求助如何对RIFE_HDv2.py网络进行v2版本的训练

[xsp1]

您好，我注意到有人提出v2.3版本权重模型的视频插帧效果似乎好于v6、v4系列的版本，经过我在自己视频上插帧验证实验也是如此，所以想自己重新训练v2版本的网络试一下，在这之前我已经跑通了RIFE.py版本的训练。但我发现RIFE_HDv2.py和RIFE.py之间存在很多差异，经过我的修改也不能直接在目前这个发行版本的train.py上直接训练，应该是代码结构如update函数返回值不同，同时v2版本应该提供flow_gt文件，但目前的RIFE.py中不需要。请问能否提供适合RIFE_HDv2.py版本训练的代码？十分感谢！

[hzwer]

正好最近有邮件也问这个问题，我在这里贴一下；我个人认为 v4 以后的版本有更大的潜力，原因是它不仅速度快了若干倍，还支持任意时刻插帧（我认为是这个功能牺牲了部分性能）
flow_gt 的相关使用可以参考 IFRNet 提供的代码

关于复现的一些信息：
https://drive.google.com/file/d/1Az-X6xMcvi2sUCGQJ9BAhEBSJtAaIKN4/view?usp=sharing

This is an unorganized training code of the v4.x version. The main difference from the open source version is

1. The augmentation in the dataset
2. The scale is randomly transformed during training

In my impression, there are only two differences between the v2.x version and this one. One is that v2 uses a larger model, and the other is that v2 is only trained at t=0.5. All the details you asked about I believe can be confirmed from the code. In our tests, the v4.x version should achieve 90% of the performance of the v2 version, and use a smaller model and support arbitrary timestep (recently we found that this is an operation that sacrifices t=0.5 performance, but Very useful).

I think the most important point is to pay attention to the LPIPS validation index of the model, which can be tested in vimeo90K and some actual collected resolution scenes above 1080p.

[xsp1]

非常感谢您的回复！我会继续研究一下，谢谢您的帮助！