🎉 Add Multitask Prompt Tuning

[mayank31398]

Adds code for the paper: https://arxiv.org/abs/2303.02861

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[mayank31398]

@pacman100
Can you take a look at this PR?
This adds the functionality to use Multitask Prompt Tuning as proposed in this paper.

[mayank31398]

Also, a working notebook has been added @pacman100 🤗

[pacman100]

Thank you so much @mayank31398 for adding MultiTask Prompt Tuning 🔥! Could you please resolve the conflicts and run make style and make quality to resolve the quality issues. Post that we can go ahead and merge this PR 🤗.

[mayank31398]

Done @younesbelkada :)

[BenjaminBossan]

Thanks so much for this PR and all the work you put in.

I haven't read the paper in detail, so I cannot comment on whether the implementation is faithful or not.

I have a few minor comments, please check my comments. Not all of that is critical. My main concern is that, if I'm not mistaken, only one prompt_tuning_init argument is tested. Therefore, I would like to see the tests extended to cover the other types.

As a nice-to-have feature, it would great if there were a few explanations in the notebook about what's going on + a link to the paper.

[BenjaminBossan]

I just noticed that there isn't any addition to the docs for MPT. Could you please add some documentation there? Or are we fine with adding it in a separate PR?

[pacman100]

Hello @mayank31398, Thank you for iterating! I believe the PR can be merged once the conflicts are resolved as there is an example showing usage and related tests. Thank you for all the work and for your kind patience 🤗

[mayank31398]

Hey @pacman100 let me add some comments in the code.
There is a some stuff that isn't clear.
I will also resolve the merge conflicts over the weekend

[mayank31398]

@pacman100 I have addressed the suggestions.
Can you run the unittests?
Thanks

[mayank31398]

@pacman100 @younesbelkada lets merge this?

[mayank31398]

Hey, guys still waiting for an update :)

[younesbelkada]

Thanks a lot for your patience @mayank31398 and sorry for the delay , I left two tiny comments, apart from that I trust other maintainers review. Can you please revert the changes in the unrelated files?
Thanks!

[mayank31398]

@younesbelkada addressed the comments

[BenjaminBossan]

@mayank31398 Just to let you know: We'll probably soon merge #807 which adds a new folder structure to the tuners. If you want to, you can already adjust your code to use the new structure. Otherwise, it's also fine as is, then we'll do the refactoring later.

[younesbelkada]

Thanks again!

[mayank31398]

@mayank31398 Just to let you know: We'll probably soon merge #807 which adds a new folder structure to the tuners. If you want to, you can already adjust your code to use the new structure. Otherwise, it's also fine as is, then we'll do the refactoring later.

I prefer leaving as is for now.
I am a little busy with some other stuff :)
Lets refactor later

[pacman100]

Thank you @mayank31398 for iterating multiple times and for your patience, LGTM! 🔥🚀✨

[junzhang-zj]

@mayank31398 Why can't I locate the distillation process of the source prompt in the code? Am I missing something? 😭

[mayank31398]

@junzhang-zj yeah, it doesn't have that.
Its not simple to add it to this repo I think.
This is a basic MPT pipeline where you pretrain a joing prompt on a bunch of task, use it for initialization of target prompt and then finetune it for the target prompt.

[junzhang-zj]

@mayank31398 Thank you for your answer, which addressed my concern about carelessness!