugger!

[LysandreJik]

Very cool to inherit the license!