daybreak-u / darknet_face_with_landmark Goto Github PK

Dear all,

I try to train this model with wider_face data in the darknet, some time, total_loss is nan when in yolo layer "iou_loss=ciou". When i change "iou_loss=diou", this problem doesnt occur.

Did anyone have same problem. Can you suggest me solve it. Thank you

Good job!I wonder the test speed.Could you provide the test speed in gpu or cpu environment?Thank you!

大佬好，请问在 AlexeyAB 的哪个commit上搞的啊？
谢谢！

*** Error in `./darknet': double free or corruption (out): 0x000000005ef54080 ***

请问有没有好工具能把模型转为caffe的？ 尝试把onnx的转为caffe时，onnx2caffe中报错LeakyRelu is not supported

version.h fatal error

您說與darknet上訓練相同
在darknet上訓練yolov3,會先需要下載darknet53.conv預訓練來進行訓練
大老您也是用darknet53.conv來做嗎?
那另外想請問 mbv3small_yolov3_face.cfg要用哪一個來進行預訓練? 是yolov3-tiny.conv.11?!

目前只能使用yolov3嗎 能使用到tiny-yolov3嗎? 或者甚至是yolov4/tiny-yolov4

謝謝

@DayBreak-u

感谢楼主分享哈^_^
尝试把代码合到darknet最新版本，发现这里的循环又用到了i
https://github.com/ouyanghuiyu/darknet_face_with_landmark/blob/2db381ee665fd16ae0dda62749e5411523371e97/src/yolo_layer.c#L724

I want to use your mobilenetv2-yolov3 to train my own dataset, do you have pretrained weights?

_

尊敬的作者，您好。
我想知道您在训练过程中数据集中label是什么格式的，是采用什么工具对人脸的关键点进行标注，谢谢。

你好，请问这项目是在C的环境下编译的吗，有没有python版本的呢？相对而言，这个darknet c版本编译比较难，不好上手啊。

您好，非常感谢您工作。我在使用yolov3进行预测后，生成的labels文件夹无法使用widerface文件夹中的evaluate.py进行测试，请问您是如何使用的呢，需要将yolo的训练结果进行转换吗？谢谢~

Hi, I was trying to make your repo but got this, how can I fix this?

./src/activations.c: In function ‘get_activation’:
./src/activations.c:59:42: error: ‘HSWISH’ undeclared (first use in this function); did you mean ‘SWISH’?
if (strcmp(s, "hswish") == 0) return HSWISH;
^~~~~~
SWISH
compilation terminated due to -Wfatal-errors.
Makefile:162: recipe for target 'obj/activations.o' failed
make: *** [obj/activations.o] Error 1

Also, what is the logic in your cfg file?
I was trying to run the model in Alexey, the filters before every yolo layer should be
(class+coords+1)*mask right? But when I change it doesn't work.

Thank you!

你好，
非常感谢你的工作。上周用你的代码训练了其他任务的检测和关键点定位模型，但是在训练的过程中，总是出现 《读取图片的过程中，线程一直等待，无法继续训练问题》，

图片内容可能不能显示，具体内容如下：
第一次训练：
2244: 2.664922，2.490456 avg loss, 0.001000 rate, 0.701170 seconds, 71808 images, 62.321252 hours left

第二次训练：

2256: 2.414187， 2.562203 avg loss, 0.001000 rate, 0.712135 seconds, 72192 images, 42.599252 hours left

如上面所示，训练了好几次，都是这种情况，也不报错，就是不继续运行了，我理解的应该是线程的问题，请问你能帮忙看看怎么解决这个问题吗？谢谢！

我还把开辟64个线程改为开辟1个线程，也是报上面的错误。

大佬开源，必定又是一个高赞项目。但此类任务pytorch等框架不是更好操作吗，模型转换也容易，大佬用darknet来做是基于什么考虑的呢？

您好 我在做转caffe及测试的时候发现 你的三个卷积层最后的输出通道数 好像和我想象的不一样 按我的理解 一个face类加xyhw confidence 是6 至少来说输出通道数应该是6的倍数 所以这里有点不太明白 您是不是后处理和yolo还不太一样

The valid label hasn't key points, only face box.
Could you please provide the validation label with key points？
Thank you!

膜拜大佬

大佬有尝试过v4里面的CSPDarknet53吗

Hi, buddy, thank your wonderful project, I was success face detect in custom images. But next I want crop the face region and output a .png file, how I modify the yolo_landmark.py(for example) demo like in the def cvDrawBoxes(detections, img ,ratio_w , ratio_h ) function, and let the predictions.jpg(yolo output) obtain a face_crop region(just I need a face local output). How can I let the yolo frame output it?

@ouyanghuiyu
你好: 第一次将人脸相关数据集移植到yolo里,不是很了解.我想请教步骤2的具体实施细节,是否要先下载对应的数据集呢.

Hi，
i use the weights provided and run the test with default parameter.
test image 'selfie.jpg' result is worse than your result.

在使用scripts/retinaface2yololandmark.py脚本将retinaface的标记文件转为yolo的格式使用

找不到数据信息

大佬你好，我用retinaface2yololandmark.py文件制作完数据集后，运行train_face_landmar.sh进行训练，然后打印出了如下错误：
Warning: in txt-label class_id = -2147483648 > classes=1 in cfg-file。 In txt-labels class_id should be [from 0 to 0]
truth.x = 很大的一个数

定位问题发现在scr/yolo_layer.c里面，然而目前还有解决。

所以有两个问题：

1. 定位的问题是不是准确的？ 2. 有啥建议的解决方案么。。。

PS： 肉眼看生成的label.txt没啥问题

I test this repo using 416*416 size. under win10 , NVIDIA GTX 1060 GPU
Each picture inference time is about average 112ms.
But when i use stand yolov3.cfg to detector ordinary object ,each picture takes about average 38 ms .
Why this network is so much slower than yolov3 network ?

when I ran this model on webcam I am getting 10 fps on 1080ti and when I typed nvidia-smi it's using nearly 400mb of GPU and CPU is also being used two core 100 percent.
here is the command I have used to run video
./darknet detector demo ./data/face.data ./cfg/mbv2_yolov3_face.cfg ./models/mbv2_yolov3_face_final.weights -c 0

您好，请问您使用的关键点检测技术是参考的哪篇论文呢？

@ouyanghuiyu good work even I was also working on a similar project on darknet. Can we use darknet to get 512 face embedding like InsightFace.

Dear author:

when I train or test this project, I will get this error! How can I fix this problem? Thanks!