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

Question

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

yangfly opened this issue 5 years ago · 58 comments

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:

Answer 1 · 2019-05-10T12:08:48.000Z

非常感谢

Answer 2 · 2019-05-11T01:29:23.000Z

Great! Thanks for sharing!

Answer 3 · 2019-05-11T09:09:28.000Z

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

Can you share the model link?

Answer 4 · 2019-05-11T12:50:32.000Z

@MyraBaba Update Google Drive

Answer 5 · 2019-05-11T19:00:00.000Z

@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

Answer 6 · 2019-05-11T19:17:54.000Z

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

Answer 7 · 2019-05-11T20:51:25.000Z

@yangfly @nttstar
Can we convert this to TVM ?

Answer 8 · 2019-05-12T09:03:50.000Z

@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.

Answer 9 · 2019-05-12T09:09:16.000Z

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

Answer 10 · 2019-05-12T09:25:26.000Z

@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

Answer 11 · 2019-05-12T09:39:08.000Z

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

Answer 12 · 2019-05-13T05:52:17.000Z

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

Answer 13 · 2019-05-14T01:01:25.000Z

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

Answer 14 · 2019-05-14T01:06:57.000Z

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

The config.py is added to BaiduYun and GoogleDrive.

Answer 15 · 2019-05-14T09:52:29.000Z

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

+1

Answer 16 · 2019-05-15T03:46:17.000Z

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

Answer 17 · 2019-05-15T10:24:34.000Z

参考test_widerface, 短边1600 @ytt1790579195

Answer 18 · 2019-05-15T10:31:44.000Z

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

Answer 19 · 2019-05-16T08:38:32.000Z

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

Answer 20 · 2019-05-20T01:56:59.000Z

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

Answer 21 · 2019-05-21T02:48:26.000Z

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

Answer 22 · 2019-05-22T15:11:20.000Z

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

Answer 23 · 2019-05-22T15:13:40.000Z

配置参数，在 train.log 中都打印出来了。

Answer 24 · 2019-05-22T15:19:34.000Z

配置参数，在 train.log 中都打印出来了。

嗯～好滴谢谢啦

Answer 25 · 2019-05-23T06:13:34.000Z

@yangfly
Can you help regarding TVM conversion ?

Answer 26 · 2019-05-24T02:59:26.000Z

这个MobileNet0.25是v1还是v2？

Answer 27 · 2019-05-28T01:22:17.000Z

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？

Answer 28 · 2019-05-31T09:16:25.000Z

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

Answer 29 · 2019-06-14T09:14:05.000Z

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

Answer 30 · 2019-06-18T07:49:08.000Z

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。

Answer 31 · 2019-06-19T12:46:42.000Z

@MyraBaba
how did you resolved the below 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

please help

Answer 32 · 2019-06-19T12:49:14.000Z

@aneesh0

I havent faced such error yet.

EDIT:1
sorry I noticed that I have :)

let me look and get back to you

Answer 33 · 2019-06-19T12:51:01.000Z

@MyraBaba
how did you resolved the below 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
please help

hello, you can print the shape of scores, i think it is 3 dementions.

Answer 34 · 2019-06-24T02:45:47.000Z

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

The config.py is added to BaiduYun and GoogleDrive.

hello,could you share the link? I can`t find in the shared links.

Answer 35 · 2019-06-26T09:59:41.000Z

if i want to save my model each 5 epochs ,how to set the param?

Answer 36 · 2019-07-02T03:34:57.000Z

@aneesh0

I havent faced such error yet.

EDIT:1
sorry I noticed that I have :)

let me look and get back to you

Have you found it? :)

Answer 37 · 2019-07-03T14:07:33.000Z

@yangfly muchas gracias

es la linea de detectores:

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

es esto correcto ? porque da error:

Archivo "/Users/xxx/Projects/new_/insight_10May/insightface/RetinaFace/retinaface.py", línea 228, en
puntajes de detección = puntajes [:, self._num_anchors ['stride% s'% s] :,:,: ]
IndexError: demasiados índices para la matriz

I got the same problem
I dont know if @yangfly @MyraBaba can tell how you solved.
thanks

Answer 38 · 2019-07-04T07:10:59.000Z

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

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

mobilenet_0_25 is just the pre-trained.
You should use mnet.25 instead.

Answer 39 · 2019-08-06T08:44:04.000Z

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

Answer 40 · 2019-08-06T08:46:54.000Z

command 2>&1 | tee train.log

Answer 41 · 2019-08-08T01:43:13.000Z

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

Answer 42 · 2019-08-13T14:50:08.000Z

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

Answer 43 · 2019-08-23T11:00:28.000Z

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

Answer 44 · 2019-09-25T07:19:59.000Z

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...

Answer 45 · 2019-12-05T09:30:21.000Z

@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

Answer 46 · 2019-12-09T04:25:11.000Z

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

Answer 47 · 2019-12-10T08:14:52.000Z

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

Answer 48 · 2020-02-16T04:16:04.000Z

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

Answer 49 · 2020-04-10T06:34:41.000Z

@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

Answer 50 · 2020-05-14T01:45:56.000Z

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？

Answer 51 · 2020-07-28T06:40:42.000Z

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，

Answer 52 · 2020-07-28T07:06:40.000Z

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

Answer 53 · 2020-07-28T08:33:24.000Z

+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 ?

Answer 54 · 2020-08-05T13:50:22.000Z

@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

Answer 55 · 2020-08-25T19:21:05.000Z

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

Answer 56 · 2020-09-03T07:13:18.000Z

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

Answer 57 · 2020-12-23T10:01:08.000Z

@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

Answer 58 · 2021-04-01T18:36:20.000Z

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