hzwer
OK, I think I have fixed it.
I don't know much about hardware, can someone help? It seems that you can send a topic to https://www.svp-team.com/forum/viewforum.php?id=7
The methods are similar. In Flownet.py, the first value returned by model inference is the optical flow list. Just take flow_list[-1][:, :2]. 
Set it to 1.0. Then flow[:, :2] will be F1->0 (Theoretically speaking)
flow[:, :2] is Ft->0, flow[:, 2:4] is Ft->1 Set timestep to 1.0. Then flow[:, :2] will be F1->0, flow[:, 2:4] will be F1->1 Set timestep to 0. Then flow[:, :2]...
I have no relevant skills ... 完全不懂
No, I only have this. 
你好，我有时候换数据集训练会遇到这种情况 通常我会调小蒸馏损失的权重系数以及增加优化器的 weight decay
那 torch.load('{}/flownet.pkl'.format(path)).keys() 会得到什么？
我不理解，看起来模型参数和 pkl 中是一致的 #PyTorch模型加载错误解决方法# 来自跃问分享 https://yuewen.cn/share/179863116760768512?utm_source=share&utm_content=web_linkcopy&version=2 很奇怪的问题