MP auto download - Githubissues

corochann commented 4 years ago

[x] include changes of current master branch (dataset class are added to support converter)
[x] write download script of materials project dataset
[x] modify data_dir path etc.
[x] include citation/license properly
[x] check code runs
[ ] merge to master branch --> this PR

corochann commented 4 years ago

Implementation design for `converter`

Currently it is difficult to unify deciding which converter to use.

Current implementation is as follows:

NumpyTupleDataset ->use converter_method_dict to switch converter based on model
Other dataset (PaddingDataset, SparseDataset) -> use dataset.converter to use converter on dataset.

I confirmed following example runs.

bash -x test_mp.sh 0: train megnet & cgcnn with mp dataset, on GPU.
bash -x test_qm9.sh 0: train megnet works well, as well as sparse gin_sparse model.
bash -x test_molnet.sh wtih modified --method megnet: train megnet with molnet.
bash -x test_tox21.sh 0: training code works without backward compatibility break.

corochann commented 4 years ago

I dropped testing py27 & py35 as well. When pymatgen is installed, it seems pandas version will change and work only when py36~ is used.

corochann commented 4 years ago

/test

pfn-ci-bot commented 4 years ago

Successfully created a job for commit 34779fc:

corochann commented 4 years ago

/test

pfn-ci-bot commented 4 years ago

Successfully created a job for commit ffe59f9:

corochann commented 4 years ago

Let me re-organize code and submit PR later.