shape = (BATCH_SIZE, N_CLASSES, IMG_SIZE, IMG_SIZE)
buffer_pool = [torch.empty(shape, pin_memory=True).share_memory_() 
               for _ in range(POSTPROC_WORKERS)]