Skip to content

Releases: chemprop/chemprop

v2.1.2

03 Mar 19:32
@akshatzalte akshatzalte
v2.1.2
1229525
Compare
Choose a tag to compare
v2.1.2 Latest
Latest

What's Changed

Important changes

  • CLI implementation of RIGR as an option in --multi-hot-atom-featurizer-mode by @akshatzalte in #1172

    A new featurization scheme, RIGR (Resonance Invariant Graph Representation), is now available. To access it via the CLI, use --multi-hot-atom-featurizer-mode rigr. This featurizer uses only resonance invariant features so it treats all resonance structures of a molecule identically. It uses a subset of the atom and bond features from the default v2 featurizer. With 60% fewer features, RIGR has shown comparable or superior performance across a variety of property prediction tasks in a forthcoming manuscript. An example Jupyter notebook is also provided.

Other changes

New Contributors

Full Changelog: v2.1.1...v2.1.2

Contributors

craabreu, akshatzalte, and 2 other contributors
Assets 2
Loading

v2.1.1

29 Jan 01:07
@KnathanM KnathanM
v2.1.1
1b8f413
Compare
Choose a tag to compare
v2.1.1

Notable changes

In #1090, we started the process of integrating logging into the core code. This will make it easier for users to control what information Chemprop prints to output. It will also make it easier for developers to include more information outputs for potential debugging.

Scipy 1.15 subtly change how logit works which caused some of our tests to fail (as the values reported were slightly different than before). The expected test values have been updated. #1142

A new example notebook has been added which demonstraits how to adapt Chemprop to work with Shapley value analysis. This is another method to lend some intepretability to Chemprop models by highlighting which atom/bond features are most impactful to the final prediction value. #938

We continue to try to make chemprop easy to use. In #1091 and #1124 we added better warnings and error messages. And in #1151 we made is easy to open the example notebooks in Google Colab. This allows people reading the docs to immediately jump in and try chemprop without needing to set up a python environment.

Bug Fixes

In #1097, we fixed a bug where the transforms for scaling extra features/descriptors were turned off during validation. This caused models trained with these extra inputs to not report accurate metrics during training, which is a problem if the "best" model is selected instead of the last model as is done in hyperparameter optimization. Training a model and using the last model was unaffected as was doing inference.

#1084 fixed a bug where R2Score did not have the attribute task_weights. This attribute is not used but is needed for compatability with other metrics

In v2.1 we transitioned to using torchmetrics for our metrics and loss functions, in part because it takes care of training across multiple nodes (DDP) automatically. Our custom metric for Matthew's correlation coefficient however was not set up the way torchmetrics expected. This was fixed in #1131.

What's Changed

Full Changelog: v2.1.0...v2.1.1

Contributors

JacksonBurns, oscarwumit, and 3 other contributors
Loading

v2.1.0

23 Oct 21:35
@shihchengli shihchengli
v2.1.0
bf00b39
Compare
Choose a tag to compare
v2.1.0

The v2.1 release adds the uncertainty quantification modules, including estimation, calibration, and evaluation (#937). For more details on uncertainty quantification in Chemprop, please refer to the documentation and the example notebook. Additionally, we switched the loss functions and metrics to torchmetrics (#1022). With this change we also changed the "val_loss" reported to be calculated the same as the training loss to make them comparable (#1020). We also changed Chemprop to use replicates instead of cross validation (#994) and batch normalization is now disabled by default (#1058).

Core code changes

  • The validation_loss_function is removed in #1023.
  • The batch norm is disabled by default in #1058
  • An new predictor, QuantileFFN, is added in #963
  • BinaryDirichletLoss and MulticlassDirichletLoss are integrated into DirichletLoss in #1066
  • The split type of CV and CV_NO_VAL are removed in #994
  • A models list of metric is now registered as children modules in #1020

CLI changes

  • Disable batch norm by default, and it can be turned on by --batch-norm #1058
  • Many CLI flags related to uncertainty quantification are added #1010
  • Quantile regression is now supported via -t regression-quantile #963
  • The cross validation (CV) is replaced with replicates. The number of replicates can be specified via --num-replicates and the flag --num-folds is deprecated #994
  • --tracking-metric is added which is the metric to track for early stopping and checkpointing #1020

New notebooks

  • An notebook showing interoperability of Chemprop featurizer w/ other libraries (DGL and PyG) #1063
  • Active learning #910
  • Uncertainty quantification #1071

