DeepXDE v1.9.0

DeepXDE supports data-parallel training of PINN now.🎉🎉🎉

Areas of improvement

• Backend tensorflow.compat.v1: Support data-parallel training of PINN via Horovod in both weak and strong scaling modes
• Backend paddle: Some updates to support the new Paddle version
• Bug fix for Float16

New APIs

• Add dde.config.set_parallel_scaling to set parallel mode
• Add dde.config.set_hvd_opt_options for hvd.DistributedOptimzer

DeepXDE v1.8.4

We stop the support of Python 3.7 from this release.

Areas of improvement

• If no backend is selected, automatically find an available backend.

DeepXDE v1.8.3

Areas of improvement

• Many improvements for backends:
  • Display more information
  • Verify if the backend is available/importable
  • Add interactive installation of Paddle
• Backend PyTorch: Model.restore supports restoring model to a specified device
• Use setuptools_scm for VCS versioning

New APIs

• Support new geometry dde.geometry.Ellipse

DeepXDE v1.8.2

This release is used to test the automated release workflow.

DeepXDE v1.8.1

Areas of improvement

• Support float16
• Backend TensorFlow supports Model.state_dict
• Migrate build config from setup.py to pyproject.toml
• Migrate CI from Travis to GitHub Actions
• Many examples and documentation improvements

DeepXDE v1.8.0

A lot of implementations for backend paddle! Feel free to use backend paddle.🎉🎉🎉

Areas of improvement

• dde.nn.FNN supports defining activation functions for each layer
• dde.geometry.PointCloud supports boundary points and normals
• Bug fix

New APIs

• Backend PyTorch: Support dde.nn.DeepONet

DeepXDE v1.7.2

Areas of improvement

• Add dde.icbc.PointSetOperatorBC
• dde.callbacks.OperatorPredictor can be used during training
• Backend PyTorch: dde.icbc.PointSetBC supports multi-component outputs
• Bug fix

DeepXDE v1.7.1

Areas of improvement

• dde.data.TripleCartesianProd and dde.data.QuadrupleCartesianProd support mini-batch for both branch and trunk nets
• Backend TensorFlow 1.x: L-BFGS dumps trainable variables and test loss

DeepXDE v1.7.0

Areas of improvement

• dde.icbc.PointSetBC supports mini-batch
• Bug fix

New APIs

• Backend paddle: Support dde.nn.DeepONet and dde.nn.DeepONetCartesianProd

API changes

• Change dde.callback.PDEResidualResampler to dde.callback.PDEPointResampler

DeepXDE v1.6.2

Areas of improvement

• Set Hammersley as the default point sampling for PINN
• Improve point sampling of dde.geometry.GeometryXTime.random_points
• dde.callback.ModelCheckpoint supports monitoring test loss
• PyTorch backend: dde.nn.PODMIONet and dde.nn.MIONetCartesianProd support multiple merge operations