graph-pes-test

Use the graph-pes-test command to test a trained model. Testing functionality is already baked into graph-pes-train, but this command allows you more fine-grained control over the testing process.

Usage

graph-pes-test test-config.yaml model_path=path/to/model.pth

graph-pes-test -h

usage: graph-pes-test [-h] [args ...]

Test a GraphPES model using PyTorch Lightning.

positional arguments:
args        Config files and command line specifications.
            Config files should be YAML (.yaml/.yml) files.
            Command line specifications should be in the form
            my/nested/key=value. Final config is built up from
            these items in a left to right manner, with later
            items taking precedence over earlier ones in the
            case of conflicts. The data2objects package is used
            to resolve references and create objects directly
            from the config dictionary.

optional arguments:
-h, --help  show this help message and exit

Example

data:
    dimers:
        +file_dataset:
            path: path/to/dimers.xyz
            cutoff: 5.0
    amorphous:
        +file_dataset:
            path: path/to/amorphous.xyz
            cutoff: 5.0

accelerator: gpu

loader_kwargs:
    batch_size: 64
    num_workers: 4

Config

class graph_pes.config.testing.TestingConfig

Configuration for testing a GraphPES model.

model_path: str

The path to the model.pt file.

data: GraphDataset | dict[str, GraphDataset]

Either:

• a single GraphDataset. Results will be logged as "<prefix>/<metric>".

• a mapping from names to datasets. Results will be logged as "<prefix>/<dataset-name>/<metric>", allowing for testing on multiple datasets.

loader_kwargs: dict[str, Any]

Keyword arguments to pass to the GraphDataLoader.

Defaults to:

loader_kwargs:
    batch_size: 2
    num_workers: 0

You should tune this to make testing faster.

torch: TorchConfig

The torch configuration to use for testing.

logger: Literal['auto', 'csv'] | dict[str, Any] = 'auto'

The logger to use for logging the test metrics.

If "auto", we will attempt to find the training config from <model_path>/../train-config.yaml, and use the logger from that config.

If "csv", we will use a CSVLogger.

If a dictionary, we will instantiate a new WandbLogger with the provided arguments.

accelerator: str = 'auto'

The accelerator to use for testing.

prefix: str = 'testing'

The prefix to use for logging. Individual metrics will be logged as <prefix>/<dataset_name>/<metric>.