dotnet / machinelearning

ML.NET is an open source and cross-platform machine learning framework for .NET.
https://dot.net/ml
MIT License
8.94k stars 1.86k forks source link

AutoML: consider adding precision in checkpoint file #6802

Open torronen opened 11 months ago

torronen commented 11 months ago

Nightly, VS 2022, .NET 7.0

The checkpoint file has columns with 3-digit precision. In some cases I see loss being exactly the same on every run in the checkpoint file. I would like to confirm it it is exactly (or almost) same prediction. Therefore I would like to see 5-digit values in this file.

I think, in difficult problems it may also help the tuner to go to correct direction.

I know it is a minor issue to go 0.01% level but I also think there is no downside to adding precision to this file.

Example Cost-frugal tuner. Are these results not improving because of the type of the data, or because the changes to params are too small? image