dee1024 / pytorch-captcha-recognition Goto Github PK

View Code? Open in Web Editor NEW

1.1K 1.1K 311.0 660 KB

基于CNN训练的一套 "端到端" 的验证码识别模型，使用深度学习+训练数据+大量计算力，纯数字识别率高达 99.99%，数字+字母识别率 96%

License: Apache License 2.0

Python 100.00%

pytorch-captcha-recognition's People

Contributors

Stargazers

Watchers

Forkers

dee2048 81724097 kawayiyi newdimitri shi1992 eglrp startime-h iweisi lyonleelpl waitingfy bigff alanmorninglight hzl-totyu qsdj hwmt2015 kenvifire cheewing cpp597455873 wangxingyuaaa hlenter hayekzh putaodoudou zyjia batermj jym66 acosin achi-migo no-bb-just-do-it buptyuxin xueran1991 starstylesky dkgee niluanwudidadi mengpengfei lllfx msw1535540 cherryxuan weichen12345 longniao teliduxing004 enableasync law52525 dttworld juxj sofool zhwei820 michael-hsu ronggh wqvoon yanmovfer innovationb1ue jack19861225 iwannapython v3ucn srayaa elliot-bia zzuniverse efwfe charliex2 chinesehuang docksw godisaboy sakunoo robinhoooood chunchunc petergwok zenistzw tianqiraf asyncins f549263766 gechangli zhliying ipxplay hanchuntao qq2737499951 junming4 chocoai alex2andre jack061 orientswift henrylee123 xxx520 17637462979 huangkai31 xyzera caroid yiwenjie007 wangyao049 youthlab lynch88024 enjoyhomelife yungsh fakegit edmt niaoshuai chayao2015 kingking888 liuyibo2017 wec4dre96 alisaxxw

pytorch-captcha-recognition's Issues

one hot code 独热码代码优化

原来的独热码编码时位置计算太依赖数字，这样每次改动图片大小时都有可能对应改动编码算法。而且解码也必须依赖数字。从这两个角度优化，新的独热码代码如下：

import numpy


CHARACTER = {'0': 0, '1': 1, '2': 2, '3': 3, '4': 4, '5': 5, '6': 6, '7': 7, '8': 8, '9': 9, 
             'A': 10, 'B': 11, 'C': 12, 'D': 13, 'E': 14, 'F': 15, 'G': 16, 'H': 17, 'I': 18,
             'J': 19, 'K': 20, 'L': 21, 'M': 22, 'N': 23, 'O': 24, 'P': 25, 'Q': 26, 'R': 27,
             'S': 28, 'T': 29, 'U': 30, 'V': 31, 'W': 32, 'X': 33, 'Y': 34, 'Z': 35}

CAPTCHA_NUMBER = 6


def one_hot_encode(value: list) -> tuple:
    """编码，将字符转为独热码
    vector为独热码，order用于解码
    """
    order = []
    vector = numpy.zeros(CAPTCHA_NUMBER * len(CHARACTER ), dtype=float)
    for k, v in enumerate(value):
        index = k * len(CHARACTER) + CHARACTER.get(v)
        vector[index] = 1.0
        order.append(index)
    return vector, order


def one_hot_decode(value: list) -> str:
    """解码，将独热码转为字符
    """
    res = []
    for ik, iv in enumerate(value):
        val = iv - ik * len(CHARACTER) if ik else iv
        for k, v in CHARACTER.items():
            if val == int(v):
                res.append(k)
                break
    return "".join(res)


if __name__ == '__main__':
    code = '0A2JYD'
    vec, orders = one_hot_encode(code)
    print(orders)
    print(vec)
    print(one_hot_decode(orders))

运行后，输出如下：

[0, 46, 74, 127, 178, 193]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
0A2JYD

