nickliqian / cnn_captcha

use cnn recognize captcha by tensorflow. 本项目针对字符型图片验证码,使用tensorflow实现卷积神经网络,进行验证码识别。
Apache License 2.0
2.76k stars 783 forks source link

训练集识别率达到的1,但是验证集的识别率一直为0 #63

Open jjeejj opened 5 years ago

jjeejj commented 5 years ago

我现在的用代码生成的训练集 20000 张。训练成功后 训练集识别率达到的1,但是验证集的识别率一直为0 temp 这是为什么呢

jjeejj commented 5 years ago

然后执行 python test_batch.py 验证识别率 。识别率只有 2% image

ZhangYiBo513 commented 5 years ago

我的也是,我修改了图片尺寸之后就变这样了,如果用作者原始代码识别率还好,但是改动之后识别率太低了

nickliqian commented 5 years ago

@jjeejj @ZhangYiBo513 如果可以,把验证码和代码发给我试一下? 另外需要注意的地方,一个是标签char_set记得改成对应的数据,另外一个是继续增加样本试试。

ZhangYiBo513 commented 5 years ago

hi~: 首选感谢解惑回复,验证码我放在附件里面了; 我要识别的验证码感觉是不用修改char_set数据,我测试的样本大概有2000张可能有点少; 还有一点我很疑惑,为什么修改图片尺寸之后图片的识别率会很低很低,只有百分之零点几,同时我还修改了cycle_stop=8000, 虽然训练的次数少了,但是识别率也不应该这么低啊?所以还需要向您请教下原因。 感谢解惑!

------------------ 原始邮件 ------------------ 发件人: "Nick Li"notifications@github.com; 发送时间: 2019年7月12日(星期五) 晚上6:45 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63)

@jjeejj @ZhangYiBo513 如果可以,把验证码和代码发给我试一下? 另外需要注意的地方,一个是标签char_set记得改成对应的数据,另外一个是继续增加样本试试。

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

nickliqian commented 5 years ago

@ZhangYiBo513 看得出你的样本相对复杂,2000张样本不太够的。另外没有收到你的附件。。。

jjeejj commented 5 years ago

现在这个项目,我已经实际在使用了,首先感谢作者 识别率一直为0,基本上都是样本不太够,训练集的数据太少

改了图片的配置大小要和实际的图片大小保持一致

我遇到一个情况是:当训练到一万次的时候,识别率还是很低的只有 0.1,但是当寻训练到 11000次左右的时候,识别率瞬间上升,达到 百分之 八,九 十的

ZhangYiBo513 commented 5 years ago

好的,我近期会在收集一批样本,然后会通过邮件的形式发给你,但是我想知道你是怎么打标签的呢? 感谢!

------------------ 原始邮件 ------------------ 发件人: "Nick Li"notifications@github.com; 发送时间: 2019年7月12日(星期五) 晚上9:37 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63)

@ZhangYiBo513 看得出你的样本相对复杂,2000张样本不太够的。另外没有收到你的附件。。。

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

ZhangYiBo513 commented 5 years ago

是的,我改变的图片尺寸就是为了和要测试的样本尺寸一致; 我想问下,你的训练集样本是自己生成的吗?还是自己采集的然后打的标签? 可能是我训练的次数太少了,我把配置训练次数改成了8000,识别率一直很低;回头我再测试下。

------------------ 原始邮件 ------------------ 发件人: "YibuMe"notifications@github.com; 发送时间: 2019年7月12日(星期五) 晚上10:13 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63)

现在这个项目,我已经实际在使用了,首先感谢作者 识别率一直为0,基本上都是样本不太够,训练集的数据太少

改了图片的配置大小要和实际的图片大小保持一致

我遇到一个情况是:当训练到一万次的时候,识别率还是很低的只有 0.1,但是当寻训练到 11000次左右的时候,识别率瞬间上升,达到 百分之 八,九 十的

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

nickliqian commented 5 years ago

@ZhangYiBo513 打标签可以通过打码平台。我是已经有训练好的模型可以识别大部分的验证码。

l976308589 commented 5 years ago

请问您训练好的模型存放在哪里,能否分享?@nickliqian @nickliqian

xhj743694 commented 5 years ago

@ZhangYiBo513 打标签可以通过打码平台。我是已经有训练好的模型可以识别大部分的验证码。

大佬推荐些打码平台吧,感觉百度出来的都好不靠谱

jerry94test commented 5 years ago

你们测试过6位数的验证码吗 我跑了20w 轮 准确率 也是有 6% 使用的默认的 py 打吗 width 200 长度6 改了这2 个参数

ZhangYiBo513 commented 5 years ago

6位没有测试过,4位我最新测试,10000张样本准确率可以达到70%,6位的话更改配置文件,和其他配置没有问题的话应该也是ok的。

------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:02 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63)

