610265158 / face_landmark

A simple method for face alignment based on wingloss and mutitask learning :)
Apache License 2.0
251 stars 80 forks source link

想问一下训练的最终指标是多少? #11

Closed XP-CHRC closed 4 years ago

XP-CHRC commented 4 years ago

请问您最后在Mobilenet V2上训练的指标是多少(IPN or ION)?

610265158 commented 4 years ago

请问您最后在Mobilenet V2上训练的指标是多少(IPN or ION)?

hi, I did not try to measure any metrics, there I mainly concern about the loss value. Sorry for that.

XP-CHRC commented 4 years ago

@610265158 非常感谢您的回复,请问最终的loss是多少呢?并且您的工程可以基于CPU训练么?我尝试了简单的更改代码使其能够运行在CPU上,但是我不是太了解分布式训练,并没有成功。

610265158 commented 4 years ago

@610265158 非常感谢您的回复,请问最终的loss是多少呢?并且您的工程可以基于CPU训练么?我尝试了简单的更改代码使其能够运行在CPU上,但是我不是太了解分布式训练,并没有成功。

loss大概在4上下, 改成cpu的话, 把train.py 40行
devices = ['/device:CPU:{}'.format(i) for i in range(cfg.TRAIN.num_gpu)] strategy = tf.distribute.MirroredStrategy(devices) cpu可能会慢一些,加载一下训好的模型参数吧

另外,我主要在shufflenet结构上跑的,mobilenet没怎么跑.

chaotiaor commented 4 years ago

你好,我在试图把你用tensorflow的模型转化为pytorch的模型,但是,模型中的关键词命名并没有在工程中找到,然后,我使用tensorbord可视化你的模型,也无法可视化出来,所以我想问一下你模型中的结构是怎么的。我用的是tensorflow 1.14的版本

610265158 commented 4 years ago

你好,我在试图把你用tensorflow的模型转化为pytorch的模型,但是,模型中的关键词命名并没有在工程中找到,然后,我使用tensorbord可视化你的模型,也无法可视化出来,所以我想问一下你模型中的结构是怎么的。我用的是tensorflow 1.14的版本

hi, 我觉得这不是一个很方便的事情, 为什么要迁移到pytorch呢

chaotiaor commented 4 years ago

我准备在工程里直接布置你训练出来的模型,但是,用opencv里的dnn板块不能直接读取,安装TensorFlow的c++版本太过费时间,所以我就打算把它转成pytorch的,要是不好搞的话,我只有用pytorch重写一次,然后再进行训练了。

---原始邮件--- 发件人: "Lz"<notifications@github.com> 发送时间: 2019年11月13日(星期三) 晚上7:38 收件人: "610265158/face_landmark"<face_landmark@noreply.github.com>; 抄送: "Comment"<comment@noreply.github.com>;"rong"<zhaortiang@foxmail.com>; 主题: Re: [610265158/face_landmark] 想问一下训练的最终指标是多少? (#11)

你好,我在试图把你用tensorflow的模型转化为pytorch的模型,但是,模型中的关键词命名并没有在工程中找到,然后,我使用tensorbord可视化你的模型,也无法可视化出来,所以我想问一下你模型中的结构是怎么的。我用的是tensorflow 1.14的版本

hi, 我觉得这不是一个很方便的事情, 为什么要迁移到pytorch呢

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

610265158 commented 4 years ago

我准备在工程里直接布置你训练出来的模型,但是,用opencv里的dnn板块不能直接读取,安装TensorFlow的c++版本太过费时间,所以我就打算把它转成pytorch的,要是不好搞的话,我只有用pytorch重写一次,然后再进行训练了。 ---原始邮件--- 发件人: "Lz"<notifications@github.com> 发送时间: 2019年11月13日(星期三) 晚上7:38 收件人: "610265158/face_landmark"<face_landmark@noreply.github.com>; 抄送: "Comment"<comment@noreply.github.com>;"rong"<zhaortiang@foxmail.com>; 主题: Re: [610265158/face_landmark] 想问一下训练的最终指标是多少? (#11) 你好,我在试图把你用tensorflow的模型转化为pytorch的模型,但是,模型中的关键词命名并没有在工程中找到,然后,我使用tensorbord可视化你的模型,也无法可视化出来,所以我想问一下你模型中的结构是怎么的。我用的是tensorflow 1.14的版本 hi, 我觉得这不是一个很方便的事情, 为什么要迁移到pytorch呢 — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

