search

2y7c3 / Super-Resolution-Neural-Operator

Super-Resolution Neural Operator, in CVPR 2023
MIT License
137 stars 12 forks source link

Cuda memory consumption #2

Open Mingyuan1997 opened 1 year ago

Mingyuan1997 commented 1 year ago

Do you have a value of expected cuda memory allocated of your code base with repsect to input image size. I tried with 3090 and it's always out of my cuda memory. Thanks!

2y7c3 commented 1 year ago

Please try to use the "batched_predict" function with a small bsize, e.g. bsize =500.

Mingyuan1997 commented 1 year ago

Thanks for your response. My issue is even worse. Even I run one demo input and the cuda gets out of memory.

Mingyuan1997 commented 1 year ago

And the cuda memory budget is still quite limited for batch_predict with a small bsize with rdn model. Especially on high-resolution images

AllenDun commented 6 months ago

I tried an input image 1920x1080, scale x2, rdn model always out of memory