Skip to content

v1.2.0
Compare
Choose a tag to compare
@HydrogenSulfate HydrogenSulfate released this 27 Nov 11:08
· 1 commit to release/1.2 since this release
v1.2.0
842ac9f

PaddleScience 1.2.0 Release Notes

PaddleScience 1.2.0 is now available with multiple feature updates, documentation, code optimizations and bug fixes. This update brings numerous enhancements that improve overall performance and usability. Some of the key updates include:

Examples

  • Add ShockWave, Bubble, DeepCFD, DeepONet, IDE, FPDE, HeatPINN, Nowcasting(inference only), TopOpt, Biharmonic2D, PhyLSTM, CFD-GCN, Deep-Spatial-Temporal, DU_CNN, PIRBN and XPINNS, in total 15 examples and their corresponding network structures.

Module

  • Add network structure including AMGNet, DeepONet, DeepPhyLSTM, UNetEx, refactor GAN for tidy and better readability code.
  • Add Siren activation function.
  • Add create_graph and retain_graph options for jacobian and hessian for saving memory in which creating reverse graph is unnecessary.
  • Add compute_sdf_derivatives option to compute discrete derivative of SDF when it is used in given equations.
  • Add MeshAirfoilDataset, MeshCylinderDataset, NPZDataset, IterableNPZDataset in dataset. Graph data is preliminarily supported with PGLBox and adapted default_collate_fn.
  • Change all equations from python closure style to sympy style with symbolic computation supported in ppsci/utils/symbolic.py.
  • Add ppsci.experimental and ppsci.probability module for supporting API call for mathematical usage.
  • Add ppsci.loss.mtl module for Multi-Task Learning algorithms, PCGrad and AGDA is included preliminarily.
  • Split original L2Rel into L2Rel and MeanL2Rel, the former is computed along entire dataset(more commonly usage), while the latter is computed element-wisely for certain usage.
  • Add run_check_mesh for checking whether ppsci.geometry.Mesh module works properly.

Improvements

  • Support running code with hydra configuration for better experimental management. For details, see user guide - hydra and hydra intro.
  • Solver.predict can return np.ndarray data instead of Tensor data by setting return_numpy=True.
  • Numerous type-hints and annotations were added for better development&reading experience.

Document and bug fix

  • Completely new home page with full example list and all important information: https://paddlescience-docs.readthedocs.io/zh/release-1.2/.
  • Training/evaluation script are provided with pretrained model at the beginning of document of each example for quick experience and reproduction.
  • Correct almost all typos in code, such as visulzie --> visualize.

We recommend upgrading paddlescience to version 1.2.0 or the develop branch.

New Contributors

Full Changelog: v1.1.0...v1.2.0

Contributors

jiamingkong, cxxly, and 20 other contributors
Assets 3
Loading