I have a machine with 1500 Gb of ram and 112 cores but no GPU. I would like to know how to calculate suitable number of num_workers in DataLoader and batch size. I have to make the inference with Inception-v3(pytorch) on 80,000,000 images. I tried with a batch size of 30k and num_workers = 4 with fine results. if I try with num_workers = 50 I don't get a good result Continue reading...