结构是一个类似shufflenet 的网络,试一下tflite? 转换的话,有时间我来搞一下,如果pytorch更方便的话,我可以帮助迁移到pytorch上面

chaotiaor commented 4 years ago

好的,非常感谢你的回复,我明天会继续尝试一下,不过,我可能会用pytorch再写一次,然后把模型转成c++,测试它是不是可以在工程中进行应用。之所以使用pytorch,主要是我对TensorFlow已经不是很熟悉了,现在的主流好像还是TensorFlow。

---原始邮件--- 发件人: "Lz"<notifications@github.com> 发送时间: 2019年11月13日(星期三) 晚上7:52 收件人: "610265158/face_landmark"<face_landmark@noreply.github.com>; 抄送: "Comment"<comment@noreply.github.com>;"rong"<zhaortiang@foxmail.com>; 主题: Re: [610265158/face_landmark] 想问一下训练的最终指标是多少? (#11)

我准备在工程里直接布置你训练出来的模型,但是,用opencv里的dnn板块不能直接读取,安装TensorFlow的c++版本太过费时间,所以我就打算把它转成pytorch的,要是不好搞的话,我只有用pytorch重写一次,然后再进行训练了。 … ---原始邮件--- 发件人: "Lz"<notifications@github.com> 发送时间: 2019年11月13日(星期三) 晚上7:38 收件人: "610265158/face_landmark"<face_landmark@noreply.github.com>; 抄送: "Comment"<comment@noreply.github.com>;"rong"<zhaortiang@foxmail.com>; 主题: Re: [610265158/face_landmark] 想问一下训练的最终指标是多少? (#11) 你好,我在试图把你用tensorflow的模型转化为pytorch的模型,但是,模型中的关键词命名并没有在工程中找到,然后,我使用tensorbord可视化你的模型,也无法可视化出来,所以我想问一下你模型中的结构是怎么的。我用的是tensorflow 1.14的版本 hi, 我觉得这不是一个很方便的事情, 为什么要迁移到pytorch呢 — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

结构是一个类似shufflenet 的网络,试一下tflite? 转换的话,有时间我来搞一下,如果pytorch更方便的话,我可以帮助迁移到pytorch上面

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

chaotiaor commented 4 years ago

你好,我使用tensorflow 2.0,使用你项目中转化模型成为tflite格式的代码进行转化,会报错,这是为什么呢

610265158 commented 4 years ago

你好,我使用tensorflow 2.0,使用你项目中转化模型成为tflite格式的代码进行转化,会报错,这是为什么呢

报什么错呢,我就是用这个转的呀

chaotiaor commented 4 years ago

这种错误, tensorflow.lite.python.convert.ConverterError: See console for info.

------------------ 原始邮件 ------------------ 发件人: "Lz"<notifications@github.com>; 发送时间: 2019年11月15日(星期五) 上午10:41 收件人: "610265158/face_landmark"<face_landmark@noreply.github.com>; 抄送: "逍遥游"<zhaortiang@foxmail.com>; "Comment"<comment@noreply.github.com>; 主题: Re: [610265158/face_landmark] 想问一下训练的最终指标是多少? (#11)

你好,我使用tensorflow 2.0,使用你项目中转化模型成为tflite格式的代码进行转化,会报错,这是为什么呢

报什么错呢,我就是用这个转的呀

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

610265158 commented 4 years ago

你转的哪个模型, 需要更详细的报错信息

chaotiaor commented 4 years ago

face_landmark_shufflenetv2_0.75 ,这个模型

错误信息如下 File "C:\Users\meice\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow_core\lite\python\convert.py", line 200, in toco_convert_protos raise ConverterError("See console for info.\n%s\n%s\n" % (stdout, stderr)) tensorflow.lite.python.convert.ConverterError: See console for info. b"'toco_from_protos' \xb2\xbb\xca\xc7\xc4\xda\xb2\xbf\xbb\xf2\xcd\xe2\xb2\xbf\xc3\xfc\xc1\xee\xa3\xac\xd2\xb2\xb2\xbb\xca\xc7\xbf\xc9\xd4\xcb\xd0\xd0\xb5\xc4\xb3\xcc\xd0\xf2\r\n\xbb\xf2\xc5\xfa\xb4\xa6\xc0\xed\xce\xc4\xbc\xfe\xa1\xa3\r\n"