你们测试过6位数的验证码吗 我跑了20w 词 准确率 也是有 6%

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

jerry94test commented 5 years ago

6位没有测试过,4位我最新测试,10000张样本准确率可以达到70%,6位的话更改配置文件,和其他配置没有问题的话应该也是ok的。 ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:02 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 你们测试过6位数的验证码吗 我跑了20w 词 准确率 也是有 6% — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

你测试一下就知道了 我测试过4 位数 在7000轮下就可以达到 0.70 我用6位数的 跑了一晚上都不行 还是只有 0.02-0.11 这是我的配置文件

{
  "root_dir": "sample/origin/",
  "image_suffix": "png",
  "characters": "0123456789abcdefghijklmnopqrstuvwxyz",
  "count": 20000,
  "char_count": 6,
  "width": 200,
  "height": 60
}
{
  "origin_image_dir": "sample/origin/",
  "new_image_dir": "sample/new_train/",
  "train_image_dir": "sample/train/",
  "test_image_dir": "sample/test/",
  "api_image_dir": "sample/api/",
  "online_image_dir": "sample/online/",
  "local_image_dir": "sample/local/",
  "model_save_dir": "model/",
  "image_width": 200,
  "image_height": 60,
  "max_captcha": 6,
  "image_suffix": "png",
  "char_set": "0123456789abcdefghijklmnopqrstuvwxyz",
  "use_labels_json_file": false,
  "remote_url": "http://127.0.0.1:6100/captcha/",
  "cycle_stop": 20000000,
  "acc_stop": 0.99,
  "cycle_save": 500,
  "enable_gpu": 1
}

安装包列表

_tflow_select             2.1.0                       gpu    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
absl-py                   0.7.1                    pypi_0    pypi
astor                     0.7.1                    py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
blas                      1.0                         mkl    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
bleach                    1.5.0                    pypi_0    pypi
captcha                   0.3                      pypi_0    pypi
certifi                   2019.3.9                 pypi_0    pypi
chardet                   3.0.4                    pypi_0    pypi
click                     7.0                      pypi_0    pypi
cudatoolkit               9.0                           1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
cudnn                     7.6.0                 cuda9.0_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
cycler                    0.10.0                   pypi_0    pypi
easydict                  1.8                      pypi_0    pypi
flask                     1.0.2                    pypi_0    pypi
gast                      0.2.2                    pypi_0    pypi
grpcio                    1.19.0                   pypi_0    pypi
h5py                      2.7.0               np113py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
hdf5                      1.8.15.1                      2    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
html5lib                  0.9999999                pypi_0    pypi
icc_rt                    2019.0.0             h0cc432a_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
idna                      2.7                      pypi_0    pypi
itsdangerous              1.1.0                    pypi_0    pypi
jinja2                    2.10.1                   pypi_0    pypi
keras-applications        1.0.8                      py_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
keras-preprocessing       1.1.0                      py_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
libprotobuf               3.6.0                h1a1b453_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
markdown                  3.1                      pypi_0    pypi
markupsafe                1.1.1                    pypi_0    pypi
matplotlib                2.1.0                    pypi_0    pypi
mkl                       2017.0.3                      0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
mock                      2.0.0                    py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
numpy                     1.16.2                   pypi_0    pypi
olefile                   0.46                     pypi_0    pypi
pbr                       1.10.0                   py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
pillow                    4.3.0                    pypi_0    pypi
pip                       19.2.1                   pypi_0    pypi
pyparsing                 2.4.0                    pypi_0    pypi
python                    3.6.9                h5500b2f_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
python-dateutil           2.8.0                    pypi_0    pypi
pytz                      2018.9                   pypi_0    pypi
requests                  2.19.1                   pypi_0    pypi
scipy                     0.19.1              np113py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
setuptools                36.4.0                   py36_1    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
six                       1.12.0                   pypi_0    pypi
sqlite                    3.29.0               he774522_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
tensorboard               1.7.0                    pypi_0    pypi
tensorflow                1.12.0          gpu_py36ha5f9131_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
tensorflow-base           1.12.0          gpu_py36h6e53903_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
tensorflow-gpu            1.12.0               h0d30ee6_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
termcolor                 1.1.0                    py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
urllib3                   1.23                     pypi_0    pypi
vc                        14.1                 h0510ff6_4    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
vs2015_runtime            14.15.26706          h3a45250_4    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
werkzeug                  0.15.2                   pypi_0    pypi
wheel                     0.29.0                   py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
wincertstore              0.2                      py36_0    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
zlib                      1.2.11               h62dcd97_3    https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
ZhangYiBo513 commented 5 years ago

