Skip to main content

Changing Profile

Profiles are classy's way of handling custom configurations. There are two ways in which you can use a profile:

  1. using a single profile file (e.g., configurations/profiles/custom.yaml), which can override every single pre-existing configuration value;
  2. using a profile file in combination with hydra's config groups.
caution

Your custom profiles must be stored under configurations/profiles, or classy won't be able to discover them.

Single-file profile​

Say that you are implementing a custom Named Entity Recognition model. In your profile, you are going to specify your model's configuration explicitly, and it will be overridden over classy's default.

configurations/profiles/single-file.yaml
model:
  _target_: src.model.MyCustomClassyNERModel
  parameter1: ...
  parameter2: ...

Then, to train using your model, you just run classy train token data/ner -n custom-experiment --profile single-file.

info

This kind of composition is very similar to hydra's default behaviour, but with an important difference: when using profiles, configurations are not additive but exclusive (if they have a _target_ property): for example, if your model defines an optim_conf, such configuration will entirely replace the previously existing optim_conf, instead of adding its config parameters to it.

Exploiting config groups​

Single file configs are prone to become very large very quickly. Thus, we extend hydra's behaviour to support config groups specification even outside the defaults list. In a nutshell, this means that you can have a config file under configurations/models/custom-model.yaml that is referred by configurations/profiles/grouped.yaml.

configurations/models/custom-model.yaml
model:
  _target_: src.model.MyCustomClassyNERModel
  parameter1: ...
  parameter2: ...
configurations/profiles/grouped.yaml
model: custom-model