Defaults for CLI predict configs: default "obs_names_n" batch_key and default PredictionWriter

[sjfleming]

I sometimes forget that I need to specify obs_names_n as a "batch_key" when writing a yaml config for prediction. Since this is always the same and always required for prediction, isn't it possible to inject it as a default when running the predict subcommand?

  batch_keys:
    ...
    obs_names_n:
      attr: obs_names

Similarly, is it possible to include a default PredictionWriter callback (that can be overwritten by the yaml) whenever the CLI subcommand is predict?

I guess my ultimate goal would be this:
For most cases, you can actually use the same yaml config file for train and predict. This might be enabled if, for example, you could give a couple extra inputs at the command line at runtime to specify things like ckpt_path.

[sjfleming]

I fooled with this for a while and it seems to be less trivial than I thought. I can add defaults for something like "data.batch_keys", but if for example I put this in NewLightningCLI.add_arguments_to_parser

            if self.subcommand_is_predict:
                parser.set_defaults(
                    {
                        "data.batch_keys": {"obs_names_n": {"attr": "obs_names"}},
                    }
                )

then that becomes the default... but if you specify any data.batch_keys in your config file (which you will), then this default gets totally overridden, not added to.

If I try to add a default just for obs_names_n directly like this

            if self.subcommand_is_predict:
                parser.set_defaults(
                    {
                        "data.batch_keys.obs_names_n": {"attr": "obs_names"},
                    }
                )

then I get some error about

jsonargparse._namespace.NSKeyError: No action for key "data.batch_keys.obs_names_n" to set its default.

which I think means that the parser considers data.batch_keys to be an input argument, but does not consider data.batch_keys.obs_names_n to be an input argument.