search

rampasek / GraphGPS

Recipe for a General, Powerful, Scalable Graph Transformer
MIT License
643 stars 114 forks source link

PCQM4Mv2 cfg to run pretrained GPS model #11

Closed rampasek closed 2 years ago

rampasek commented 2 years ago

Added config file to run inference on PCQM4Mv2 using a pretrained GPS model.

To run the inference:

  1. Download and unzip: https://www.dropbox.com/s/677clfz3cng8xsi/pretrained-GPS-pcqm4m.zip?dl=1 You should now have a folder pretrained/pcqm4m-GPS+RWSE.medium in the root of the project. This is the last epoch checkpoint from the original run used for the GPS paper preprint. It was trained using a random split of the official train set to train and validation set, taking the official valid set as the test set. The official test-dev and test-challenge sets are not used.

  2. Run: python main.py --cfg configs/GPS/pcqm4m-GPS-inference.yaml Note: This will download and process the PCQM4Mv2 dataset first (~1h, when you run it for the first time), then we need to precompute Laplacian decompositions (~2h on 4 core CPU), and finally run the inference.

This checkpoint achieves MAE: 0.0860 on what is here called the test set, which is the official valid set.

rampasek commented 2 years ago

Note, this PR also closes #6 original request.