baratilab / diffusion-based-fluid-super-resolution Goto Github PK

In train_ddpm/runners/diffusion_tub.py the loss is computed as

loss = loss_registry[config.model.type](model, x, t, e, b, x_offset.item(), x_scale.item()).

This calls def conditional_noise_estimation_loss(model, x0: torch.Tensor, t: torch.LongTensor, e: torch.Tensor, b: torch.Tensor, x_scale, x_offset, keepdim=False, p=0.1) in train_ddpm/functions/losses.py and thus swaps x_scale with x_offset. It would be much appreciated if the authors could double check if they used a correct version for the results in the paper or these results suffer from this mismatch. Thank you in advance.

Hello, may I inquire whether you utilize a 256256 to 256256 mapping during the training process of the diffusion model? Is it only during the testing phase that you directly employ the model, without retraining, for the reconstruction of 3232, 6464, and sparse data?

Hi! I started directly from step 2 and ran python main.py --config kmflow_re1000_rs256.yml --seed 1234 --sample_step 1 --t 240 --r 30 in the main directory of this repo, but got the following message:

Traceback (most recent call last):
File "main.py", line 112, in
sys.exit(main())
File "main.py", line 94, in main
args, config, logger, log_dir = parse_args_and_config()
File "main.py", line 40, in parse_args_and_config
dir_name = 'recons_{}_t{}_r{}_lam{}'.format(config.data.data_kw,
AttributeError: 'Namespace' object has no attribute 'data_kw'

Can you please provide the scripts used in your experiments in order for me to reproduce the results?

Hi,
I tried to use your code, but it seems that Py torch 1.7 is not the rights version since fft2 didn't exist in the mentioned version and you used it in your code.

Thanks