- Python >= 3.6
- PyTorch >= 1.2
- Horovod >= 0.19
- opencv-python
- h5py
- Python >= 3.6
- TensorFlow == 1.14
- Horovod >= 0.16
Download the classificatoin dataset from 10.13139/OLCF/1510313 via Globus. If you have access to OLCF machines, the preprocessed data are available at
classification: /gpfs/alpine/world-shared/stf011/junqi/stemdl/classification/data
reconstruction: /gpfs/alpine/world-shared/stf011/junqi/stemdl/dataThe train-test data split is with respect to different materials.
- Clone the repo:
git clone --recursive https://github.com/at-aaims/stemdl-benchmark- Run the space-group classification:
classification/source/run.sh is the working job script on Summit, and the training is launched via
jsrun -n<NODES> -a1 -g1 -c42 -r1 -b none python sgcl_runner.py --epochs 10
--batch-size 32
--train-dir <path-to-train-dataset>
--val-dir <path-to-val_dataset> - Run the potential reconstruction:
reconstruction/job.lsf is the working job script on Summit, and the training is launched via
CMD="python -u stemdl_run.py --hvd_fp16
--hvd_group 1
--nvme
--filetype "lmdb"
--data_dir <path-to-train-data>
--fp16
--mode "train"
--batch_size 1
--network_config "json_files/network_FCDenseNet_pool_avg.json""
--ilr 1.e-6
--bn_decay 0.1
--scaling 1.0
--input_flags "json_files/input_flags.json"
--hyper_params "json_files/hyper_params.json" "
jsrun -n<NODES> -a 6 -c 42 -g 6 -r 1 --bind=proportional-packed:7 --launch_distribution=packed stdbuf -o0 utils/launch.sh "${CMD}"- A universal classifier for the space group of solid-state materials.
-
Due to the intrinsic imbalance of the crystal space group distribution in nature, the classes in the dataset are also imbalanced. We use both top1 accuracy and F1 score (Macro) to measure the model performance and the example implementation provides a baseline validation top1 accuracy ~0.57 and F1 score ~0.43. Note that the train-test data split is with respect to different materials, i.e. the validation is on materials hasn't been seen by the model.
-
Considering the application of the pre-trained model at the edge, other metrics of interest are the model size and inference time. Generally, for the same performance metric as above, the smaller the model size, the better.