Ways to export/import network weights

Question

Ways to export/import network weights

samuelvenzi opened this issue 4 years ago · comments

Does Goro support any way of exporting/saving and then importing the trained model?

Patrick Barker · Answer 1 · Tue Jan 26 2021 06:53:49 GMT+0800 (China Standard Time)

Gorgonia has a way of doing this https://gorgonia.org/how-to/save-weights/ I haven't worked this in as first class citizen, but you should be able to call Learnables() to get the differentiable weights and use gob to save them.

If you wanted to write a cleaner method for this I would be happy to review the PR!

Bruno Andreghetti Dantas · Answer 2 · Thu Jan 28 2021 07:59:19 GMT+0800 (China Standard Time)

Saving the Learnables() was pretty straight-forward, but is there any method to import them after loading them from the file?

Patrick Barker · Answer 3 · Thu Jan 28 2021 11:24:19 GMT+0800 (China Standard Time)

https://github.com/aunum/goro/blob/master/pkg/v1/model/model.go#L606 is probably the best example, that could be refactored to use a LoadLearnables() or SetLearnables() method.

Samuel Venzi · Answer 4 · Sat Jan 30 2021 07:23:53 GMT+0800 (China Standard Time)

Hey, I just created a SetLearnables() method inspired by CloneLearnablesTo(). I was trying, at first, to do the same thing outside Goro, but the need to access some unexported fields, like trainBatchChain, made me go for a fork and a PR #3. If you could review it please @pbarker.

Patrick Barker · Answer 5 · Sat Jan 30 2021 11:10:19 GMT+0800 (China Standard Time)

thanks for the contribution! closing with merge #3