ETA on pytorch training code?

bigscience-workshop / t-zero

Reproduce results and replicate training fo T0 (Multitask Prompted Training Enables Zero-Shot Task Generalization)

Apache License 2.0

456 stars 53 forks source link

ETA on pytorch training code? #13

Open NtaylorOX opened 2 years ago

NtaylorOX commented 2 years ago

Hope it is okay to add this as an issue - this is amazing work and would love to see the training implementation in pytorch. I see it is in the works, and wondered if this is still the case and whether an ETA is on the horizon?

VictorSanh commented 2 years ago

Hello @NtaylorOX Thanks for your interest! As far as I know, no one (including me) has this item on their high-priority todo so it's hard to share with you a hard ETA. A loose (and more reasonable) horizon would be a few weeks. In the meantime, you should feel free to either:

use the tf implementation
adapt the single task fine-tuning codebase (there some non trivial work involved) and contributed a PR!

ari9dam commented 2 years ago

Hi @VictorSanh , I was wondering if there is any plan to publish the pytorch training code.

VictorSanh commented 2 years ago

thanks @ari9dam for the ping! Priorities changed a (big) bit since last time, and at the moment, there is no plan of officially extending the T0 replication codebase to PyTorch unfortunately.

Xnhyacinth commented 5 months ago

So the program is now canceled?