v0.3.0
New features
- Classifier (#124, #137, #151), supports
predict
,predict_proba
methods - Regressor (#129, #137, #151), supports
predict
method - BaseForwardModel (#137, #151) and add
save_pickle
andload_pickle
to it (#139) - Add
target_index
toparse
method ofCSVFileParser
andSDFFileParser
(#131) - Add
extract_total_num
toCSVFileParser
andSDFFileParser
(#131) - Add
target_index
option toget_qm9
(#131) - Add
train_target_index
,val_target_index
, andtest_target_index
toget_tox21
(#131) - Add
dropout_ratio
toRSGCN.__init__
(#146).
Changes that break compatibility
chainer_chemistry.dataset.preprocessors.weavenet_preprocessor.DEFAULT_NUM_MAX_ATOMS
has been removed. Usechainer_chemistry.WEAVE_DEFAULT_NUM_MAX_ATOMS
instead (#127).
Bug fixes
- Fix the behavior of
concat hidden
argument of GGNN (#117, #118) - Fix the order of transpose of GGNN (#143)
- It was a serious bug that degraded the performance of GGNN. (Thank you @ir5 for the bug report!)
- Fix examples to run with Python 2.7 (#144) (Thank you @mihaimorariu)
Example
- Add an example of using own dataset (#114)
- Add an example of Classifier (include saving model by pickle) using Tox21 dataset (#124, #125, #136, #141, #152, #156)
- Add prediction script to the QM9 example (#140)
- Add an example of Regressor (include saving model by pickle) using QM9 dataset (#129, #152, #156).
Others
- Improve document of concat_mols (#116)
- Import without installing RDKit (#119, #123)
- Add introductory slide to README (#130)
- Speed up example test scripts (#132)
- Add tests that check invariance of Graph CNN under graph isomorphism (#145)
- Add tests of
RSGCN
(#146) - Remove import of matplotlib (#148)
- Add a comment on model portability (#150)
- Apply flake8 (#155)
Also thank you @mihaimorariu (#144) and @ir5 for the implementation, documentation, bug report and example improvements!