样本是自己生成的吗? 我之前用生成的验证码改了尺寸以后也出现了这种状况,不过后来用了实际的样本,训练之后效果就好了,具体原因我也不清楚。

------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:20 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63)

6位没有测试过,4位我最新测试,10000张样本准确率可以达到70%,6位的话更改配置文件,和其他配置没有问题的话应该也是ok的。 … ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:02 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 你们测试过6位数的验证码吗 我跑了20w 词 准确率 也是有 6% — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

你测试一下就知道了 我测试过4 位数 在7000轮下就可以达到 0.70 我用6位数的 跑了一晚上都不行 还是只有 0.02-0.11 这是我的配置文件 { "root_dir": "sample/origin/", "image_suffix": "png", "characters": "0123456789abcdefghijklmnopqrstuvwxyz", "count": 20000, "char_count": 6, "width": 200, "height": 60 } { "origin_image_dir": "sample/origin/", "new_image_dir": "sample/new_train/", "train_image_dir": "sample/train/", "test_image_dir": "sample/test/", "api_image_dir": "sample/api/", "online_image_dir": "sample/online/", "local_image_dir": "sample/local/", "model_save_dir": "model/", "image_width": 200, "image_height": 60, "max_captcha": 6, "image_suffix": "png", "char_set": "0123456789abcdefghijklmnopqrstuvwxyz", "use_labels_json_file": false, "remote_url": "http://127.0.0.1:6100/captcha/", "cycle_stop": 20000000, "acc_stop": 0.99, "cycle_save": 500, "enable_gpu": 1 }

安装包列表 _tflow_select 2.1.0 gpu https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main absl-py 0.7.1 pypi_0 pypi astor 0.7.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main blas 1.0 mkl https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free bleach 1.5.0 pypi_0 pypi captcha 0.3 pypi_0 pypi certifi 2019.3.9 pypi_0 pypi chardet 3.0.4 pypi_0 pypi click 7.0 pypi_0 pypi cudatoolkit 9.0 1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cudnn 7.6.0 cuda9.0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cycler 0.10.0 pypi_0 pypi easydict 1.8 pypi_0 pypi flask 1.0.2 pypi_0 pypi gast 0.2.2 pypi_0 pypi grpcio 1.19.0 pypi_0 pypi h5py 2.7.0 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free hdf5 1.8.15.1 2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free html5lib 0.9999999 pypi_0 pypi icc_rt 2019.0.0 h0cc432a_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main idna 2.7 pypi_0 pypi itsdangerous 1.1.0 pypi_0 pypi jinja2 2.10.1 pypi_0 pypi keras-applications 1.0.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main keras-preprocessing 1.1.0 py_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main libprotobuf 3.6.0 h1a1b453_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main markdown 3.1 pypi_0 pypi markupsafe 1.1.1 pypi_0 pypi matplotlib 2.1.0 pypi_0 pypi mkl 2017.0.3 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free mock 2.0.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free numpy 1.16.2 pypi_0 pypi olefile 0.46 pypi_0 pypi pbr 1.10.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free pillow 4.3.0 pypi_0 pypi pip 19.2.1 pypi_0 pypi pyparsing 2.4.0 pypi_0 pypi python 3.6.9 h5500b2f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main python-dateutil 2.8.0 pypi_0 pypi pytz 2018.9 pypi_0 pypi requests 2.19.1 pypi_0 pypi scipy 0.19.1 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free setuptools 36.4.0 py36_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free six 1.12.0 pypi_0 pypi sqlite 3.29.0 he774522_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorboard 1.7.0 pypi_0 pypi tensorflow 1.12.0 gpu_py36ha5f9131_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-base 1.12.0 gpu_py36h6e53903_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-gpu 1.12.0 h0d30ee6_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main termcolor 1.1.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free urllib3 1.23 pypi_0 pypi vc 14.1 h0510ff6_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main vs2015_runtime 14.15.26706 h3a45250_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main werkzeug 0.15.2 pypi_0 pypi wheel 0.29.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free wincertstore 0.2 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free zlib 1.2.11 h62dcd97_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main

