Replicate 0.3.0
This is the first big release since we launched Replicate, and we've got a smorgasbord of new features for you.
In case you're landing here for the first time, Replicate is a version control system for machine learning models.
Some highlights:
Files are uploaded in the background
When you pass files to replicate.init() or experiment.checkpoint(), these will now be uploaded in the background while your training continues. Using Replicate won't make your training any slower.
This also involved a lot of behind-the-scenes improvements which have made things faster and more robust. For example, opening files from Python is now much faster.
Easier to read experiment list
The output of replicate ls has been redesigned so it works better with more experiments and params:
Check out individual files
You can now check out individual files instead of everything inside a checkpoint and its experiment. For example, you can run this to just check out the file model.pth:
replicate checkout 3b453f0 --path model.pth
Thanks to @VastoLorde95 for working on this!
PyTorch Lightning integration
Replicate now includes a callback that makes it much easier to use with PyTorch Lightning. Full details are in the documentation.
Thanks @gan3sh500 for working on this!
And a few more things
Some more features:
experiment.checkpoint()automatically generates the step number if you don't specify it. (Thanks @justinchuby!)- The Python version is now recorded when you run your experiment. One more step towards making your training environment reproducible. (Thanks @kvthr!)
And a few bugs squashed:
- Checkpoints are no longer counted as best if the primary metrics is
NaNorNone. (Thanks @ryanbloom!) - Fix ID generation when random seed is fixed.
Thanks to everyone who suggested features, complained about broken things, and help build things!
Run pip install -U replicate to upgrade. Or, add replicate~=0.3.0 to requirements.txt if you're using that.