[Task]: Configuration sharing (like _BASE_ in detectron2) + kwargs overwriting

[DominikFilipiak]

What is the motivation for this task?

1. When doing experiments, we often make small changes to the YAML config by hand and see what's the output of the model.
   These configs will share the vast majority of their parameters.
   Since the number of configs quickly rises up, it's easy to get lost.
   Therefore, I think it would be nice to have a config inheritance mechanism.
   So, we'd have a base config, and a set of additional configs.
   We pass additional config to the train/test script and pull everything from base (location defined by, say, _BASE_ keyword in additional config) and overwrite keys defined in additional.
   This forms a combined config, which is the final one.

2. Additionally, if a user will provide a keyword argument (e.g. --ops model.n_features=50 model.backbone=resnet18) to the script, it should overwrite values in the combined config.
   I think it should be implemented in train/test scripts

Describe the solution you'd like

It's nicely done in detectron2 (both of my remarks are, actually) with their YACS configs.

In short, in their YAML config, there's a top-level key _BASE_, which specifies the base config path.
You may also want to check load_yaml_with_base function definition.

Here's an example -- _BASE_ is defined on the first line.

With regard to the second remark, they provide scripts which can be used like this (see again YACS configs).

/demo.py --config-file config.yaml [--other-options] \
  --opts MODEL.WEIGHTS /path/to/weights INPUT.MIN_SIZE_TEST 1000

Additional context

No response

[samet-akcay]

@DominikFilipiak, we are already working on this, which will be available in v1.0.