MLflowCallback to log run_name argument

[HenryMaguire]

🚀 Feature request

When using the MLflowCallback (set as default for Trainer), I would like to log the run_name argument passed to TrainingArguments as the Run Name on the MLflow dashboard. Currently runs are being logged as nameless.
E.g. see below.

Motivation

Trainer makes training 🤗 models so easy and MLflow is great for organising experiments/caching artifacts. I would like to make it easier to organise experimental runs and make research easier, particularly for larger teams. This feature would be a very simple patch on the original PR #8016

example usage

training_args = TrainingArguments(
    label_names=['labels_t1', 'labels_t2'],
    output_dir='./runs,          # output directory
    run_name='multitask_clf_<run_name>',
)
trainer = Trainer(
    model=model,        # the instantiated 🤗 Transformers model to be trained
    args=training_args,        # training arguments, defined above
    train_dataset=train_dataset,        # training dataset
)

Your contribution

I can submit a PR

[sgugger]

Please don't hesitate to suggest a PR, run_name is there just for this reason!

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[ksopyla]

It will be useful to have this, and it will organize all the experiments better.

[yangdong02]

Is there any update on this issue and #12841? I have a very simple one-line solution of passing args.run_name (which currently serves for wandb) to mlflow.start_run that can fix this. I can submit a PR in case of need.

[pratikchhapolika]

@HenryMaguire can you send link to your notebook or colab?