• TorchScript compatibility has been added to export TorchANI models through torch.jit. Users can now use C++ API for deployments. (#303, #305, #306, #307, #308, #326, #327).
• Some APIs are changed due to the compatibility issue with TorchScript:
  • AEVComputer input is changed, cell and pbc are now keyword arguments. (#303)
  • Ensemble is now hardcoded to have a size of 8. (#307)
  • torchani.nn.Sequential is added to include type annotations for JIT. (#307)
• An example of how the models can be exported using PyTorch JIT has been provided (#328).
• All the unit tests and checks have been moved to GitHub Actions. (#309, #310, #313, #314, #317, #318, #319, #322, #323, #324)
• Added a script for profiling the training on NVIDIA GPUs using Nsight System (#325)
• Bug fixed in the dimensions of self_energies for a dataset containing only one element (#302)

• The package name of PyTorch has changed from torch-nightly to torch, we update it respectively (#295, #294)
• Add new experimental data API that has much less memory usage and much better performance (#284, #299)
• Example files now reproduce NeuroChem results after taking out the outlier energy conformers (#287)
• Accelerate angular AEV computation and reduce memory cost (#290)
• Remove all large files and stop using git-lfs (#289)

• Support PyTorch new torch.bool datatype. (#278)
  Warning: Boolean tensor is a breaking change introduced in PyTorch, TorchANI has to be updated to run on the latest PyTorch, otherwise it will produce wrong results.
• Improve TorchANI to be more close to NeuroChem. Update examples respectively. Also, provide helper functions. (#283, #282, #280, #279, #275, #261, #259, #255, #245, #263)
• Add example for energy/force training. (#249, #240, #238, #233)
• Bug fixed in AdamW implementation. (#261)
• Dataset related API has changed, see the new API doc for detail. Its implementations are also improved. (#230, #231, #232, #236, #237, #250, #257, #272)
• Builtin models API has changed, and implementation improved. (#252, #266)
• Fix a bug of not having torchani.data when ignite is not installed. (#256)

• Experimental support for weight decay (#223)
• Add example of training ANI models by manually writing training loops that has the same configuration as NeuroChem (#224)
• Use torch.utils.tensorboard to replace tensorboardX (#224)
• h5py and pytorch-ignite-nightly are removed from dependencies, so that inference-only application do not need to install them (#224)
• MAE vs MaxAE are clarified in NeuroChem trainer (#225)
• Support mode computation in vibrational_analysis(#228)

• Fix bug of analytical stress of PBC (#221)
• Add support for analytical hessian and vibrational analysis (#222)

• Fix a bug on PBC (#214)
• Add support for analytical stress (#218)

• Fix ase interface on GPU (#212)

• Further performance improvements (#202, #204, #205, #206,#207)
• Use a stable API torch.unique_consecutive to replace the temporary unique in most recent nightly PyTorch (#209)

• Rewrite AEVComputer to improve performance and reduce memory usage (#197)

Note that this release use a very new PyTorch operator torch.repeat_interleave and a very new temporary API torch._unique_dim2_temporary_will_remove_soon. Please make sure you update your nightly PyTorch before updating TorchANI. New version of TorchANI will be released when torch._unique_dim2_temporary_will_remove_soon is removed to make sure TorchANI always work with latest nightly PyTorch.

• Inference on Python 2.7 (#171 , #172)