单尺度0.791! InsightFace赛高！RetinaFace MobileNet0.25同人模型公开

[yangfly]

RetinaFace 牛逼！:two_hearts: InsightFace 赛高！:two_hearts:

开放 mobilenet0.25 版 RetinaFace 同人版模型
更完整的模型、日志和WiderFace测试截图 百度云 提取码:nzof GoogleDrive

模型说明，使用的是Gluon Model Zoo 的标准版 MobileNet0.25 预训练模型。（没有 fast downsampling，模型大小 1.68Mb）
Batch-size 32x2（两块1080Ti），其他参数都是 RetinaFace 默认的。

WIDER Face Hard 单尺度测试：0.791

WIDER Face Hard 多尺度测试：0.825

欢迎调参大佬，公开更给力的结果。:beer:

[BrownOfSummer]

非常感谢

[nttstar]

Great! Thanks for sharing!

[MyraBaba]

@yangfly how can we download mobileNet model ? Dropbox better in Europe ?

Can you share the model link?

[yangfly]

@MyraBaba Update Google Drive

[MyraBaba]

@yangfly thanks a lot

is the detector line :

detector = RetinaFace('/Users/xxx/Downloads/mnet.25/mobilenet_0_25' , 0 , gpuid , 'net3')

is this correct ? because it gives error:

File "/Users/xxx/Projects/new_/insight_10May/insightface/RetinaFace/retinaface.py", line 228, in detect
scores = scores[:, self._num_anchors['stride%s'%s]:, :, :]
IndexError: too many indices for array

[Neltherion]

@yangfly Thanks for uploading the models. Could you explain what the difference is between mobilenet_0_25-0000 and mnet.25-0000 ?

[dexception]

@yangfly @nttstar
Can we convert this to TVM ?

[yangfly]

@yangfly Thanks for uploading the models. Could you explain what the difference is between mobilenet_0_25-0000 and mnet.25-0000 ?

mobilenet_0_25-0000 is the pre-trained model from Gluon Model Zoo and mnet.25-0000 is the detector that I trained.

[Neltherion]

@yangfly Ah... I get it... thanks... (I was thinking they BOTH are Detector models)

[MyraBaba]

@yangfly quick question:

When I try mobile model mode cpu is using 4 cores and perform very well and fast, but normal model using only one core and pretty slooooow in cpu.

any idea why ? how can I push to use all available cores ?

Best

[yangfly]

Might helpful. https://mxnet.incubator.apache.org/versions/master/faq/perf.html#intel-cpu

[vgg4resnet]

has anyone convert it to caffe or tensorflow pb ，will u share it？

[alphinside]

Hi, @yangfly can you share with us the config.py used for training ? Thank you

[yangfly]

Hi, @yangfly can you share with us the config.py used for training ? Thank you

The config.py is added to BaiduYun and GoogleDrive.

[BrownOfSummer]

has anyone convert it to caffe or tensorflow pb ，will u share it？

+1

[yantiantianlx]

1 测试 （WIDER Face Hard 单尺度测试：0.791） 这个时的scale和阈值是怎么设定的。
2 测试widerface时， scale一直为1.0， 阈值0.02， easy为0.897，medium为0.874， hard 为 0.703, hard曲线有点差

[nttstar]

参考test_widerface, 短边1600 @ytt1790579195

[yangfly]

我没有修改 test_widerface.py 和 retinaface.py 中的任何参数

[JasonZhao001]

Hi, @yangfly can you share with us the config.py used for training ? Thank you

The config.py is added to BaiduYun and GoogleDrive.

Hi, @yangfly 我发现在你提供的config中使用的是network.mnet.pretrained = 'model/mobilenet025fd0' #78，这个mobilenet025fd0代表的是fast downsample的mobilenet0_25吧。从哪里能获取到呢

[JasonZhao001]

@nttstar Do you know where to get the pretrained model named "mobilenet025fd0"?

