David Pitt

This PR adds batched mode to the FNOGNO and IntegralTransform for cases when the mesh stays constant across a batch. Tests for the FNOGNO have been updated to add this...

Deepcopy base model

2

The models in `neuralop.models` subclass `neuralop.models.BaseModel`, which adds lots of useful functionality but does not implement the `__deepcopy__` method necessary to copy a torch module's architecture and state. This PR...

@juliusberner suggests we add 1d positional embedding. The only pos-embedding we currently support by default is 2D.

Connected to #280: we previously discussed adding `torchvision`-style dataset files that include the ability to download data for sample problems, e.g. Navier-Stokes or Darcy flow. This PR adds this functionality...

Would be nice to set up NCCL backend and have DDP run in pure python

Since the `SpectralConv` refactor, we now depend on newer source from `tltorch` than is available in the latest tagged release on PyPI. Instead of installing `tensorly-torch` from pip, to pass...

This has been mentioned a few times. Would be good to bring this in soon

Adds `MSELoss`, refactors finite difference methods out of `data_losses`, makes `MSELoss` the default option for `Trainer.train_loss`

#284 mentions that logging no longer happens by default because we moved logging to Callbacks. I think a default logging callback would be useful

Quick fix explaining how mode indexing works.