一、整体架构 1.1 累加参数设置 nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp[