UIP_EVAL

Training folder structure

configs contains the YAML configuration files used by the matsciml experiment parser. At a high level, a single experiment YAML defines the scope of the experiment (e.g. task, model, and dataset) imperatively. The experiment is then composed by passing definitions for each component, i.e. a path to the LiPS dataset YAML file. An example call looks like this:

python matsciml/experiments/training_script.py \
	-e configs/experiments/faenet_lips_force.yaml \
	-m configs/models/faenet_pyg.yaml \
	-d configs/datasets/lips.yaml \
	-t configs/trainer.yaml

Adding an experiment

1. Copy one of the experiment YAML configs; no hard and fast rule for naming scheme, but to start off we have <model>_<dataset>_<task>.yml just for the ease of access.
2. Modify the keys in the experiment YAML config - the keys must match what are defined in the other configs (e.g. lips refers to the name of the YAML file)
3. Update trainer.yaml as needed: in particular, set the wandb entity to yours!
4. Update the dataset YAML file as needed: pay attention to batch size, and paths.

Common tweaking parameters

• Batch size (per DDP worker) is modified in the dataset YAML.
• Number of workers, epochs, callbacks are configured in trainer.yaml
• Learning rate is configured in the model YAML.