v0.5.0

This is the release note of v0.5.0. See here for the complete list of solved issues and merged PRs.

Independent graph convolution layers

• Separate links to modules (#270, #301, #305)
  • This is for users to define their own network easily.

New features

Model, Link

• Add RelGCN (#269, #316)

• Add RelGAT (#217, #299, #302, #315)

• GGNN: support num_edge_type as argument (#294)

• Add StandardScaler link (#309)

• Add GraphMLP (#295)

• GGNN input size invariant support (#297)

• NFP input size invariant support (#296)

Functions

• softmax with mask (#306)

Dataset

• Add PDBBind dataset (in MoleculeNet) (#255)
• Add Zinc 250k dataset (#276)
• Add kekulize option in preprocessor (#262)

Saliency modules

• Add saliency modules (#283, #284, #286)

Utils

• util function for save/load files to json (#303)

Evaluator

• Add R2 score evaluator (#253)

Splitter

• Add time order splitter (#254)

Changes that break compatibility

• always return canonical smiles from parser (#264)
• move construct_discrete_edge_matrix to common (#260)
• Separate links to modules (#270)

Bug fixes

• Inverse scaling for predicted values (#250, Thank you @kazuyaujihara)
• Fix FutureWarning in permute_adj (#277)
• report metrics as scalars (#282)
• Rename device_id inside BaseForwardModel for chainer v6 (#280)

Tests

• Improvements on daily test (#243)
• Improvements on slow preprocessor test (#263)

Example

• Add example for own dataset (#249)
• update shell script for examples (#266, #273, #275)
• QM9
  • Scale output labels in QM9 example (#256)
  • Evaluate each label separately (#314)

Others

• Fix documents (#241, #242)
• Support features indexer with length 0 (#246)
• Update README (#265)
  • Fix rdkit version (#247)
  • Add research projects list (#248)