I'd like a convenient way to specify dataset-wise parameters, but with possible default experiment-level values.
E.g. training_parameters with lr and batch size at the top-level of a configuration, and then be able to specify training_parameters with a different batch size for one dataset and use the default otherwise.
First thoughts: make all top-level things accessible wrt to the dataset namespace and then update any coincident entries, and otherwise add.
I'd like a convenient way to specify dataset-wise parameters, but with possible default experiment-level values.
E.g. training_parameters with lr and batch size at the top-level of a configuration, and then be able to specify training_parameters with a different batch size for one dataset and use the default otherwise.
First thoughts: make all top-level things accessible wrt to the dataset namespace and then update any coincident entries, and otherwise add.