Update allennlp to newer version
marekrydlewski opened this issue · comments
First, thanks for your amazing work.
Do you plan to update allennlp dependency to a newer version? (e.g. 2.x.x)? This project uses an archaic version 0.8.4 which causes quite a few security problems & it is problematic to work with Python newer than 3.7.
I've tried to make the update on my own, but there are several changes in allennlp API as well as additional tweaks of allennlp classes in this repository makes it hard for someone not accustomed to the codebase.
The old version of allennlp stops us from using GECToR in our company, thank you!
Hi @marekrydlewski
Thanks for the feedback. Indeed, transition to the newest AllenNLP would require significant code rewrites.
Unfortunately, we do not plan to update it. From the very beginning, this repository was treated as a "research paper repository".
At the same time, if you have the opportunity to update the code (possibly collaborating with others), we would be happy to review your changes. This was previously done to update the transformer version. Thanks
Thanks, @skurzhanskyi There is a chance we will try to update the code, await pull requests then
Nice to hear it 🔥
Hi @marekrydlewski
Im wondering how far you've done for updating the code. Our team is also interested to do the same thing.
@marekrydlewski @damien2012eng Hey, I have a re-implementation of the GECToR code, which drops all AllenNLP dependencies and use vanilla pytorch to construct the code. For acceleration and distributed training, I use deepspeed to simplify the whole training process. I test all the codes with my experiments, it turns out to be much faster and achieves compatible performance compared to this repo. If you're interested, you can check my repo here: https://github.com/Jason3900/FastGECToR . @skurzhanskyi And if it's useful, I can also make a merge request to your repo. BTW, the AllenNLP is in maintenance mode.
@Jason3900 This is really cool. Thanks for sharing!
@Jason3900, great news! You did a big piece of work, and I see that code differs significantly. For reproducibility, I suggest that we treat your work as a separate repository and add the link to it in our README. WDYT?
@skurzhanskyi That would be great! Thanks!
@damien2012eng You are welcome. Glad to be of help!
@Jason3900 Thank you for your work!
@skurzhanskyi You are welcome! Thanks for the excellent work on GEC!
@skurzhanskyi Hey! Sorry to bother you again. My team would like to public my project on https://github.com/cofe-ai/fast-gector . And I'll maintain it from now on. Therefore, is it possible to update your readme to change the link?
Sure, will do