Quickstart¶
This page loads a published Rem3Di model and turns a small dataset of molecules into descriptor vectors.
Prerequisites¶
- Installed
remedi. - A published model directory (see layout).
- A MoleculeDataset (zarr) to embed. Don't have one? See Prepare a dataset, or use one of the example datasets.
Get descriptors¶
RemediCalculator goes from a model directory to descriptors: it loads the
checkpoint and embeds a dataset for you.
from remedi.evaluation.benchmark.descriptors import RemediCalculator
from remedi.data_handling.dataset.molecule_dataset import MoleculeDataset
# 1. Point the calculator at a published model directory.
calc = RemediCalculator(
model_dir="/path/to/published_model",
device="cuda", # or "cpu"
)
# 2. Open a dataset and embed it.
dataset = MoleculeDataset.open_existing_dataset_from_dir(
"/path/to/dataset_zarr"
)
descriptors = calc.calculate(dataset) # numpy array (N_molecules, D)
print(descriptors.shape)
descriptors has one row per molecule; feed it straight into any downstream model.
Running on a different machine than the model was trained on? The MACE path baked into the config may not exist locally, so pass
RemediCalculator(model_dir=..., mace_model_path="/local/MACE.model"). See Concepts: Foundation-model frontend.
Next steps¶
- Train a downstream model on these descriptors.
- Evaluate a model across many benchmarks at once.
- Notebook:
examples/01_get_descriptors.ipynb.