Skip to content

Latest commit

 

History

History
31 lines (31 loc) · 2.11 KB

TODO.md

File metadata and controls

31 lines (31 loc) · 2.11 KB
  • add resulting images to log ?
  • CHECK Convolution classes & add flat_meshgrid
  • tools to visualize dataset
  • decorator for forward method that detect numpy input, disable grad & reshape? See also reshape in ReactionProblem
  • diff, total_variation_norm, norm in another module?
  • consistant way of specifying domain bounds, kernel size, ... in the command-line (e.g. [0,1]x[0,1], 15x15, maybe event [0,1]³ or [0,1]^3 ?!)
  • more shapes and operations, fix elongate, express everything from basic shape + op (sphere <=> rounded dot, rectangle <=> elongated dot, ...)
  • batch_size, batch_shuffle, lr, train_N, val_N and dataloader method in Problem?
  • dataset generation and loss calculation on multiple steps in a dedicated class (eg EvolutionProblem)?
  • 64 bits precision support (see Lightning-AI/pytorch-lightning#2497)
  • adding dtype in every possible class (eg Domain)
  • optimization, speed and memory footprint, eg: dataset generated on the fly instead of fully stored, using broadcasting in domain.X and K, lazy expression using Keops, TorchScript, ...
  • load checkpoint from GPU to CPU (use map_location or always save to cpu?)
  • investigate high cpu load (eg on cluster24-math)
  • method for saving model (without trainer)
  • what about the training and validation dataset when resuming ? (with seed fixed)
  • add more tests (model learning, checkpointing, loading, ...)
  • AllenCahnSplitting also directly from model (without checkpoint)? (how to save hyperparameters?)
  • faster fftconv with padding != 'circular' using scipy.fftpack.next_fast_len
  • reload hyperparameters when reloading from checkpoint ?
  • dedicated module
  • remove utils functions from trainer
  • epoch=#-v#.ckpt ?! (collide with auto saving initial state in trainer.Trainer)
  • common animation script or functions
  • animation features for allen_cahn_problem
  • a hparams.yaml is written in default_root_dir ?!
  • --config option for nnpf without action specified
  • TensorBoardScalar can merge multiple files
  • fix Tracing issues
  • train action populates command-line arguments from class constructor
  • check 64 bits, also when loading