Chainer-chemistry

This is the release note of v0.7.0. See [here](https://github.com/chainer/chainer-chemistry/milestone/8?closed=1) for the complete list of solved issues and merged PRs.

**Note that this is planned to be the final major release.** As [announced in chainer blog](https://chainer.org/announcement/2019/12/05/released-v7.html), further development will be limited to only serious bug-fixes and maintenance.

Highlights
- Graph Conv Neural Network with Sparse pattern is added & SNS dataset is supported to try big sparse graph data training (398, thanks knshnb!)
- See [blog (Japasene)](https://tech.preferred.jp/ja/blog/graph-neural-network-with-sparse-data-structure/) and [slideshare](https://www.slideshare.net/pfi/20190930-pfn-internship-2019-extension-of-chainerchemistry-for-large-amp-sparse-graphkenshin-abe) for further information
- MEGNet, CGCNN model is supported (396, thanks nd-02110114!)

Changes that break compatibility
- Use shifted softplus in SchNet (395)

New features

Models
- SparseRelGCN (398, thanks knshnb!)
- SparseGIN (398, thanks knshnb!)
- GIN with COO matrix (398, thanks knshnb!)
- MEGNet (396, 405, 409, thanks nd-02110114!)
- CGCNN (396, 405, thanks nd-02110114!)

Dataset
- SNS Dataset support is added to try big sparse graph data
- Citation Network: cora, citeseer (398, 402, thanks knshnb!)
- Reddit (398, 402, thanks knshnb!)

Dataset class

- SparseGraphDataset (398, thanks knshnb!)
- PaddingGraphdataset (398, thanks knshnb!)

See table below for usage

| | Chemical | Network
----|----|----
| adjacency matrix | `NumpyTupleDataset` | **`PaddingGraphDataset`** |
| scatter operation | **`SparseGraphDataset`** | **`SparseGraphDataset`** |
| sparse matmul | not supported |**`PaddingGraphDataset (use_coo=True)`** |

Function
- Add shifted softplus (395)
- Used in SchNet

Examples
- `network_graph` example is added for `cora, citeseer, reddit` dataset training (398, thanks knshnb!)

This is the release note of v0.6.0. See [here](https://github.com/pfnet-research/chainer-chemistry/milestone/7?closed=1) for the complete list of solved issues and merged PRs.

Changes that break compatibility
- [Big Change] Unify arguments in model, update and readout (368)

- return svg text instead of ipython SVG object in `MolVisualizer` and `SmilesVisualizer` (388)
- Support ChainerX (376)
- fix typo (358, Thank you shllln)

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

New features

Model, Link
- Add Message Passing Neural Network(MPNN) (312)
- Add SparseGGNN (342, Thank you ir5)
- Add GNN-FiLM (386)
- Add GraphWarpModule(GWM) (321, 340, 368, 372)
- Add GWMGraphConvModel
- Add Graph Isomorphism Network(GIN) (321, 337, 345, 371, Thank you TaizoAyase)

Scaler
- FlowScaler (318)
- Support more than 3-dim array by StandardScaler (348)
- MinMaxScaler (351)
- MaxAbsScaler (350)

Utils
- GraphConvPredictor (317)
- set_up_predictor (336)
- AutoPrintReport (373)
- Train Utils (382)

Docker
- Add official dockerfile (364, 365, 374, 381, Thank you cks-coil, zaltoprofen)

Development Utils
- typing module (331)
- document checker (332)

Bug-fix
- addHs in SchNet preprocess (390)
- weave preprocessor (360)
- fix the type of return value from R2ScoreEvaluator (339)

Tests
- Use FlexCI (378, 381)
- GWMPreprocessor tests (383)
- Fix some tests (377)

This is the release note of v0.5.0. See [here](https://github.com/pfnet-research/chainer-chemistry/milestone/6) 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)

This is the release note of v0.4.0. See [here](https://github.com/pfnet-research/chainer-chemistry/milestone/5?closed=1) 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](https://arxiv.org/abs/1703.00564) 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)

Bug fixes
- 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)

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

Support version
- Remove Chainer v2 (194) and add Chainer v3 from Travis CI tests (195)
- Freeze RDKit version in Travis CI to 2017.09.3.0 (227)
- Add description of supported library versions (232)

Example
- 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)

Others
- Change assertion messages for preprocessors (162, Thank you mihaimorariu)
- Fix typo in `datasets/qm9.py` (176, Thank you natsukium)
- Retrieve version number in setup.py without imp (198)
- Add .pytest_cache to .gitignore (199)
- Fix document formats (218)

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` and `load_pickle` to it (139)
- Add `target_index` to `parse` method of `CSVFileParser` and `SDFFileParser` (131)
- Add `extract_total_num` to `CSVFileParser` and `SDFFileParser` (131)
- Add `target_index` option to `get_qm9` (131)
- Add `train_target_index`, `val_target_index`, and `test_target_index` to `get_tox21` (131)
- Add `dropout_ratio` to `RSGCN.__init__` (146).

Changes that break compatibility
- `chainer_chemistry.dataset.preprocessors.weavenet_preprocessor.DEFAULT_NUM_MAX_ATOMS` has been removed. Use `chainer_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!

This is the release of v0.2.0.

New features
- Add `RSGCN` (Renormalized Sepectral Graph Convolutional Network) (89 thank you anaruse!)
- Add `BalancedSerialIterator` for imbalanced data training (59)
- Add `ROCAUCEvaluator` for binary classification task evaluation (62)
- Add `self_connection` option in `construct_adj_array` (100)

Changes that break compatibility
- `return_smiles` return `numpy.ndarray`, instead of `list` (79)
- `csv_file_parser.parse` and `sdf_file_parser.parse` methods return `dict` that contains dataset and smiles, instead variable length `tuple` (94)

Bug fixes
- QM9 dataset ordering (107)

Example
- Support `SchNet` inference example in tox21 example (103)
- Add `RSGCN` example in QM9, tox21 example (89, 104)
- Add `BalancedSerialIterator`, `ROCAUCEvaluator` sample usage in tox21 example (60, 62)

Others
- Dataset exploration notebook in QM9, tox21 example (78)
- Add many tests.

Also thank you amaotone (50), kazuyaujihara (85), msakai (88) and anaruse (89) for the implementation, documentation and example improvements!