Yueyi wang
Yueyi wang
> 您好, > > 1. 1-shot实验我们只使用了1张GPU,5-shot实验使用了2/4张GPU,但GPU数量对性能的影响很小。请避免使用resume(中断训练,重新加载),这会在一定程度上产生影响。另外,多卡训练只需根据GPU数量对应修改batch_size,学习率保持不变。 > 2. 不同的实验环境可能会引起结果差异。即使固定了随机种子,也较难保证获得完全相同的结果,例如"双线性插值"过程存在的随机性不受种子影响,这对于多轮迭代的训练过程影响更大。 > 3. 这里我的建议是尝试1次以上的复现,并更换随机种子进行训练/测试。如仍不能取得论文中的指标,可以在今后的工作中呈现您复现的指标,这并不会产生对论文评估的潜在负面影响,可参考ASGNet (CVPR'21) 对PFENet的复现。 请问两张卡的batch_size在lr保持不变的情况下,您是设置的多少?
> 1. Yes, you could and you will train 200epoch started at 100e. > 2. It's warning due to API change in pytorch, don't worry. Thanks for your reply. I...
> Any exception trace log @66Kevin ? Sorry, I don't get any other exception trace log. So I am very confused that why it was terminated.
> Thank for the nice suggestion! We have fixed this. Thanks for your reply! There is also a question about Series Informed Activation Function. Why the number of stacked activation...