Will YOLOv4 be supported?

[chouxianyu]

https://arxiv.org/pdf/2004.10934.pdf https://github.com/AlexeyAB/darknet YOLOv4 has appeared recently!

[loveunk]

1027 They are still working on it.

[glenn-jocher]

@chouxianyu thanks for the links! Yes, we actually have a completely separate repository under development with simpler training for everyone's custom datasets which produces markedly better results than this one. This should be publicly released in early May, and then we will direct all visitors here to the new repo!

[chouxianyu]

OK! Looking forward to the new repository and thanks for your outstanding work!

[Alex866850]

Hi, I'm a big fan of yours.I have been following your project since last year.No doubt, ultralytics- yolov3 it is extremely friendly and effective for those of us who are new to scientific research.Today, I heard that yolov4 came out. If you can integrate some of the functions or mechanisms of v4, I think it is simply too strong and powful. It must be another big storm.

[glenn-jocher]

@cxr12345 thanks for the comment! Yes, we've learned a whole lot in the last year or so about making training robust and friendly to everyone's custom datasets, and we have folded all these lessons learned into a new repo recently that is much simpler to use and more user friendly, and also produces better models! It's private at the moment, but should be public in early May.

[glenn-jocher]

Our main goal is to trend towards an automl-type solution, where the training adapts to custom datasets with a minimal amount of input from the user. This includes things like defining your model and dataset more simply with *.yaml files, automatically adjusting anchors based on dataset labels, hyperparameter adjustment based on the custom dataset etc. And training to the highest mAPs possible more automatically, because we know its very frustrating trying to find just the right settings etc.

[tyb197]

@glenn-jocher That's wonderful!!! I just began to learn how to train yolov3 network and plan to make some improvements ｏｎ my own task. Thank you for your previous excellent works which gave me a high starting point so that I can learn quickly. However, yolov4 has released today and it was great. Looking forward to this tutorial repo!

[jveitchmichaelis]

@glenn-jocher this sounds great, is this repository going to be deprecated? I've got a few usability improvements I was going to PR here (e.g. better plots during training, user-specified output directories, training on arbitrary channel data, etc), but it sounds like it may be worth waiting for the new release. Thanks for your work so far, this has been a really useful base to work on top of (and sometimes hard to keep up with merging commits!)

[glenn-jocher]

@jveitchmichaelis yes we are aiming to deprecate this repo eventually and move on to a new repo, though a lot of the code will graduate on to the new repo as well. The main parts that will change are the model definition and architecture methods (i.e. models.py, /cfg, /data). Most of the rest will be retained.

n-channel functionality is not currently supported because the user base is too small, but single-channel image training is supported (these are automatically imported as 3-channel images by opencv).

[liguoyu666]

I am very much looking forward to this tutorial repo

[QQ2737499951]

关于pytorch-yolov4，可加群857449786 注明（pytorch yolov4）共同学习，谢谢

[liguoyu666]

关于pytorch-yolov4，可加群857449786 注明（pytorch yolov4）共同学习，谢谢

付费群？打扰了

[Hwijune]

Please add pruning.

[wang-xinyu]

I have tried yolov4 inference in this repo, it works!

Download yolov4.weights from AlexeyAB/darknet. And run detect.py

python detect.py --cfg cfg/yolov4.cfg --weights weights/yolov4.weights --device 0 --img-size 608

It works properly.

And I have implemented yolov4 in tensorrt based on this repo.

[coakfulo]

付费群门槛也是有点高哦

发自我的iPhone

在 2020年4月28日，12:04，大禹 notifications@github.com 写道：

 关于pytorch-yolov4，可加群857449786 注明（pytorch yolov4）共同学习，谢谢

付费群？打扰了

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[glenn-jocher]

@wang-xinyu oh awesome!!

[glenn-jocher]

@wang-xinyu I updated readme for yolov4 tensorrt b0b52eec53e43548430f823e4d60032fab163228

[wang-xinyu]

@glenn-jocher Thanks! :)

[ZihaoCheng001]

@wang-xinyu Nice, the yolov4 from AlexeyAB/darknet really works well!

[Alex866850]

关于pytorch-yolov4，可加群857449786注明（pytorch yolov4）共同学习，谢谢 There is a charge for this group. Why? Did you make yolov5?

[Alex866850]

关于pytorch-yolov4，可加群857449786 注明（pytorch yolov4）共同学习，谢谢

这个群需要收费？？为啥呢，你们是做出了yolov5了吗？

[QQ2737499951]

关于pytorch-yolov4，可加群857449786 注明（pytorch yolov4）共同学习，谢谢

这个群需要收费？？为啥呢，你们是做出了yolov5了吗？

pytorch-yolov4复现了，可以win10下训练自己数据集，效果不错。仅此而已。

[liguoyu666]

关于pytorch-yolov4，可加群857449786 注明（pytorch yolov4）共同学习，谢谢

这个群需要收费？？为啥呢，你们是做出了yolov5了吗？

pytorch-yolov4复现了，可以win10下训练自己数据集，效果不错。仅此而已。

再过一个月，各位大佬的开源版本就出来了，你就继续藏着捏着吧

[dengfenglai321]

关于pytorch-yolov4，可加群857449786 注明（pytorch yolov4）共同学习，谢谢

打扰了

[XiaoLaoDi]

Our main goal is to trend towards an automl-type solution, where the training adapts to custom datasets with a minimal amount of input from the user. This includes things like defining your model and dataset more simply with *.yaml files, automatically adjusting anchors based on dataset labels, hyperparameter adjustment based on the custom dataset etc. And training to the highest mAPs possible more automatically, because we know its very frustrating trying to find just the right settings etc.

@glenn-jocher Great work ! Is this (new repo) available now ?

[github-actions[bot]]

This issue is stale because it has been open 30 days with no activity. Remove Stale label or comment or this will be closed in 5 days.

[glenn-jocher]

Latest models available at our new repo: https://github.com/ultralytics/yolov5

[antimo22]

Sorry but I'm a little bit confused. I want to train the YOLOv4-tiny model on my custom dataset, using the ultralytics implementation. But there are some problems, whether I use the v5 or v3 repo.

Regarding the v5 repo, there are some problems about the format files (YOLOv4 has cfg and weights file, while the repo wants yaml and pt file). Though I've found the yaml file in other repo, I'm not able to find the pt version of the weights file.

Regarding the v3 repo, If i use the cfg, weights format (and the yaml for the --data argument) I have different errors, like:

yaml.scanner.ScannerError: mapping values are not allowed here in "./models/yolov4-tiny.cfg", line 128, column 65

or some other things related to file incompatibility.

Given that someone said here that he was able to use YOLOv4 with this implementation, I want to know if I have to use a specific version, or some specific files to make things work.

[glenn-jocher]

@antimo22 we had a trio of YOLOv4 PRs in YOLOv5 from users but we were not able to validate their performance against the official paper and they have not been merged. They are:

• https://github.com/ultralytics/yolov5/pull/153
• https://github.com/ultralytics/yolov5/pull/2651
• https://github.com/ultralytics/yolov5/pull/2681

You can use these as a starting point if you'd like, though they are a bit out of date. If you succeed in implementing YOLOv4 in the YOLOv5 repo please submit a PR to help other users. Thank you!