High GPU usage

[cnavarrete]

I'm running the docker and I can barely run 16 epochs on my 12GB NVIDIA TITAN X (Pascal) before it crashes with your classic cuda out of memory error.

I ran the training with the following command.

python3 train.py --weights yolov4.weights --data /datasets/anabalon_osses_final_vis/data.yaml --epochs 350 --img-size 640 --batch-size 2 --cfg cfg/yolov4.cfg  --device 1

The error

RuntimeError: CUDA out of memory. Tried to allocate 4.46 GiB (GPU 0; 11.91 GiB total capacity; 943.14 MiB already allocated; 3.48 GiB free; 7.78 GiB reserved in total by PyTorch)

I was seriously taken aback by the high gpu consumption. Anyone knows the cause? I haven't had this problem with yolov3 (ultralycs repo), v5 or v7.