Skip to content

First experiment

This walkthrough uses examples/va_estimation, because it is a real plugin package in this repository. It is a template for the framework workflow, not a zero-download demo: its configs require the VA datasets, features, and paths listed in the example README.

Project shape

The important pieces are:

examples/va_estimation/
├── configs/
│   ├── multimodal_train.yaml
│   └── multimodal_test.yaml
├── src/
│   └── chimera_plugin.py
└── pyproject.toml

pyproject.toml exposes the plugin as va_estimation = "chimera_plugin:register". The register function imports the modules that populate Chimera ML registries.

Configure

Install the plugin and edit the dataset, feature, annotation, and optional notification paths in the config:

python -m pip install -e examples/va_estimation
chimera-ml validate-config \
  --config-path examples/va_estimation/configs/multimodal_train.yaml

An experiment config has named components such as data, model, loss, and optimizer, each using name plus params. metrics, callbacks, and logging are lists of the same shape. See Configuration for the contract.

Train

chimera-ml train \
  --config-path examples/va_estimation/configs/multimodal_train.yaml

The CLI creates a per-run BuildContext, builds each component from its registry, and calls Trainer.fit. The checkpoint and snapshot callbacks write under the configured log root. The exact output name is controlled by callback parameters and the generated run name.

Evaluate

chimera-ml eval \
  --config-path examples/va_estimation/configs/multimodal_test.yaml \
  --checkpoint-path path/to/checkpoint.pt

Evaluation loads either a checkpoint containing model_state_dict or a raw state dictionary, then evaluates the merged train/validation/test loader splits that the datamodule exposes. The checkpoint must match the model strictly.

Next steps