FAIR-Chem Core

2.19.0 verified Sat May 09 auth: no python

fairchem-core is the core package of the FAIR Chemistry library, providing machine learning models (e.g., GemNet-OC, DimeNet++) for chemistry and materials science. Version 2.19.0 supports Python 3.11 to 3.13. The library is actively maintained and released every few months.

pip install fairchem-core

Common errors

error ImportError: cannot import name 'OC20Dataset' from 'fairchem.core.data' ↓

cause Incorrect import path; the dataset classes are in `fairchem.core.datasets`.

fix

Change import to from fairchem.core.datasets import OC20Dataset.

error ModuleNotFoundError: No module named 'fairchem' ↓

cause `fairchem-core` is not installed, or is installed in a different Python environment.

fix

Run pip install fairchem-core and ensure you are using the correct Python interpreter.

error RuntimeError: Expected tensor to be on GPU, but found tensor on CPU ↓

cause Model expects a CUDA tensor, but data was loaded on CPU.

fix

Move model and data to GPU: model = model.cuda(); data = {k: v.cuda() for k, v in batch.items()}.

Warnings

breaking The package structure changed from `fairchem` to `fairchem.core` in version 2.0. Code using `from fairchem import ...` will break. ↓

fix Update imports to use `fairchem.core` instead of `fairchem`.

gotcha Requires PyTorch with CUDA. CPU-only will fail if GPU expects CUDA tensors. ↓

fix Install PyTorch with CUDA: `pip install torch --index-url https://download.pytorch.org/whl/cu118`

deprecated The `OC22Dataset` is deprecated in favor of `OC20Dataset` with appropriate parameters. ↓

fix Use `OC20Dataset` with `task='oc22'` if available.

Imports

OC20Dataset

wrong

from fairchem.core.data import OC20Dataset

correct

from fairchem.core.datasets import OC20Dataset

The correct module is `datasets`, not `data`.

Quickstart

Minimal example: load OC20 S2EF dataset, create a model, and run a forward pass.

import torch
from fairchem.core.models import GemNetOC
from fairchem.core.datasets import OC20Dataset

dataset = OC20Dataset("s2ef", split="train", subset="200k", root="./data")
loader = torch.utils.data.DataLoader(dataset, batch_size=32, shuffle=True)
model = GemNetOC()
output = model(next(iter(loader)))
print(output)