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

jerry94test commented 5 years ago

样本是自己生成的吗? 我之前用生成的验证码改了尺寸以后也出现了这种状况,不过后来用了实际的样本,训练之后效果就好了,具体原因我也不清楚。 ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:20 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 6位没有测试过,4位我最新测试,10000张样本准确率可以达到70%,6位的话更改配置文件,和其他配置没有问题的话应该也是ok的。 … ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:02 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 你们测试过6位数的验证码吗 我跑了20w 词 准确率 也是有 6% — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread. 你测试一下就知道了 我测试过4 位数 在7000轮下就可以达到 0.70 我用6位数的 跑了一晚上都不行 还是只有 0.02-0.11 这是我的配置文件 { "root_dir": "sample/origin/", "image_suffix": "png", "characters": "0123456789abcdefghijklmnopqrstuvwxyz", "count": 20000, "char_count": 6, "width": 200, "height": 60 } { "origin_image_dir": "sample/origin/", "new_image_dir": "sample/new_train/", "train_image_dir": "sample/train/", "test_image_dir": "sample/test/", "api_image_dir": "sample/api/", "online_image_dir": "sample/online/", "local_image_dir": "sample/local/", "model_save_dir": "model/", "image_width": 200, "image_height": 60, "max_captcha": 6, "image_suffix": "png", "char_set": "0123456789abcdefghijklmnopqrstuvwxyz", "use_labels_json_file": false, "remote_url": "http://127.0.0.1:6100/captcha/", "cycle_stop": 20000000, "acc_stop": 0.99, "cycle_save": 500, "enable_gpu": 1 } 安装包列表 _tflow_select 2.1.0 gpu https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main absl-py 0.7.1 pypi_0 pypi astor 0.7.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main blas 1.0 mkl https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free bleach 1.5.0 pypi_0 pypi captcha 0.3 pypi_0 pypi certifi 2019.3.9 pypi_0 pypi chardet 3.0.4 pypi_0 pypi click 7.0 pypi_0 pypi cudatoolkit 9.0 1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cudnn 7.6.0 cuda9.0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cycler 0.10.0 pypi_0 pypi easydict 1.8 pypi_0 pypi flask 1.0.2 pypi_0 pypi gast 0.2.2 pypi_0 pypi grpcio 1.19.0 pypi_0 pypi h5py 2.7.0 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free hdf5 1.8.15.1 2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free html5lib 0.9999999 pypi_0 pypi icc_rt 2019.0.0 h0cc432a_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main idna 2.7 pypi_0 pypi itsdangerous 1.1.0 pypi_0 pypi jinja2 2.10.1 pypi_0 pypi keras-applications 1.0.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main keras-preprocessing 1.1.0 py_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main libprotobuf 3.6.0 h1a1b453_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main markdown 3.1 pypi_0 pypi markupsafe 1.1.1 pypi_0 pypi matplotlib 2.1.0 pypi_0 pypi mkl 2017.0.3 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free mock 2.0.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free numpy 1.16.2 pypi_0 pypi olefile 0.46 pypi_0 pypi pbr 1.10.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free pillow 4.3.0 pypi_0 pypi pip 19.2.1 pypi_0 pypi pyparsing 2.4.0 pypi_0 pypi python 3.6.9 h5500b2f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main python-dateutil 2.8.0 pypi_0 pypi pytz 2018.9 pypi_0 pypi requests 2.19.1 pypi_0 pypi scipy 0.19.1 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free setuptools 36.4.0 py36_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free six 1.12.0 pypi_0 pypi sqlite 3.29.0 he774522_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorboard 1.7.0 pypi_0 pypi tensorflow 1.12.0 gpu_py36ha5f9131_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-base 1.12.0 gpu_py36h6e53903_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-gpu 1.12.0 h0d30ee6_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main termcolor 1.1.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free urllib3 1.23 pypi_0 pypi vc 14.1 h0510ff6_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main vs2015_runtime 14.15.26706 h3a45250_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main werkzeug 0.15.2 pypi_0 pypi wheel 0.29.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free wincertstore 0.2 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free zlib 1.2.11 h62dcd97_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