识别符号需要调整哪些模型参数

数字和字母识别都很精准但是加上了加减符号从新训练后识别数字没问题识别符号也会被认为是数字 setting中添加了符号样本也相应的标注了还需要调整model的参数么新手求问

run captcha_predict.py got an error after change the png photo in predict

Traceback (most recent call last):
File "/Users/james/pythonproject/pytorch-captcha-recognition/captcha_predict.py", line 34, in
main()
File "/Users/james/pythonproject/pytorch-captcha-recognition/captcha_predict.py", line 19, in main
for i, (images, labels) in enumerate(predict_dataloader):
File "/Users/james/pythonproject/airasiago/venv/lib/python2.7/site-packages/torch/utils/data/dataloader.py", line 345, in next
data = self._next_data()
File "/Users/james/pythonproject/airasiago/venv/lib/python2.7/site-packages/torch/utils/data/dataloader.py", line 385, in _next_data
data = self._dataset_fetcher.fetch(index) # may raise StopIteration
File "/Users/james/pythonproject/airasiago/venv/lib/python2.7/site-packages/torch/utils/data/utils/fetch.py", line 44, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/Users/james/pythonproject/pytorch-captcha-recognition/my_dataset.py", line 24, in getitem
label = ohe.encode(image_name.split('')[0]) # 为了方便，在生成图片的时候，图片文件的命名格式 "4个数字或者数字_时间戳.PNG", 4个字母或者即是图片的验证码的值，字母大写,同时对该值做 one-hot 处理
File "/Users/james/pythonproject/pytorch-captcha-recognition/one_hot_encoding.py", line 22, in encode
vector[idx] = 1.0
IndexError: index 152 is out of bounds for axis 0 with size 144

训练时loss在降低，但实际预测时基本全错！问题在哪？

训练1w张图片，loss率最后也在0.01，使用测试集一只显示0，请问大佬怎么回事？

修改版本：https://github.com/pprp/captcha_identify.torch

@dee1024 感谢您的工作，我研究了以下您的代码，针对您的库中的不足之处进行了部分改进和个性化处理。
这是链接：https://github.com/pprp/captcha_identify.torch
使用您的库：

training set : 10k
test set: 1k

达不到96%的要求，请教您具体是哪个地方有问题？

另外，我主要的改动的点在以下几个方面：

添加了更多torchvision中支持的模型
改了一下文件的名称
支持了GPU，当然cpu也可以
添加了以下功能：训练完成一个epoch之后进行测试，（需要保证test和train中的模型一致）
添加了以下功能：将每次得到的测试结果写入results.txt文件，运行torch_util.py得到results.png可视化准确率。
RES152为基础网络进行训练，混合数字和大写字符只能达到94%，还达不到原作者的识别率，希望得到赐教

目前最好的是RES101,RES152为基础的网络，可以达到94%。

loss 在降低，但是Test Accuracy of the model =0.00000%

pytorch-captcha-recognition git:(master) ✗ python captcha_train.py
init net
('epoch:', 0, 'step:', 0, 'loss:', 0.7110539674758911)
('epoch:', 1, 'step:', 0, 'loss:', 0.20382478833198547)
('epoch:', 2, 'step:', 0, 'loss:', 0.2277430146932602)
('epoch:', 3, 'step:', 0, 'loss:', 0.2373160570859909)
('epoch:', 4, 'step:', 0, 'loss:', 0.22053484618663788)
('epoch:', 5, 'step:', 0, 'loss:', 0.21534274518489838)
('epoch:', 6, 'step:', 0, 'loss:', 0.19580186903476715)
('epoch:', 7, 'step:', 0, 'loss:', 0.1724429875612259)
('epoch:', 8, 'step:', 0, 'loss:', 0.15848731994628906)
('epoch:', 9, 'step:', 0, 'loss:', 0.1449590027332306)
('epoch:', 10, 'step:', 0, 'loss:', 0.14423207938671112)
('epoch:', 11, 'step:', 0, 'loss:', 0.1371624767780304)
('epoch:', 12, 'step:', 0, 'loss:', 0.12939313054084778)
('epoch:', 13, 'step:', 0, 'loss:', 0.12587422132492065)
('epoch:', 14, 'step:', 0, 'loss:', 0.12168151885271072)
('epoch:', 15, 'step:', 0, 'loss:', 0.12061356008052826)
('epoch:', 16, 'step:', 0, 'loss:', 0.11823482811450958)
('epoch:', 17, 'step:', 0, 'loss:', 0.11660336703062057)
('epoch:', 18, 'step:', 0, 'loss:', 0.11383380740880966)
('epoch:', 19, 'step:', 0, 'loss:', 0.11137279868125916)
('epoch:', 20, 'step:', 0, 'loss:', 0.10716967284679413)
('epoch:', 21, 'step:', 0, 'loss:', 0.10241743177175522)
('epoch:', 22, 'step:', 0, 'loss:', 0.10174597799777985)
('epoch:', 23, 'step:', 0, 'loss:', 0.09972652792930603)
('epoch:', 24, 'step:', 0, 'loss:', 0.09792148321866989)
('epoch:', 25, 'step:', 0, 'loss:', 0.09143608063459396)
('epoch:', 26, 'step:', 0, 'loss:', 0.09200140088796616)
('epoch:', 27, 'step:', 0, 'loss:', 0.09041691571474075)
('epoch:', 28, 'step:', 0, 'loss:', 0.08658407628536224)
('epoch:', 29, 'step:', 0, 'loss:', 0.08212457597255707)
save last model
➜ pytorch-captcha-recognition git:(master) ✗ python captcha_test.py
load cnn net.
Test Accuracy of the model on the 3 test images: 0.000000 %