[yangfly]

@nttstar Do you know where to get the pretrained model named "mobilenet025fd0"?
I didn't use mobilenet025fd0.

[Charrin]

mnet0.25的anchor设置貌似没有scale step，默认使用了RAC_SSH的配置，anchor的大小与论文中Table2不一致，对吗（默默过来确认一下

[yangfly]

配置参数，在 train.log 中都打印出来了。

[Charrin]

配置参数，在 train.log 中都打印出来了。

嗯～好滴谢谢啦

[ashuezy]

@yangfly
Can you help regarding TVM conversion ?

[zhouwei5113]

这个MobileNet0.25是v1还是v2？

[shuyikong]

你好，请问为什么我将你训练好的mnet.25在cpu上做推断时得到的结果与gpu不一致，且差别很大呢？

[yangfly]

1. TVM conversion: You may convert by yourself for the different hardware platforms.
2. MobileNet version: v1 0.25
3. 结果不一致：你可能需要 debug 下，比如中间结果，配置参数.

[jamessmith90]

I think we can run this model on TensorRT.

Mobilenet Support is available.
https://mxnet.incubator.apache.org/_modules/mxnet/gluon/model_zoo/vision.html

@yangfly
Can you confirm this will work ?

https://mxnet.incubator.apache.org/versions/master/tutorials/tensorrt/inference_with_trt.html

[baishiruyue]

hello @yangfly
我在mxnet框架下面测试你提供的mnet.25模型，发现耗时和retinaface-R50模型的耗时相差不大，甚至还要更耗时，并且两个模型相差100多M，它们所占用的gpu显存也相差不大，我对这个问题很疑惑，是我哪些设置有问题吗？希望您的指教，多谢！

我用的gpu是titan xp。

[oftenliu]

配置参数，在 train.log 中都打印出来了。
请问下train.log是怎么生成的呀

[yangfly]

command 2>&1 | tee train.log

[wanjinchang]

@yangfly Good job!!When I test this model,I found that it consumes so much memory about 3G,could you tell me why?

[Dockyin]

请问下mnet.25模型在手机上测试过吗？
我测试时发现mobilenet_0_25可以在android平台上给出预测结果，虽说只有1*1000维度；但是使用mnet.25的话，加载模型时就会报错。
Android平台的方案如下：https://github.com/apache/incubator-mxnet/tree/master/amalgamation

[skyeyeCV]

@yangfly Did you only use wider face dataset when training?What is the batch size when you training?Thanks a lot!

[israelwei]

Did you use different learning rate & learning rate schedule while training on mobilenet (compared to resnet)? I don't see these values in the config files...

[gripdestiny]

@yangfly 你好，参考了你提供的训练日志，我发现FG acc 开始会往下降到0，然后在回升。这个现象是正常的么，我遇到的问题是FG acc下降到0以后就不会再回升，你有遇到过这问题么

[yangfly]

抱歉，没有遇到这种现象。------------------ 原始邮件 ------------------ 发件人: "gripdestiny"<notifications@github.com> 发送时间: 2019年12月5日(星期四) 下午5:30 收件人: "deepinsight/insightface"<insightface@noreply.github.com>; 抄送: "杨帆"<yf.dl@qq.com>;"Mention"<mention@noreply.github.com>; 主题: Re: [deepinsight/insightface] 单尺度0.791! InsightFace赛高！RetinaFace MobileNet0.25同人模型公开 (#669) @yangfly 你好，参考了你提供的训练日志，我发现FG acc 开始会往下降到0，然后在回升。这个现象是正常的么，我遇到的问题是FG acc下降到0以后就不会再回升，你有遇到过这问题么 &mdash; You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[lbin]

https://github.com/lbin/Retinaface_Mobilenet_Pytorch 单尺度相同设置hard达到85.55%

[brealisty]

hello @yangfly
我在mxnet框架下面测试你提供的mnet.25模型，发现耗时和retinaface-R50模型的耗时相差不大，甚至还要更耗时，并且两个模型相差100多M，它们所占用的gpu显存也相差不大，我对这个问题很疑惑，是我哪些设置有问题吗？希望您的指教，多谢！

