search

soumik12345 / wandb-addons

Weights & Biases Addons is a repository consisting of additional unitilities and community contributions for supercharging your Weights & Biases workflows.
https://geekyrakshit.dev/wandb-addons/
23 stars 7 forks source link

Integration with `weave` and `WeightWatcher` #36

Open soumik12345 opened 1 year ago

soumik12345 commented 1 year ago

Develop a framework-agnostic integration with weave and WeightWatcher.

The idea is to create weave panels automatically using the results of analysis and summary we get from analyzing a model using WeightWatcher.

shivance commented 1 year ago

Hi @soumik12345 can I contribute to this issue?

soumik12345 commented 1 year ago

Hi @soumik12345 can I contribute to this issue?

Hey @shivance Thanks for being willing to contribute! Please feel free to raise a PR linking this issue!

shivance commented 1 year ago

@ayulockin @soumik12345 It would be really helpful if you could lay out the initial roadmap which i should follow, as I'm new to it !

soumik12345 commented 1 year ago

@ayulockin @soumik12345 It would be really helpful if you could lay out the initial roadmap which i should follow, as I'm new to it!

@shivance The basic idea behind the integration is to create detailed and insightful Weights & Biases dashboards out of the analysis results that we get from WeightWatcher. Consider the integration to be simple functions like, log_details() and log_summary() that would log the respective data to Weights & Biases and make use of plots and tables/weave for visualization.

A sample use could be as follows:

import weightwatcher as ww
from wandb_addons.weightwatcher import log_details, log_summary

import torchvision.models as models

model = models.vgg19_bn(pretrained=True)
watcher = ww.WeightWatcher(model=model)
details = watcher.analyze()
log_details(details)
summary = watcher.get_summary(details)
log_summary(summary)