Save/Load Model?

[anchen1011]

Will there be an API for saving and loading models in the future?
It was brought up in #6951 but not actually considered carefully.

Currently we have only:

1. Save/Load checkpoints, which is epoch-dependent
2. Save/Load params, which is not exactly save/load models, and is very unstable cross versions

Thanks!

[ThomasDelteil]

If your model is hybridizable, and you want to shelf it for a while, the recommended way would be to use .export('prefix', epoch=0).
That way you can reload it later in a symbol block using SymbolBlock.imports() or in a different language binding altogether.

If you are iterating with your model, the recommend way is save_params/load_params for the parameters, and save the model definition file in a .py file or simply pickle it.

You can check this upcoming tutorial. https://github.com/indhub/mxnet/blob/ef1a5fb8ecfaf9f85c26a468cdbcfbfbab58c423/docs/tutorials/gluon/save_load_params.md

[lanking520]

Thank you for submitting the issue! I'm labeling it so MXNet community members can help resolve it. @nswamy can you add 'Question', 'Disscussion' here?

[ThomasDelteil]

@szha @piiswrong is it safe to pickle a model to store the network definition? And later reload the params using net.load_parameters() ? Could we imagine having a net.to_pickle() in the future? What would be the gotchas to consider when doing this ?

[Ishitori]

@anchen1011 Did it work for you? If yes, please, close the issue.

[thomelane]

@szha @piiswrong any comments to the above? otherwise can we close this issue due to not hearing back from author, thanks!

@anchen1011 can be reopened if .export doesn't work for you, cheers.