zhaoyanglijoey / yolov3

Pytorch implementation of real time object detection algorithm YOLOv3
MIT License
21 stars 14 forks source link

Cuda Out of Memory #2

Closed SurionAndrew closed 4 years ago

SurionAndrew commented 5 years ago

Thanks for the great Darknet implementation, It works really well

I am struggling to get batch classification to work on my gtx 1050ti (4gb). I can only use batches = 1,2,3. if I increase this number I get the following "RuntimeError: CUDA out of memory. Tried to allocate 20.00MB (GPU 0; 3.94 Gib total capacity; 2.74 Gib already allocated; 35.31mb free; 159.00Mib cached).

But using subdivisions doesn't help. is there perhaps not place where memory needs to freed or am i just completed limited by hardware..

Also how could we speed up inference ? eg setting it to FP16 rather than 32 Regards Andrew