Darknet: Training in multy-gpu

[a13483685]

my device have 4 2080ti Gpus ,but when I set the subdivisions=32 in yolov4.cfg and training .shown CUDA Error:out of memory:Resource temoprarily unavailiable in terminal. And then ,I changed another device(1 GPU 1080ti),also have that question. How can I solution the problem? thank you very much!

[WongKinYiu]

for width=height=512, you can train with batch=64 and subdivisions=32 on 1080ti.
for width=height=416, you can train with batch=64 and subdivisions=16 on 1080ti.

[AlexeyAB]

@WongKinYiu

for width=height=512, you can train with batch=64 and subdivisions=32 on 1080ti.

I think preferably to use this way (since we use random=1), isn't it?

[WongKinYiu]

@AlexeyAB

yes, and my multi-gpu training on 4 2080ti will finish in this week, I will cheak if the hyper-parameter should be modify or not.