我使用脚本重新训练样本居然可以了 难道是一批样本训练了 3000轮就重新生生一批 可能是这样

ZhangYiBo513 commented 5 years ago

.....具体原因需要,可能需要作者看下。

------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:26 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63)

样本是自己生成的吗? 我之前用生成的验证码改了尺寸以后也出现了这种状况,不过后来用了实际的样本,训练之后效果就好了,具体原因我也不清楚。 … ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:20 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 6位没有测试过,4位我最新测试,10000张样本准确率可以达到70%,6位的话更改配置文件,和其他配置没有问题的话应该也是ok的。 … ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:02 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 你们测试过6位数的验证码吗 我跑了20w 词 准确率 也是有 6% — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread. 你测试一下就知道了 我测试过4 位数 在7000轮下就可以达到 0.70 我用6位数的 跑了一晚上都不行 还是只有 0.02-0.11 这是我的配置文件 { "root_dir": "sample/origin/", "image_suffix": "png", "characters": "0123456789abcdefghijklmnopqrstuvwxyz", "count": 20000, "char_count": 6, "width": 200, "height": 60 } { "origin_image_dir": "sample/origin/", "new_image_dir": "sample/new_train/", "train_image_dir": "sample/train/", "test_image_dir": "sample/test/", "api_image_dir": "sample/api/", "online_image_dir": "sample/online/", "local_image_dir": "sample/local/", "model_save_dir": "model/", "image_width": 200, "image_height": 60, "max_captcha": 6, "image_suffix": "png", "char_set": "0123456789abcdefghijklmnopqrstuvwxyz", "use_labels_json_file": false, "remote_url": "http://127.0.0.1:6100/captcha/", "cycle_stop": 20000000, "acc_stop": 0.99, "cycle_save": 500, "enable_gpu": 1 } 安装包列表 _tflow_select 2.1.0 gpu https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main absl-py 0.7.1 pypi_0 pypi astor 0.7.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main blas 1.0 mkl https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free bleach 1.5.0 pypi_0 pypi captcha 0.3 pypi_0 pypi certifi 2019.3.9 pypi_0 pypi chardet 3.0.4 pypi_0 pypi click 7.0 pypi_0 pypi cudatoolkit 9.0 1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cudnn 7.6.0 cuda9.0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cycler 0.10.0 pypi_0 pypi easydict 1.8 pypi_0 pypi flask 1.0.2 pypi_0 pypi gast 0.2.2 pypi_0 pypi grpcio 1.19.0 pypi_0 pypi h5py 2.7.0 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free hdf5 1.8.15.1 2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free html5lib 0.9999999 pypi_0 pypi icc_rt 2019.0.0 h0cc432a_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main idna 2.7 pypi_0 pypi itsdangerous 1.1.0 pypi_0 pypi jinja2 2.10.1 pypi_0 pypi keras-applications 1.0.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main keras-preprocessing 1.1.0 py_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main libprotobuf 3.6.0 h1a1b453_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main markdown 3.1 pypi_0 pypi markupsafe 1.1.1 pypi_0 pypi matplotlib 2.1.0 pypi_0 pypi mkl 2017.0.3 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free mock 2.0.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free numpy 1.16.2 pypi_0 pypi olefile 0.46 pypi_0 pypi pbr 1.10.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free pillow 4.3.0 pypi_0 pypi pip 19.2.1 pypi_0 pypi pyparsing 2.4.0 pypi_0 pypi python 3.6.9 h5500b2f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main python-dateutil 2.8.0 pypi_0 pypi pytz 2018.9 pypi_0 pypi requests 2.19.1 pypi_0 pypi scipy 0.19.1 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free setuptools 36.4.0 py36_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free six 1.12.0 pypi_0 pypi sqlite 3.29.0 he774522_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorboard 1.7.0 pypi_0 pypi tensorflow 1.12.0 gpu_py36ha5f9131_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-base 1.12.0 gpu_py36h6e53903_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-gpu 1.12.0 h0d30ee6_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main termcolor 1.1.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free urllib3 1.23 pypi_0 pypi vc 14.1 h0510ff6_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main vs2015_runtime 14.15.26706 h3a45250_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main werkzeug 0.15.2 pypi_0 pypi wheel 0.29.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free wincertstore 0.2 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free zlib 1.2.11 h62dcd97_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