请问模型在哪里找，或是怎样训练

请问有训练好的模型吗？如果没有，能否请教一下怎样去训练呢...求教

loss在下降，单识别结果几乎全错

大概需要训练多少次？

我在训练时，控制台展示是这样，请问是参数设置有问题吗

epoch: 0 step: 9 loss: 0.18992629647254944
epoch: 0 step: 19 loss: 0.1413472294807434
epoch: 0 step: 29 loss: 0.13513770699501038
epoch: 0 step: 39 loss: 0.13097789883613586
epoch: 0 step: 49 loss: 0.13270880281925201
epoch: 0 step: 59 loss: 0.13205519318580627
epoch: 0 step: 69 loss: 0.13079266250133514
epoch: 0 step: 79 loss: 0.13159562647342682
epoch: 0 step: 89 loss: 0.13097992539405823
epoch: 0 step: 99 loss: 0.1281210333108902
save model
epoch: 0 step: 109 loss: 0.13041895627975464
epoch: 0 step: 119 loss: 0.12769392132759094
epoch: 0 step: 129 loss: 0.12638987600803375
epoch: 0 step: 139 loss: 0.126255601644516
epoch: 0 step: 149 loss: 0.12566044926643372
epoch: 0 step: 157 loss: 0.12796930968761444
epoch: 1 step: 9 loss: 0.12493212521076202
epoch: 1 step: 19 loss: 0.12400683015584946
epoch: 1 step: 29 loss: 0.12240520119667053
epoch: 1 step: 39 loss: 0.12404748797416687
epoch: 1 step: 49 loss: 0.11832032352685928
epoch: 1 step: 59 loss: 0.1218739002943039
epoch: 1 step: 69 loss: 0.11750295013189316
epoch: 1 step: 79 loss: 0.1162392869591713
epoch: 1 step: 89 loss: 0.1177394762635231
epoch: 1 step: 99 loss: 0.11434363573789597
save model
epoch: 1 step: 109 loss: 0.11179210245609283
epoch: 1 step: 119 loss: 0.11091677844524384
epoch: 1 step: 129 loss: 0.11243681609630585
epoch: 1 step: 139 loss: 0.11096695810556412
epoch: 1 step: 149 loss: 0.1121552512049675
epoch: 1 step: 157 loss: 0.10759076476097107
epoch: 2 step: 9 loss: 0.10809783637523651
epoch: 2 step: 19 loss: 0.10550016164779663
epoch: 2 step: 29 loss: 0.10933805257081985
epoch: 2 step: 39 loss: 0.1027115136384964
epoch: 2 step: 49 loss: 0.10497838258743286
epoch: 2 step: 59 loss: 0.10278748720884323
epoch: 2 step: 69 loss: 0.10579609125852585
epoch: 2 step: 79 loss: 0.10177668184041977
epoch: 2 step: 89 loss: 0.09883224219083786
epoch: 2 step: 99 loss: 0.10240185260772705
save model