CI/CD

  • Ray can be tested on Python 3.12 #1064
  • USE_LIBUV: 0 is added into the CI workflow #1065

Backwards Compatibility Note

Models trained with v2.0 will not load properly in v2.1 due to the loss functions file being moved. A conversion script is provided to convert a v2.0 model to one compatible with v2.1. Its usage is python chemprop/utils/v2_0_to_v2_1.py <v2_0.pt> <v2_1.pt>

data.make_split_indices now always returns a nested list. Previously it would only return a nested list for cross validation. We encourage you to use data.make_split_indices(num_replicates=X) where X is some number greater than 1, to train on multiple splits of your data to get a better idea of the performance of your architecture. If you do use only one replicate, you will need to unnest the list like so:

train_indices, val_indices, test_indices = data.make_split_indices(mols)
train_data, val_data, test_data = data.split_data_by_indices(
    all_data, train_indices, val_indices, test_indices
)
train_data, val_data, test_data = train_data[0], val_data[0], test_data[0]

What's Changed

Full Changelog: v2.0.5...v2.1.0

Contributors

jonwzheng, JacksonBurns, and 7 other contributors
Loading

v2.0.5

13 Oct 02:00
@shihchengli shihchengli
v2.0.5
09b764b
Compare
Choose a tag to compare
v2.0.5

We continue to enhance and improve the functionality and usability of Chemprop. If there are things you'd like to see addressed in a future update, please open an issue or PR.

Core code changes

We discovered that our Noam learning rate scheduler does not match what was originally proposed. The current scheduler does work well though, so it was decided to not change the definition. Instead the scheduler was renamed and refactored to be more clear. By @shihchengli in #975
Work on uncertainty quantification methods revealed that our previous prediction tensor return dimensions would cause difficulty down the line. Now we have placed uncertainty into a separate dimension. By @hwpang in #959
The BinaryDirichletFFN and MulticlassDirichletFFN predictors were added early in the v2 development, but not tested. Now they have been tested and corrected. By @shihchengli in #1017
The RDKit 2D molecular featurizer was added back by popular demand. The versions used in v1 are available as well as a version that uses all available molecular features in rdkit.Chem.Descriptors. By @KnathanM in #877

CLI changes

  • Log statistical summary of training, validation, and test datasets by @donerancl in #882
  • Change the default verbose level to INFO by @shihchengli in #953
  • Save both probabilities and class label for multiclass classification by @shihchengli in #987
  • Add --remove-checkpoints flag to opt out of saving checkpoints by @shihchengli in #1014
  • Add --class-balance flag to train CLI by @shihchengli in #1011
  • Save target column names in model for use at inference by @hwpang in #935
  • Fix save-smiles-splits not working with rxn. columns as column header by @jonwzheng in #998

Transfer learning

  • Add new example notebook for transfer learning by @joelnkn in #904
  • Use pre-train output scaler to scale training data in CLI by @KnathanM in #1051
  • Add --checkpoint and --freeze-encoder flags in train CLI for transfer learning by @shihchengli in #1007

Documentation

  • Fixed typos in CLI reference and standardized formatting by @donerancl in #880
  • Example Notebook for Classification by @twinbrian in #1047
  • Improve frzn-ffn-layers description and update doc for transfer learning by @oscarwumit in #993
  • add transform tests by @KnathanM in #955
  • Add documentation for how to use a separate splits file (CLI) by @KnathanM in #1041

Other small bug fixes

  • Convert v1 models trained on GPU by @KnathanM in #978
  • Fix hpopting Notebook and CLI for Windows by @JacksonBurns in #1034
  • Update multiclass data to be compatible with rdkit 2024.09.1 by @jonwzheng in #1037
  • Define task_weights if it is None in MulticlassClassificationFFN by @shihchengli in #988
  • change installed torch version on windows actions again by @KnathanM in #1016
  • Update batch norm freezing to freeze running stats by @joelnkn in #952
  • Pass map_location through load_submodules() to torch.load() by @shihchengli in #1029
  • fix no-header-rows in predict command error by @sunhwan in #1001

New Contributors

Full Changelog: v2.0.4...v2.0.5

Contributors

sunhwan, jonwzheng, and 8 other contributors
Loading

v2.0.4

31 Jul 01:03
@shihchengli shihchengli
v2.0.4
2459752
Compare
Choose a tag to compare
v2.0.4

Enhancements and New Features

This release introduces several enhancements and new features to Chemprop. A notable addition is a new notebook demonstrating Monte Carlo Tree Search for model interpretability (see here). Enhancements have been made to the output transformation and prediction saving mechanisms for MveFFN and EvidentialFFN. Additionally, users can now perform predictions on CPU even if the models were trained on GPU. Users are now also warned when not using the TensorBoard logger, helping them to be aware of available logging tools for better monitoring.

