Dev list before release

[pzhanggit]

• PRs @aprokop

  • Update models to allow conditioning #1
  • Add Flow3D data loader #9

• GINO @aprokop

• SOLPS-autoregressive training @TsChala

  • SOLPS data, input actuator and autoregressive training #15

• SOLPS-additional input arguments @TsChala

  • SOLPS data, input actuator and autoregressive training #15

• SOLPS-update dataset file @TsChala

  • SOLPS data, input actuator and autoregressive training #15

• SR-Metrics like SSIM and other util functions @TsChala

  • Spatial gradient loss #20
  • Add SSIM metric #22

• SR-data augmentation (? not planned for now) @TsChala

  • @TsChala: Let's put this on hold for now. The current implementation doesn't work with sequence parallelism. Also we have to be careful when doing augmentation for vector fields (e.g., velocity) since the directions need to be transformed too.

• SR-update dataset file @TsChala

  • Update SR data reading #19

• Preserving sampler for each epoch (multi-epoch data loader) @TsChala

  • Add multi epoch data loader #12

• Adaptive tokenization @pzhanggit

  • mv adapt implementation from gitlab; upgrade to 3D #13

• Cleanup functions in pretraining branch @pzhanggit

• mixed 2D and 3D in turbt

  • Moving capabilities handling heterogeneous datasets simultaneously from pretraining branches #26

• flexible batch sampler

  • Moving capabilities handling heterogeneous datasets simultaneously from pretraining branches #26

• rm RMSInstanceNorm (already done in pretraining) @pzhanggit

  • Remove RMSinstanceNorm #14

• Clean up model forward functions @pzhanggit

  • Clean up model forward functions and fix bugs #17

• Add module to handle unstructured mesh @pzhanggit

  • add graph modules to support unstructured data #23

• TODO

  • Add a fine-tuning example @TsChala
  • CI tests (?) @aprokop
  • Reorganize datasets in project shared @pzhanggit
  • Example datasets (coming soon!)
  • Unify the DDP settings to run on different platforms @aprokop and @TsChala
  • Dictionary arguments & some function cleaning
  • Formalize package versions (FlashAttention; GINO) @aprokop @TsChala