Yi Zhu

Yes, you can. Instead of using photometric loss at [here](https://github.com/bryanyzhu/Hidden-Two-Stream/blob/master/models/multiframe_MotionNet/train.prototxt#L1333-L1347), you can load GT optical flow images and compute L1 loss.

是的，必须先build好才能使用。

This PR looks good. May I ask what is the difference between this PR and PR #1727 ? It seems this PR doesn't have the corresponding model definitions. Should I...

Hi, we use a copy from [here](https://academictorrents.com/details/184d11318372f70018cf9a72ef867e2fb9ce1d26). There are 240618 training and 19404 validation videos. There is a simple [readme](https://github.com/dmlc/gluon-cv/blob/master/scripts/action-recognition/ARXIV.md) for reproduction using GluonCV, and I'm still editing it. Please...

Hi @StevenJokess This setting file should be generated on your end, which contains the paths of the videos, duration of the videos and the label of the videos. Please read...

I have tested the script half a year ago, but not recently. Thank you for pointing out the missing url issue, I will fix it. In terms of generating `k400_train.txt`,...

Hi @blackmagicianZ , are you using the videos from the torrent website? If not, could you please open another issue and attach these videos? We can look into the issue...

Just curious, I saw two "clipp.hpp" files, is it possible to reuse it instead of putting it in two folders? Thank you.

Agree, I will try to add it.

@chinakook @zhreshold Right now, the training of semantic segmentation depends on sync batchnorm, which utilizes `dataparallel`, https://github.com/dmlc/gluon-cv/blob/master/scripts/segmentation/train.py#L163-L164 And our implementation of `dataparallel` depends on thread, https://github.com/dmlc/gluon-cv/blob/master/gluoncv/utils/parallel.py#L3 I searched online for...