识别率百分之四十

十万张测试图片，test的正确率在百分之四十左右，是不是过拟合了？如何修正参数

不同背景，不同字体得验证码都要单独训练吗

不同背景，不同字体得验证码都要单独训练吗，这个太麻烦了，什么情况下可以达到通用啊

模型训练好了，如何用一张图片去调用预测结果

实在不会用，用captcha_predict.py可以，但是我想在其它地方单独传入一张图片名称得到返回结果如何做呢？

预测的时候居然也带dropout？

我发现对同一张图多次预测的结果不同，看到模型里有用dropout，是不是预测的时候也有……

如何直接调用跑所有的图片的识别结果啊

对于部署pytorch的人一头蒙，说到的端到端呢

训练损失为0 测试损失很大

你好我在训练时损失为0 但是测试时损失很大准确率为 0

如果是未知字符个数的验证码，4-7个字符的，应该如何修改呀，固定4个5个6个7个的都能跑

可以用自己的6位验证码集训练吗

图片大小设置

RuntimeError: size mismatch, m1: [64 x 8960], m2: [1344 x 1024] at /pytorch/aten/src/TH/generic/THTensorMath.cpp:940
设置了自己定义的图片大小后出现不匹配的错误，请问如何解决？

训练的时候报错

File "C:\Users\leping.zhao\AppData\Local\Programs\Python\Python36\lib\site-pac
kages\torch\utils\data\dataloader.py", line 165, in default_collate
return torch.stack(batch, 0, out=out)
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimensio
n 0. Got 160 and 164 in dimension 3 at c:\new-builder_2\win-wheel\pytorch\aten\s
rc\th\generic/THTensorMath.cpp:3616

当验证码只要30个的时候没有报错，但是把数量改成1W或者5W生成验证后，然后训练的时候就报这个错误了，麻烦看下怎么改

模型成功率为0？

使用captcha_gen.py脚本生成了1万张图片。然后直接使用captcha_train.py脚本，参数是num_epochs = 30
batch_size = 100
learning_rate = 0.001。
脚本执行完毕后使用test脚本，但是模型成功率为0.
请问，需要训练多少次才能达到0.96的正确率。

效率问题

你这个是单线程的,效率问题太大了吧

Hello, can you help me with captcha_test?

When i run the test it says im missing a model.pkl file and i dont know where to find it or generate it.

关于验证码的应用

您好

请问有没有可能有其他合作的可能

我们想做关于一些银行登入验证码的解析

能否提供一个预训练模型

自己的训练效果很差，能否提供一个预训练模型？

QQ群不存在?

单张图片预测

当我想把模型部署到服务器上的时候，准备读取客户post上来的image文件类型，尝试使用

image_root = './dataset/predict/1129_1554564483.jpg'
image = Image.open(image_root)
image = image.resize((60, 160))
if transform is not None:
    image = transform(image)
        
vimage = Variable(image.cuda())
predict_label = cnn(vimage.cuda())
c0 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, 0:captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
c1 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, captcha_setting.ALL_CHAR_SET_LEN:2 * captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
c2 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, 2 * captcha_setting.ALL_CHAR_SET_LEN:3 * captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
c3 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, 3 * captcha_setting.ALL_CHAR_SET_LEN:4 * captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
    
predict_label = '%s%s%s%s'(c0,c1,c2,c3)
print(predict_label)