我使用脚本重新训练样本居然可以了 难道是一批样本训练了 3000轮就重新生生一批 可能是这样

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

jerry94test commented 5 years ago

测试结果: 12/100 100个样本识别耗时6.816540002822876秒,准确率12.0% 删除 sample 文件夹 重新生产 训练1000 轮效果居然出其的好 如果是这个原因 这个问题应该可以关闭了

jerry94test commented 5 years ago

.....具体原因需要,可能需要作者看下。 ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:26 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 样本是自己生成的吗? 我之前用生成的验证码改了尺寸以后也出现了这种状况,不过后来用了实际的样本,训练之后效果就好了,具体原因我也不清楚。 … ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:20 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 6位没有测试过,4位我最新测试,10000张样本准确率可以达到70%,6位的话更改配置文件,和其他配置没有问题的话应该也是ok的。 … ------------------ 原始邮件 ------------------ 发件人: "jerry94test"notifications@github.com; 发送时间: 2019年8月1日(星期四) 晚上7:02 收件人: "nickliqian/cnn_captcha"cnn_captcha@noreply.github.com; 抄送: "单只鸳鸯"1057964760@qq.com;"Mention"mention@noreply.github.com; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63) 你们测试过6位数的验证码吗 我跑了20w 词 准确率 也是有 6% — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread. 你测试一下就知道了 我测试过4 位数 在7000轮下就可以达到 0.70 我用6位数的 跑了一晚上都不行 还是只有 0.02-0.11 这是我的配置文件 { "root_dir": "sample/origin/", "image_suffix": "png", "characters": "0123456789abcdefghijklmnopqrstuvwxyz", "count": 20000, "char_count": 6, "width": 200, "height": 60 } { "origin_image_dir": "sample/origin/", "new_image_dir": "sample/new_train/", "train_image_dir": "sample/train/", "test_image_dir": "sample/test/", "api_image_dir": "sample/api/", "online_image_dir": "sample/online/", "local_image_dir": "sample/local/", "model_save_dir": "model/", "image_width": 200, "image_height": 60, "max_captcha": 6, "image_suffix": "png", "char_set": "0123456789abcdefghijklmnopqrstuvwxyz", "use_labels_json_file": false, "remote_url": "http://127.0.0.1:6100/captcha/", "cycle_stop": 20000000, "acc_stop": 0.99, "cycle_save": 500, "enable_gpu": 1 } 安装包列表 _tflow_select 2.1.0 gpu https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main absl-py 0.7.1 pypi_0 pypi astor 0.7.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main blas 1.0 mkl https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free bleach 1.5.0 pypi_0 pypi captcha 0.3 pypi_0 pypi certifi 2019.3.9 pypi_0 pypi chardet 3.0.4 pypi_0 pypi click 7.0 pypi_0 pypi cudatoolkit 9.0 1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cudnn 7.6.0 cuda9.0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main cycler 0.10.0 pypi_0 pypi easydict 1.8 pypi_0 pypi flask 1.0.2 pypi_0 pypi gast 0.2.2 pypi_0 pypi grpcio 1.19.0 pypi_0 pypi h5py 2.7.0 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free hdf5 1.8.15.1 2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free html5lib 0.9999999 pypi_0 pypi icc_rt 2019.0.0 h0cc432a_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main idna 2.7 pypi_0 pypi itsdangerous 1.1.0 pypi_0 pypi jinja2 2.10.1 pypi_0 pypi keras-applications 1.0.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main keras-preprocessing 1.1.0 py_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main libprotobuf 3.6.0 h1a1b453_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main markdown 3.1 pypi_0 pypi markupsafe 1.1.1 pypi_0 pypi matplotlib 2.1.0 pypi_0 pypi mkl 2017.0.3 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free mock 2.0.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free numpy 1.16.2 pypi_0 pypi olefile 0.46 pypi_0 pypi pbr 1.10.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free pillow 4.3.0 pypi_0 pypi pip 19.2.1 pypi_0 pypi pyparsing 2.4.0 pypi_0 pypi python 3.6.9 h5500b2f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main python-dateutil 2.8.0 pypi_0 pypi pytz 2018.9 pypi_0 pypi requests 2.19.1 pypi_0 pypi scipy 0.19.1 np113py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free setuptools 36.4.0 py36_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free six 1.12.0 pypi_0 pypi sqlite 3.29.0 he774522_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorboard 1.7.0 pypi_0 pypi tensorflow 1.12.0 gpu_py36ha5f9131_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-base 1.12.0 gpu_py36h6e53903_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main tensorflow-gpu 1.12.0 h0d30ee6_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main termcolor 1.1.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free urllib3 1.23 pypi_0 pypi vc 14.1 h0510ff6_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main vs2015_runtime 14.15.26706 h3a45250_4 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main werkzeug 0.15.2 pypi_0 pypi wheel 0.29.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free wincertstore 0.2 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free zlib 1.2.11 h62dcd97_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread. 我使用脚本重新训练样本居然可以了 难道是一批样本训练了 3000轮就重新生生一批 可能是这样 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