我用的gpu是titan xp。

耗时问题你解决了吗？我用了@lbin，的模型，先热加载（执行一次forward），再做预测，时间耗时非常非常短。实时完全没问题

[AdaAkanishi]

@yangfly 你好，参考了你提供的训练日志，我发现FG acc 开始会往下降到0，然后在回升。这个现象是正常的么，我遇到的问题是FG acc下降到0以后就不会再回升，你有遇到过这问题么

你好，我想问一下你解决了这个问题嘛

[BenjaminLiu1003]

老哥，我用你的mobilenet_0_25-symbol.json和最新的retinaface的代码训练，得到的mnet.25-symbol.json和你自己训练得到的网络结构有一点局部区别，请问是怎么回事啊？

[chenzhengnan]

hello @yangfly
我在mxnet框架下面测试你提供的mnet.25模型，发现耗时和retinaface-R50模型的耗时相差不大，甚至还要更耗时，并且两个模型相差100多M，它们所占用的gpu显存也相差不大，我对这个问题很疑惑，是我哪些设置有问题吗？希望您的指教，多谢！

我用的gpu是titan xp。

你好，请问你在gpu和cpu下用R50测试图片单张耗时多久，

[chenzhengnan]

hello @yangfly
我在mxnet框架下面测试你提供的mnet.25模型，发现耗时和retinaface-R50模型的耗时相差不大，甚至还要更耗时，并且两个模型相差100多M，它们所占用的gpu显存也相差不大，我对这个问题很疑惑，是我哪些设置有问题吗？希望您的指教，多谢！
我用的gpu是titan xp。

耗时问题你解决了吗？我用了@lbin，的模型，先热加载（执行一次forward），再做预测，时间耗时非常非常短。实时完全没问题

你好，我在cpu上用r50单张测试耗时要5s左右，请问你是怎么提速的

[dexception]

+1 Waiting for fine tuned optimized version with tensorRT or TVM.

Can be quantize it ? fp16 or int8 with minimal loss in accuracy ?
or an AutoTune version ?

[jinpeijie217]

@yangfly 您好，有个问题，在使用test_widerface.py时，怎么分开测试easy，medium，hard三种测试集呢？或者怎么分开测试验证集呢？图片中的recall all是不是就是最终参考的指标？ 谢谢您的时间

[ashuezy]

@yangfly What is the output shape of this mobilenet model ?

[geoffzhang]

请问，有测试过mnet0.25使用cpu在VGA分辨率下的时间吗？我在ncnn框架下测试时间为50ms（单线程），和文章相差较大 @leo2105 @yangfly @ashuezy @chenzhengnan

[meejeong]

@yangfly, I executed retinaface based on mobilenet0.25 (you trained) from onnx file.
I checked the weight of the onnx file was the same as the pretrained model.
When I checked confidence (face_rpn_cls_prob_reshape_stride32, face_rpn_cls_prob_reshape_stride16, face_rpn_cls_prob_reshape_stride8), the value is too weired.
Can you share a preprocessing method? or do you have similar problems?

00000 = {float} 0.9999858140945435
00001 = {float} 0.9999855756759644
00002 = {float} 0.9999837875366211
00003 = {float} 0.9999805688858032
00004 = {float} 0.9999803304672241
00005 = {float} 0.9999790191650391
00006 = {float} 0.9999761581420898
00007 = {float} 0.9999642372131348
00008 = {float} 0.999963641166687
00009 = {float} 0.999963641166687
00010 = {float} 0.999962568283081
00011 = {float} 0.999962329864502
00012 = {float} 0.9999611377716064
00013 = {float} 0.9999608993530273

[cyrusbehr]

@yangfly What is the license on the models your trained? Is it MIT?