Tone

I suggest that you can set fixed input tensors and compare outputs from different implementations to check it.

Note that our 8x frame interpolation model IFRNet_GoPro.pth is trained from scratch on a relatively small GoPro dataset, which may generate undesirable results. It is better to train 8x frame...

May be you should replace from models.FastFlowNet import FastFlowNet with from models.FastFlowNet_v2 import FastFlowNet in benchmark.py. Also, please refer to the [README.md](https://github.com/ltkong218/FastFlowNet/blob/main/README.md) file to configure the correct environment.

Thanks for your interest. As the figure shows, the number of orange square is 53, which are the considered search grids when building CDDC.

Do not run `python setup.py build`, and follow [Higher CUDA and PyTorch versions Support](https://github.com/ltkong218/FastFlowNet?tab=readme-ov-file#higher-cuda-and-pytorch-versions-support).

Thanks for your interest! For training on a machine equipped with single GPU, please try to run the following command: $ python -m torch.distributed.launch --nproc_per_node=1 train_vimeo90k.py --world_size 1 --model_name 'IFRNet'...

What experiment environment are you using? (PyTorch, CUDA version and GPU card)

This problem has been solved as https://github.com/ltkong218/IFRNet/pull/31.

Maybe you should install requirements for [pytorch-liteflownet](https://github.com/ltkong218/IFRNet/tree/main/liteflownet) correctly.

You can refer to [ncnn Implementation of IFRNet](https://github.com/ltkong218/IFRNet?tab=readme-ov-file#ncnn-implementation-of-ifrnet). It may be helpful.