This is the release note of v0.4.0. See here for the complete list of solved issues and merged PRs.

New features

MoleculeNet dataset support (#165, #179, #181, #182, #187, #205, #209, #216, #230, #233)
- Examples using the MolculeNet dataset are also provided.
- Settings recommended in the original paper is used.
Splitter support
- RandomSplitter (#196)
- StratifiedSplitter (#201)
- ScaffoldSplitter (#202)
Loss function implementation which deals with NaN
- mean_squared_error (#190)
- mean_absolute_error (#193)
More evaluators
- BatchEvaluator (#210)
- PRCAUCEvaluator (#210)
More parsers
- DataFrameParser (#203, #207)
- SMILESParser (#204)
Add return_is_successful option to parsers (#219 #220)
Add extract_total_num (#221)
Add raise_value_error to ROCAUCEvaluator (#157)

Changes that break compatibility

Set default argument for the script option of examples/own_dataset/train.py (#171)
Fix indexing behavior in NumpyTupleDataset when data_index is a list of length 1 (#200)
Modify total count behavior of SDFFileParser, which only affects to logging output (#220)

Use non-isomeric SMILES (#237, #240)
Remove NaN value in an adjacency matrix when out_size is larger than the number of atoms (#169, Thank you @mihainorariu)

Improvements on example test scripts (#158, #188)
Other test improvements
- #170, #197, #208, #223, #234
- #163, #168, #174, #175 (Thank you @mihaimorariu)

Model Evaluation script
- Tox21 (#147)
- QM9 (#189, #225, #226 Thank you @mihaimorariu)
Remove InferenceLoop from the Tox21 example (#184, #222)
Remove conditional import of Classifier from the Tox21 and QM9 examples (#185)