Bug Fixes

Several bugs have been fixed in this release, including issues related to Matthews Correlation Coefficient (MCC) metrics and loss calculations, and the behavior of the CGR featurizer when the bond features matrix is empty. The task_weights parameter has been standardized across all loss functions and moved to the correct device for MCC metrics, preventing device mismatch errors.

What's Changed

  • Standardize task_weights in LossFunction across all loss functions by @shihchengli in #941
  • Improve output transformation and prediction saving for MveFFN and EvidentialFFN by @shihchengli in #943
  • Enable CPU prediction for GPU-trained models by @snaeppi in #950
  • Fix Issues in MCC Metrics and Loss Calculations by @shihchengli in #942
  • Fix docs building by pinning sphinx-argparse by @jonwzheng in #964
  • Add Monte Carlo Tree search notebook for interpretability by @hwpang in #924
  • Fix CGR featurizer behavior when bond features matrix is empty by @jonwzheng in #958
  • Fix Failing CI for torch==2.4.0 on Windows ray[tune] Tests by @JacksonBurns in #971
  • warn users when not using tensorboard logger by @JacksonBurns in #967
  • Bug: Move task_weights to 'device' for MCC metrics by @YoochanMyung in #973

New Contributors

Full Changelog: v2.0.3...v2.0.4

Contributors

YoochanMyung, jonwzheng, and 4 other contributors
Loading

v2.0.3

28 Jun 22:46
@KnathanM KnathanM
v2.0.3
e5aae85
Compare
Choose a tag to compare
v2.0.3

Notable changes

The mfs argument of MoleculeDatapoint was removed in #876. This argument accepted functions which generated molecular features to use as extra datapoint descriptors. When using chemprop in a notebook, users should first manually generate their molecule features and pass them into the datapoints using x_d which stands for (extra) datapoint descriptors. This is demonstrated in the extra_features_descriptors.ipynb notebook under examples. CLI users will see no change as the CLI will still automatically calculate molecule features using user specified featurizers. The --features-generators flag has been deprecated though in favor of the more descriptive --molecule-featurizers. Available molecule features can be found in the help text generated by chemprop train -h.

The default aggregation was changed to norm in #946. This was meant to be change in version 2.0.0, but got missed. Norm aggregation was used in all the benchmarking of version 1 as it performs better than mean aggregation when predicting properties that are extensive in the number of atoms.

More documentation for the CLI hpopt and fingerprint commands have been added and can be viewed here and here.

The individual predictions of an ensemble of models are now automatically averaged and the individual predictions are saved in a separate file. #919

What's Changed

Full Changelog: v2.0.2...v2.0.3

Contributors

jonwzheng, hwpang, and 2 other contributors
Loading

v2.0.2 Adding Document Modules and hpopt Enhancement

14 Jun 22:51
@shihchengli shihchengli
v2.0.2
f64881f
Compare
Choose a tag to compare
v2.0.2 Adding Document Modules and hpopt Enhancement

In this release, we have included numerous notebooks to document modules. Chemprop may be used in python scripts, allowing for greater flexibility and control than the CLI. We recommend first looking through some of the worked examples to get an overview of the workflow. Then further details about the creation, customization, and use of Chemprop modules can be found in the module tutorials.

New CLI Features

Improved --model-path CLI

Previously --model-path could take either a single model file or a directory containing model files. Now it can take any combination of checkpoint files (.ckpt), model files (.pt), and directory containing model files. Directories are recursively searched for model file (.pt). Chemprop will use all models given and found to make predictions (#731).

Improvements for hpopt CLI

Some flags related to Ray Tune (i.e., --raytune-temp-dir, --raytune-num-cpus, --raytune-num-gpus, and --raytune-max-concurrent-trials) have been added. You can use the CLI to initiate your Ray instance using these flags. (#918)

Bug fix

An incorrect max learning rate was used when writing the config file after hyperparameter optimization. This is now fixed (#913).