谢谢 我想我的问题已经解决了

bao17634 commented 4 years ago

我的训练集图片准确度都是接近为1了,但是验证集准确度却始终是0.5左右;麻烦问一下各位你们出现类似情况没有。

第3240次训练 >>> [训练集] 字符准确率为 0.99844 图片准确率为 0.99219 >>> loss 0.0003599718 [验证集] 字符准确率为 0.82200 图片准确率为 0.53000 >>> loss 0.0003599718 第3250次训练 >>>

Esme01 commented 4 years ago

我的是这样:

第3840次训练 >>> [训练集] 字符准确率为 0.99609 图片准确率为 0.98438 >>> loss 0.0233225022 [验证集] 字符准确率为 0.49500 图片准确率为 0.06000 >>> loss 0.0233225022 第3850次训练 >>> [训练集] 字符准确率为 1.00000 图片准确率为 1.00000 >>> loss 0.0228977930 [验证集] 字符准确率为 0.39000 图片准确率为 0.03000 >>> loss 0.0228977930 第3860次训练 >>> [训练集] 字符准确率为 1.00000 图片准确率为 1.00000 >>> loss 0.0211946052 [验证集] 字符准确率为 0.44500 图片准确率为 0.01000 >>> loss 0.0211946052 第3870次训练 >>> [训练集] 字符准确率为 1.00000 图片准确率为 1.00000 >>> loss 0.0216046404 [验证集] 字符准确率为 0.43500 图片准确率为 0.03000 >>> loss 0.0216046404

bao17634 commented 4 years ago

你样本多少

------------------ 原始邮件 ------------------ 发件人: "Esme01"<notifications@github.com>; 发送时间: 2019年12月18日(星期三) 上午10:15 收件人: "nickliqian/cnn_captcha"<cnn_captcha@noreply.github.com>; 抄送: "Encircle"<1763417523@qq.com>; "Comment"<comment@noreply.github.com>; 主题: Re: [nickliqian/cnn_captcha] 训练集识别率达到的1,但是验证集的识别率一直为0 (#63)

我的是这样:

第3840次训练 >>> [训练集] 字符准确率为 0.99609 图片准确率为 0.98438 >>> loss 0.0233225022 [验证集] 字符准确率为 0.49500 图片准确率为 0.06000 >>> loss 0.0233225022 第3850次训练 >>> [训练集] 字符准确率为 1.00000 图片准确率为 1.00000 >>> loss 0.0228977930 [验证集] 字符准确率为 0.39000 图片准确率为 0.03000 >>> loss 0.0228977930 第3860次训练 >>> [训练集] 字符准确率为 1.00000 图片准确率为 1.00000 >>> loss 0.0211946052 [验证集] 字符准确率为 0.44500 图片准确率为 0.01000 >>> loss 0.0211946052 第3870次训练 >>> [训练集] 字符准确率为 1.00000 图片准确率为 1.00000 >>> loss 0.0216046404 [验证集] 字符准确率为 0.43500 图片准确率为 0.03000 >>> loss 0.0216046404

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

Terry-mine commented 3 years ago

测试结果: 12/100 100个样本识别耗时6.816540002822876秒,准确率12.0% 删除 sample 文件夹 重新生产 训练1000 轮效果居然出其的好 如果是这个原因 这个问题应该可以关闭了

您好,我想训练一个6位数包含大小写字母以及数字的模型,样本数为2w张,但在训练了1w次之后,验证集的图片准确率仍然为0.01,我想知道该如何解决

Terry-mine commented 3 years ago

第10570次训练 >>> [训练集] 字符准确率为 0.98698 图片准确率为 0.92969 >>> loss 0.0230094753 [验证集] 字符准确率为 0.40667 图片准确率为 0.00000 >>> loss 0.0230094753