Very slow when using pretrained model weights.

[Blue0rigin]

Hi, I am trying to use your pre-trained model weights to infer the scanpath for my own text (around 150 long sentences sampled from my big corpus). I directly use "CUDA_VISIBLE_DEVICES=0 python -m scandl2_pkg.test" by replacing the sample sentences with mine. But the running speed is extremly slow. Any advice to speed up the inference using your pre-trained models so that I can analyze a huge amount of sentences in my corpus?

Thanks a lot!

[Blue0rigin]

btw. I am using Nvidia A100 40G for the inference.

[Blue0rigin]

I also noticed that you have implemented ddim in your code, but it doesn't work when I set use_ddim=True. I am not sure if ddim in DiffuSeq is adapted for your ScanDL. Any advice on how to use ddim in Scandl? Thanks a lot!

[dongsenzhang]

I also encountered the problem of slow speed. Has there been any solution to it?