Training on GPU about ptec

[valentinbuc]

Hey @AlphaKh,

I would recommend installing a Windows Subsystem for Linux (WSL). Once installed, you can open the WSL terminal and run the bash commands as provided in the readme.

Alternatively, you can also open the .sh files and run the python commands in your windows command prompt. That will only be a few more commands to run, but will have the same effect.

[valentinbuc]

I am glad to hear that, would be happy to hear about the results of your project!

For calculating the performance, we used the classification_report from sklearn.metrics. You can see how we implemented it here https://github.com/EQTPartners/PTEC/blob/main/sectors/utils/evaluation.py

The computational costs were a bit more complex to calculate. We used torch.profiler.profile, you can see how we implemented it here: https://github.com/EQTPartners/PTEC/blob/main/sectors/experiments/test_flops.py
As the profiler itself uses a lot of memory, this was only done for a small set of samples, and then we extrapolated the FLOPs used on these samples to the complete prompt-tuning and inference process.

[valentinbuc]

If you are limited in computing resources, it may indeed be ideal to work with the bigscience/bloom models!

If you don't aim for an exact reproduction but also want to extend the study a bit, it should also be interesting to try out some of the newer and smaller models like gemma 2B. That would probably require some adjustments to the codebase however!

[AlphaKh]

We want to know how you calculated the performance metrics and computational costs in your results section. Could you share any specific methods or insights to help us replicate your experimental setup? We've only experimented on BigScience dataset instead of the larger HuggingLLAMA dataset because of the resource limitations and how much parameters it got. Just so you know, we're university students working on a project, and we chose your paper to reproduce. I know not doing HuggingLLAMA will effect the module but is it still possible to calculate the results?

[AlphaKh]

One last thing regarding the results, what are we supposed to compare exactly because as I said we didn't do any experiments on Huggingllama because it's too big with 7b parameters. So are we supposed to compare the methods done on the experiments for bigscience?

[AlphaKh]

Thank you so much for your recommendations, would you mind if you record yourself for 30s and in these 30s If you don't mind complementing our work integrity and ethics some things along the line "They contacted me about the models testing and training error'' This will be used in out presentation to help us stand out and we appreciate all the effort you helped us in.

[valentinbuc]

Your team definitely put great effort into this project, and I can endorse that you contacted me to clarify questions. I would be curious to read your report and have a look at your findings!