610265158 commented 4 years ago

face_landmark_shufflenetv2_0.75 ,这个模型

错误信息如下 File "C:\Users\meice\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow_core\lite\python\convert.py", line 200, in toco_convert_protos raise ConverterError("See console for info.\n%s\n%s\n" % (stdout, stderr)) tensorflow.lite.python.convert.ConverterError: See console for info. b"'toco_from_protos' \xb2\xbb\xca\xc7\xc4\xda\xb2\xbf\xbb\xf2\xcd\xe2\xb2\xbf\xc3\xfc\xc1\xee\xa3\xac\xd2\xb2\xb2\xbb\xca\xc7\xbf\xc9\xd4\xcb\xd0\xd0\xb5\xc4\xb3\xcc\xd0\xf2\r\n\xbb\xf2\xc5\xfa\xb4\xa6\xc0\xed\xce\xc4\xbc\xfe\xa1\xa3\r\n"

刚测试了,没有遇到这个问题啊, ubuntu 16.04 , 没有测试windows

chaotiaor commented 4 years ago

我是在windows下运行的

------------------ 原始邮件 ------------------ 发件人: "Lz"<notifications@github.com>; 发送时间: 2019年11月15日(星期五) 上午10:57 收件人: "610265158/face_landmark"<face_landmark@noreply.github.com>; 抄送: "逍遥游"<zhaortiang@foxmail.com>; "Comment"<comment@noreply.github.com>; 主题: Re: [610265158/face_landmark] 想问一下训练的最终指标是多少? (#11)

face_landmark_shufflenetv2_0.75 ,这个模型

错误信息如下 File "C:\Users\meice\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow_core\lite\python\convert.py", line 200, in toco_convert_protos raise ConverterError("See console for info.\n%s\n%s\n" % (stdout, stderr)) tensorflow.lite.python.convert.ConverterError: See console for info. b"'toco_from_protos' \xb2\xbb\xca\xc7\xc4\xda\xb2\xbf\xbb\xf2\xcd\xe2\xb2\xbf\xc3\xfc\xc1\xee\xa3\xac\xd2\xb2\xb2\xbb\xca\xc7\xbf\xc9\xd4\xcb\xd0\xd0\xb5\xc4\xb3\xcc\xd0\xf2\r\n\xbb\xf2\xc5\xfa\xb4\xa6\xc0\xed\xce\xc4\xbc\xfe\xa1\xa3\r\n"

刚测试了,没有遇到这个问题啊, ubuntu 16.04 , 没有测试windows

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

610265158 commented 4 years ago

我是在windows下运行的 ------------------ 原始邮件 ------------------ 发件人: "Lz"<notifications@github.com>; 发送时间: 2019年11月15日(星期五) 上午10:57 收件人: "610265158/face_landmark"<face_landmark@noreply.github.com>; 抄送: "逍遥游"<zhaortiang@foxmail.com>; "Comment"<comment@noreply.github.com>; 主题: Re: [610265158/face_landmark] 想问一下训练的最终指标是多少? (#11) face_landmark_shufflenetv2_0.75 ,这个模型 错误信息如下 File "C:\Users\meice\Anaconda3\envs\tensorflow\lib\site-packages\tensorflow_core\lite\python\convert.py", line 200, in toco_convert_protos raise ConverterError("See console for info.\n%s\n%s\n" % (stdout, stderr)) tensorflow.lite.python.convert.ConverterError: See console for info. b"'toco_from_protos' \xb2\xbb\xca\xc7\xc4\xda\xb2\xbf\xbb\xf2\xcd\xe2\xb2\xbf\xc3\xfc\xc1\xee\xa3\xac\xd2\xb2\xb2\xbb\xca\xc7\xbf\xc9\xd4\xcb\xd0\xd0\xb5\xc4\xb3\xcc\xd0\xf2\r\n\xbb\xf2\xc5\xfa\xb4\xa6\xc0\xed\xce\xc4\xbc\xfe\xa1\xa3\r\n" 刚测试了,没有遇到这个问题啊, ubuntu 16.04 , 没有测试windows — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

把模型发我,我试一下,

chaotiaor commented 4 years ago

发到你的邮箱吗

610265158 commented 4 years ago

发到你的邮箱吗

luckynote commented 4 years ago

请问您最后在Mobilenet V2上训练的指标是多少(IPN or ION)?

IPN Method Common(%) Challenging(%) Fullset(%)
PIG 0.75x 5.57 9.01 6.25
ION Method Common(%) Challenging(%) Fullset(%)
PIG 0.75 4.02 6.24 4.46