以上来处理我的图片，但遇到了输入维度不匹配的问题，然而我使用测试中的dataloader传入

for i, (images, labels) in enumerate(test_dataloader):
    image = images
    vimage = Variable(image.cuda())
    predict_label = cnn(vimage.cuda())

    c0 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, 0:captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
    c1 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, captcha_setting.ALL_CHAR_SET_LEN:2 * captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
    c2 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, 2 * captcha_setting.ALL_CHAR_SET_LEN:3 * captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
    c3 = captcha_setting.ALL_CHAR_SET[np.argmax(predict_label[0, 3 * captcha_setting.ALL_CHAR_SET_LEN:4 * captcha_setting.ALL_CHAR_SET_LEN].data.cpu().numpy())]
    predict_label = '%s%s%s%s' % (c0, c1, c2, c3)
    true_label = one_hot_encoding.decode(labels.numpy()[0])
    print("predict: %s true: %s"%(predict_label, true_label))

便不会发生任何问题，请问我该如何处理

新手，请问为何要对结果进行独热编码啊

如题

加群了

大佬, 麻烦通过一下, 我加群了

求助自定义验证码中的字符，无法训练

我想问下如果修改验证码位数为6，怎么修改参数呢

小白一枚，看见头晕。

gpu似乎没有被使用

训练时cpu占用100%而gpu占用仅1%

这是什么问题啊

Traceback (most recent call last):
File "/Users/haiyuan/Downloads/pytorch-captcha-recognition-master-2/captcha_train.py", line 43, in
main()
File "/Users/haiyuan/Downloads/pytorch-captcha-recognition-master-2/captcha_train.py", line 23, in main
for i, (images, labels) in enumerate(train_dataloader):
File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 346, in next
data = self.dataset_fetcher.fetch(index) # may raise StopIteration
File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 47, in fetch
return self.collate_fn(data)
File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/torch/utils/data/_utils/collate.py", line 80, in default_collate
return [default_collate(samples) for samples in transposed]
File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/torch/utils/data/_utils/collate.py", line 80, in
return [default_collate(samples) for samples in transposed]
File "/Library/Frameworks/Python.framework/Versions/3.7/lib/python3.7/site-packages/torch/utils/data/_utils/collate.py", line 56, in default_collate
return torch.stack(batch, 0, out=out)
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 160 and 164 in dimension 3 at ../aten/src/TH/generic/THTensor.cpp:689

为什么每次训练只能30张，请问怎么修改

init net
epoch: 0 step: 3 loss: 0.2350926697254181
epoch: 1 step: 3 loss: 0.18072116374969482
epoch: 2 step: 3 loss: 0.14990365505218506
epoch: 3 step: 3 loss: 0.1484384387731552
epoch: 4 step: 3 loss: 0.12944868206977844
epoch: 5 step: 3 loss: 0.136171355843544
epoch: 6 step: 3 loss: 0.13556182384490967
epoch: 7 step: 3 loss: 0.1339406818151474
epoch: 8 step: 3 loss: 0.12743189930915833
epoch: 9 step: 3 loss: 0.12321759760379791
epoch: 10 step: 3 loss: 0.11995768547058105
epoch: 11 step: 3 loss: 0.12512315809726715
epoch: 12 step: 3 loss: 0.11607347428798676
epoch: 13 step: 3 loss: 0.1194273829460144
epoch: 14 step: 3 loss: 0.11752907186746597
epoch: 15 step: 3 loss: 0.11002203077077866
epoch: 16 step: 3 loss: 0.11969199776649475
epoch: 17 step: 3 loss: 0.11192518472671509
epoch: 18 step: 3 loss: 0.10305792093276978
epoch: 19 step: 3 loss: 0.10709385573863983
epoch: 20 step: 3 loss: 0.10055689513683319
epoch: 21 step: 3 loss: 0.09390119463205338
epoch: 22 step: 3 loss: 0.10000652819871902
epoch: 23 step: 3 loss: 0.08290736377239227
epoch: 24 step: 3 loss: 0.08297861367464066
epoch: 25 step: 3 loss: 0.07633257657289505
epoch: 26 step: 3 loss: 0.08562453091144562
epoch: 27 step: 3 loss: 0.06919705867767334
epoch: 28 step: 3 loss: 0.06607756018638611
epoch: 29 step: 3 loss: 0.07599730044603348
save last model

直接就用cpu训练，把我的电脑cpu给爆了 >_<

为啥不默认GPU的