GPU memory for zero-shot inference

[L-J-Yuan]

It's a great work !
When I tried to run demo_web.py for zero-shot inference, I encountered the error of CUDA out of memory on an 24G GPU card.
I wonder if the inference was also done on A100 for you ?

[splionar]

Hi, we use 40GB A100 GPU for the inference. You can set a smaller --batch_query here:

numcc/demo_web.py

Line 270 in 999f6b0

parser.add_argument('--batch_query', default=48000, type=int, help='batch query for repulsive UDF')

. A value >20k should be good.