keepsake
keepsake copied to clipboard
replicate
Version control for machine learning
# Why? We haven't talked to an ML practitioner using Windows yet, but we know it is becoming a more popular platform and we should support it. Please chime in...
This is currently hard to scan and parse, particularly on Colab where there is no CSS styling at all. Some simple improvements: - Some basic styling - Format params as...
# Why Often, a set of experiments are grouped together as a particular line of enquiry. For example, you might be doing a hyperparameter sweep, where all the experiments are...
# Why Replicate is a useful way of passing models from training into production. # How Document how to load models in an inference server, and perhaps integration with production...
We should validate the names you put in a filter string, e.g. `replicate ls --filter "foo
If a command fails to run, it spews tracebacks instead of printing clearly what the output of the command was.
It looks like MD5 hashes are being used for [file paths](https://github.com/replicate/replicate/blob/main/go/pkg/repository/disk.go#L211-L222). Given what has already been learned about this algorithm, is it worth using something more collision resistant like [SHA256](https://pkg.go.dev/crypto/sha256)...
# Why Our support for versioning training data is [merely some documentation about how to do that with `params`](https://replicate.ai/docs/guides/training-data). Training data is so fundamental to machine learning, and is such...
# Why This would be useful to integrate with spreadsheet-based workflows. # How Probably something like this: ``` replicate ls --csv ```
[We provide some general advice about how to version data](https://replicate.ai/docs/guides/training-data), but we should give some specific examples about how to version data from things like https://dvc.org/ or https://feast.dev/ .