Question: How can we serialize parameters that shold be kept for inference?

[bilzard]

The document says the serialization API will be removed in future.
https://brentyi.github.io/tyro/api/tyro/extras/_serialization/#tyro.extras._serialization.to_yaml

However, in some use case like architecture search, we want to define model parameter like number of layer, depth etc. in config.
In this situation, we also need these parameter in inference. So I think we should keep these parameter in file.

The question is, what we can do in this situation? Are there more sophisticated solution than just write custom code to serialize these parameters on our own?

[brentyi]

If you're using standard dataclasses for configuration there are a lot of options!

In my own code I've been using the standard yaml.dump for writing to file:
https://github.com/brentyi/egoallo/blob/4663a8ec5bc4a0875a09ed39ef103c34ac2a718f/1_train_motion_prior.py#L98

And then yaml.load for reading:
https://github.com/brentyi/egoallo/blob/4663a8ec5bc4a0875a09ed39ef103c34ac2a718f/src/egoallo/inference_utils.py#L27-L29

dacite is also very popular!