LUYue

Thank you for your reply! My script: [extract_opticalflow_2.txt](https://github.com/open-mmlab/denseflow/files/5897245/extract_opticalflow_2.txt) My video folder (`ls 01_023`): [src_video_folder.txt](https://github.com/open-mmlab/denseflow/files/5897244/src_video_folder.txt) Output: When I set `-s=0`: [denseflow_s0.txt](https://github.com/open-mmlab/denseflow/files/5897237/denseflow_s0.txt) (The frames are only copied.) `-s=1`: [denseflow_s1.txt](https://github.com/open-mmlab/denseflow/files/5897240/denseflow_s1.txt)

The output log: ``` (torch17) luyue@gpu-Super-Server:~/workspace/Projects/MMDetSeg_2/OpticalFlow$ python extract_opticalflow_2.py "/home/luyue/workspace/Projects/MMDetSeg_2/data/SHTCampus/Train/01_023", frames ≈ 1298 push frames gray, video_flow_idx 0, batch_size 512 push frames gray, video_flow_idx 511, batch_size 512 push frames gray, video_flow_idx...

I still got 600 pairs of flow images. `ls` video_dir: [src_600_01023.txt](https://github.com/open-mmlab/denseflow/files/5897927/src_600_01023.txt) `ls` flow_dir: [flow_600_01023.txt](https://github.com/open-mmlab/denseflow/files/5897928/flow_600_01023.txt)

I find that the frames are not clear so I extract the frames of the raw video again. The old frames are extracted by `ffmpeg` command. The new frames are...

I'm sorry for that I misunderstood what you said about `-s=0`. Thanks again.

I met the same problem with you, do you solve the problem yet?

I have forgotten how I solved the problem. Maybe it's due to different versions of pytorch.

@yangyangtiaoguo 兄弟够拼的！我提这个issue的时候，也按照论文中的邮箱，给所有作者发了一封邮件，但是到现在也没人回复我。  我在ShanghaiTech和Ped2上复现的结果低一点，但是也不如Avenue上低这么多。这篇论文既然在Avenue上达到了SOTA表现，最应该放出的模型应该是Avenue上的才对。但是作者只给了Ped2和ShanghaiTech上的模型，确实不应该。

> 请问一下楼主你知道代码中哪一部分是DPU嘛？我阅读代码感觉就是直接编码器到解码器没有看到DPU部分的存在呀？ 应该就是那个meta_prototype.py中的吧

> 请问你们是怎么训练的呀？是直接就作者的代码，训练1000epoch么？我看论文写的是epoch=60，我跑完之后，ped2只有95.1。低很多。avenue也是。 我是训练1000 epoch，速度也不算很慢