Docker Image for the qmpy Environment

Someone had trouble configuring the qmpy environment (#1, #2). So, I decided to create a Docker image for the qmpy environment.

Changes from the Previous Release

• Published the Docker image for building the OQMD v1.2 dataset: tonyy999/qmpy-v1.2 (64a1781)
• Added allow_pickle=True to np.load() (0c9c065, 3b407e1)
• Updated the dataset instruction (0a56f88, 7db8313, 4a64f16, db3306a, a7ee460)
• Added Colab notebook links (2a910bb, 0e215c3)

Testing the Docker Image

I used a GCP virtual machine to test the Docker image. A 200GB SSD was chosen as a single boot persistent disk of the VM.

1. Importing the OQMD v1.2 and execution of oqmd_data.py

   • VM: t2d-standard-1 (1 AMD Milan CPU, 4GB RAM)
   • The importing took a hour, and the data extraction by oqmd_data.py took 5 hours 43 minutes.
   • The memory consumption was lower than about 2 GB.
   • The log is available at this link.

2. Execution of mp_graph.py

   • VM: c2d-standard-8 (8 AMD Milan CPUs, 32GB RAM)
   • The graph generation by mp_graph.py took 210 minutes.
   • The memory consumption was lower than 2 GB.
   • The log is available at this link.

3. Execution of oqmd.py

   • VM: t2d-standard-2 (2 AMD Milan CPUs, 8GB RAM)
   • The dataset creation by oqmd.py took about 30 minutes.
   • The memory consumption was lower than 6 GB.
   • The log is available at this link.

4. Validation of the created dataset

   • A validation tool was used to validate the created dataset.
   • Found were 15 mismatches in spacegroup, but the official dataset would have some incorrect space groups due to bugs of the Spglib around 2019. Errata for space groups is available at this link. This problem does not affect the dataset validity because the space groups are not used to transform the other data.

Requirement: PyTorch 1.1+

This version was checked on Colab notebook with PyTorch 1.10:

PyTorch 1.1+ is required to use this version. No user warning appears.

Downloadable OQMD dataset and Colab

• The OQMD v1.2 dataset available at this link.
• A Colab notebook for training a CGNN model:

Initial Release

v1.0.0

Version 1.0.0