I'm trying to build model detection and want to use YOLO-V8 as pre-train, The total images are around 750 images. The reference I replicated was from here. I used CPU Machine n1-standard-4 with 4 NVIDIA TESLA T4. The result of running time still remaining the same with n1-standard-4 and 2 NVIDIA TESLA T4 by around 6 hours. I noticed the difference in GPU utilization only ~25% for 4 NVIDIA TESLA T4 and ~90% for 2 NVIDIA TESLA T4. I've read the documentation here, from my understanding If I use n1-standard-4 which provide 4vCPU and 15GB memory, the vCPU and memory og the GPU should exceed the CPU Machine. My aim is to speed up training process of ~6 hours to only less than 3 hours. How do I supposed to do? Continue reading...
