Deepxde

Areas of improvement

- `FNN` supports arguments `use_bias` and `kernel_constraint`, and layer normalization
- Change L-BFGS option `gtol` from 1e-5 to 1e-8
- Improve `saveplot`

New APIs

- Add a new Data `Constraint`
- Add metric `mean_squared_error`

We stop the support of Python 2.7 from this release.

Areas of improvement

- Vectorize geometry related methods
- `PDE` supports resample PDE residual points via callback `PDEResidualResampler`
- Implement `Polygon.boundary_normal()`

New APIs

- Add multifidelity DeepONet

API changes

- Geometry-related methods (`inside`, `on_boundary`, `on_initial`, `boundary_normal`, `periodic_point`) now use the 2d input of shape N by d, not a 1d vector.

Areas of improvement

- `PDE` supports excluded points
- Add `PointSetBC` for training using exact solution.
- Add `Hypersphere.boundary_normal()`
- Add `Triangle.inside()`

API changes

- `PDE` doesn't support numpy.array `X` as the third parameter. To define pde with extra coefficients, use `PDE(..., auxiliary_var_function=...)`, see an example at Lorenz_inverse_forced_Colab.ipynb.

Areas of improvement

- `Model.predict()` supports PDE with two or three arguments
- `PDE` supports different loss functions for different errors
- `Hypercube.uniform_points()` uses the same dx in each dimension

New APIs

- Add metric: `nanl2_relative_error`, `zero`
- Add callback: `Timer`

New module

- Add new module `dde.grad` to compute gradients via `dde.grad.jacobian()` and `dde.grad.hessian()`

Areas of improvement

- Bug fix: `GeometryXTime.periodic_point()`

Areas of improvement

- `Model.save()` supports pickle
- `PDE`
- supports `pde` as `None`
- supports `bcs` as the empty list `[]`
- accepts `train_distribution` = 'uniform' or 'pseudo' or 'sobol'
- `OpNN` supports user-defined branch net
- `OpDataSet` supports mini-batch training

API changes

- Rename `Model._print_model()` to `Model.print_model()`

New APIs

- Add `Model.state_dict()`
- Add `PDE.resample_train_points()`