What's Changed

Full Changelog: v2.0.1...v2.0.2

Contributors

jonwzheng, kevingreenman, and 3 other contributors
Loading

v2.0.1 First Patch

06 Jun 17:14
@shihchengli shihchengli
v2.0.1
33df735
Compare
Choose a tag to compare
v2.0.1 First Patch

New CLI Features

Caching in CLI

MolGraphs are now created (by featurizing molecules) and cached at the beginning of training by default in the CLI. If you wish to disable caching, you can use the --no-cache flag, which will featurize molecules on the fly instead. (#903)

Change the default trial scheduler in HPO

We changed the default trial scheduler for HPO from AsyncHyperBand to FIFO, as it is the default in Ray and was used in version 1. You can switch the trial scheduler back to AsyncHyperBand by using --raytune-trial-scheduler AsyncHyperBand if needed. (#896)

Support Optuna in HPO

You can use optuna as a HPO search algorithm via --raytune-search-algorithm optuna. (#888)

CLI Bug Fixes

HPO-related bugs

In #873, we changed the search space for the initial and final learning rate ratios and max_lr to avoid very small (~10^-10) learning rates and also ensured that some hyperparameters are saved as integers instead of floating-point numbers (e.g., batch_size). In #881, we addressed the bug concerning the incompatibility of the saved config file with the training config. In #836, we shut down Ray processes after HPO completion to avoid zombie processes. For those encountering issues with Ray processes, we suggest you start Ray outside of the Python process.

DDP-related bugs

In #884, we resolved the issue where metrics were not synchronized across processes and disabled the distributed sampler during testing in DDP.

Backwards incompatibility note

In #883, we fixed the bug related to unused parameters in DDP. Models created via the CLI in v2.0.0 without additional atomic descriptors cannot be used via the CLI in v2.0.1. You will need to first remove message_passing.W_d.weight and message_passing.W_d.bias from the model file's state_dict to make it compatible with the current version.

What's Changed

New Contributors

Full Changelog: v2.0.0...v2.0.1

Contributors

JacksonBurns, kevingreenman, and 6 other contributors
Loading

v2.0.0 Stable Release

23 Apr 17:41
@kevingreenman kevingreenman
v2.0.0
73d8948
Compare
Choose a tag to compare
v2.0.0 Stable Release

This is the first stable release of Chemprop v2.0.0, with updates since the v2.0.0-rc.1 release candidate in early March.

The primary objectives of v2.0.0 are making Chemprop more usable from within Python scripts, more modular, easier to maintain and develop, more compute/memory efficient, and usable with PyTorch Lightning. Some features will not be migrated from v1 to v2 (e.g. web, sklearn). Some v1 features will be added in later versions of v2 (v2.1+) (e.g. uncertainty, interpret, atom- and bond-targets); see milestones here. The new version also has substantially faster featurization speeds and much higher unit test coverage, enables training on multiple GPUs, and works on Windows (in addition to Linux and Mac). Finally, the incorporation of a batch normalization layer is expected to result in smoother training and improved predictions. We encourage all Chemprop users to try using v2.0.0 to see how it can improve their workflows.

v2 documentation can be found here.

There are v2 tutorial notebooks in the examples/ directory.

A helpful transition guide from Chemprop v1 to v2 can be found here. This includes a side-by-side comparison of CLI argument options, a list of which arguments will be implemented in later versions of v2, and a list of changes to default hyperparameters.

Note that if you install from source, the primary branch of our repository has been renamed from master to main.

Due to development team bandwidth, Chemprop v1 will no longer be actively developed, so that we can focus our efforts on v2. Bug reports and questions about v1 are still welcome to benefit users who haven't yet made the switch to v2, but bug reports will not be fixed by the development team.

Please let us know of any bugs you find, questions you have, or enhancements you want in Chemprop v2 by opening an issue.

Loading

Final Patch for Version 1

12 Apr 17:47
@kevingreenman kevingreenman
v1.7.1
02c7602
Compare
Choose a tag to compare
Final Patch for Version 1

This is the final release of chemprop v1. All future development will be done on chemprop v2. The development team is still happy to answer questions about v1, but no new feature requests or PRs for v1 will be accepted. Users who identify bugs in v1 are still encouraged to open issues to report them - they will be tagged as v1-wontfix to signify that we won't be publishing fixes for them in official chemprop releases, but the bugs can still be open to community discussion.

We encourage all users to try migrating their workflows over to chemprop v2 (available now as a release candidate, stable version planned to be released within the next week) and let us know of any issues you encounter. All v1 releases will remain available on PyPI, and the v1 source code will remain available in this GitHub organization.

What's Changed

New Contributors

Full Changelog: v1.7.0...v1.7.1

Contributors

JacksonBurns, shihchengli, and willspag
Loading