请问，您的人脸旋转对齐归一化是怎么做的？

The training pictures of CASIA-Webface are gray or color ?

Hi, @Tonyfy

Thanks for your cool code and it saved my life in replicating the work of Lightened CNN.

I am a newbie in this field. Some problems make me confused when I tried to evaluated the trained caffemodel. :(

The script lfw.py generates the prediction and saved it to predict.txt. (Am I right?)

In your evaluation code here, it seems that some file's missing. Could you kindly provide the suggested align_face.py?

Thanks in advance.

Cheers.

您好，请问LCNN可以对人脸数据库进行对齐处理吗？代码中的ReadMe中的内容有些看的不是很明白，如果您能写的更详细一点，或者写一些比较详细的使用步骤的博客，对我们这种小白将会有很大的帮助，感谢

请问accurate是您提供的测试代码吗？怎么加载caffemodel？如果不是的话，可以提供一下吗，挺急的，谢谢

你好 tony，我在训练的时候加上-gpu all 想用我这8个GPU一起训练 可是结果就报错 不支持多GPU 请问下这怎么解决

hi. @Tonyfy
i am studying about lighted cnn.

i want see deploy.prototxt
Do you support deploy.prototxt??

thank you.

你好，我用opencv 3.3的dnn模块，读取prototxt和caffemodel，结果报错了。

您好，我手上有CASIA数据集，我想问下你的训练图片和测试图片都是从CASIA里面拿的吗，这样会不会影响测试的结果？

我的test accuracy到80%的时候，LFW的准确率就已经有95%了。

I0914 09:33:36.987599 6584 sgd_solver.cpp:106] Iteration 1851200, lr = 0.000174488
I0914 09:33:43.461599 6584 solver.cpp:228] Iteration 1851300, loss = 0.183091
I0914 09:33:43.461599 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:33:43.461599 6584 sgd_solver.cpp:106] Iteration 1851300, lr = 0.000174482
I0914 09:33:49.933599 6584 solver.cpp:228] Iteration 1851400, loss = 0.606006
I0914 09:33:49.933599 6584 solver.cpp:244] Train net output #0: accuracy = 0.8125
I0914 09:33:49.933599 6584 sgd_solver.cpp:106] Iteration 1851400, lr = 0.000174476
I0914 09:33:56.406599 6584 solver.cpp:228] Iteration 1851500, loss = 0.571156
I0914 09:33:56.406599 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:33:56.406599 6584 sgd_solver.cpp:106] Iteration 1851500, lr = 0.000174469
I0914 09:34:02.905799 6584 solver.cpp:228] Iteration 1851600, loss = 0.418948
I0914 09:34:02.905799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:34:02.905799 6584 sgd_solver.cpp:106] Iteration 1851600, lr = 0.000174463
I0914 09:34:09.381799 6584 solver.cpp:228] Iteration 1851700, loss = 0.0690969
I0914 09:34:09.381799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:34:09.381799 6584 sgd_solver.cpp:106] Iteration 1851700, lr = 0.000174456
I0914 09:34:15.836799 6584 solver.cpp:228] Iteration 1851800, loss = 0.274711
I0914 09:34:15.836799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:34:15.836799 6584 sgd_solver.cpp:106] Iteration 1851800, lr = 0.00017445
I0914 09:34:22.301800 6584 solver.cpp:228] Iteration 1851900, loss = 0.556112
I0914 09:34:22.301800 6584 solver.cpp:244] Train net output #0: accuracy = 0.75
I0914 09:34:22.301800 6584 sgd_solver.cpp:106] Iteration 1851900, lr = 0.000174444
I0914 09:34:28.700799 6584 solver.cpp:337] Iteration 1852000, Testing net (#0)
I0914 09:34:29.266799 6584 solver.cpp:391] Test loss: 3.77606
I0914 09:34:29.266799 6584 solver.cpp:404] Test net output #0: accuracy = 0.4675
I0914 09:34:29.294800 6584 solver.cpp:228] Iteration 1852000, loss = 0.070474
I0914 09:34:29.294800 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:34:29.294800 6584 sgd_solver.cpp:106] Iteration 1852000, lr = 0.000174437
I0914 09:34:35.765799 6584 solver.cpp:228] Iteration 1852100, loss = 0.226924
I0914 09:34:35.765799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:34:35.765799 6584 sgd_solver.cpp:106] Iteration 1852100, lr = 0.000174431
I0914 09:34:42.221799 6584 solver.cpp:228] Iteration 1852200, loss = 0.648706
I0914 09:34:42.221799 6584 solver.cpp:244] Train net output #0: accuracy = 0.78125
I0914 09:34:42.221799 6584 sgd_solver.cpp:106] Iteration 1852200, lr = 0.000174425
I0914 09:34:48.687799 6584 solver.cpp:228] Iteration 1852300, loss = 0.377885
I0914 09:34:48.687799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:34:48.687799 6584 sgd_solver.cpp:106] Iteration 1852300, lr = 0.000174418
I0914 09:34:55.157799 6584 solver.cpp:228] Iteration 1852400, loss = 0.530284
I0914 09:34:55.157799 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:34:55.157799 6584 sgd_solver.cpp:106] Iteration 1852400, lr = 0.000174412
I0914 09:35:01.607800 6584 solver.cpp:228] Iteration 1852500, loss = 0.257938
I0914 09:35:01.607800 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:35:01.607800 6584 sgd_solver.cpp:106] Iteration 1852500, lr = 0.000174405
I0914 09:35:08.076799 6584 solver.cpp:228] Iteration 1852600, loss = 0.27629
I0914 09:35:08.076799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:35:08.076799 6584 sgd_solver.cpp:106] Iteration 1852600, lr = 0.000174399
I0914 09:35:14.526799 6584 solver.cpp:228] Iteration 1852700, loss = 0.328374
I0914 09:35:14.526799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:35:14.526799 6584 sgd_solver.cpp:106] Iteration 1852700, lr = 0.000174393
I0914 09:35:20.991799 6584 solver.cpp:228] Iteration 1852800, loss = 0.405679
I0914 09:35:20.991799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:35:20.991799 6584 sgd_solver.cpp:106] Iteration 1852800, lr = 0.000174386
I0914 09:35:27.446799 6584 solver.cpp:228] Iteration 1852900, loss = 0.544654
I0914 09:35:27.446799 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:35:27.446799 6584 sgd_solver.cpp:106] Iteration 1852900, lr = 0.00017438
I0914 09:35:33.837800 6584 solver.cpp:337] Iteration 1853000, Testing net (#0)
I0914 09:35:34.394799 6584 solver.cpp:391] Test loss: 3.56771
I0914 09:35:34.394799 6584 solver.cpp:404] Test net output #0: accuracy = 0.495
I0914 09:35:34.420799 6584 solver.cpp:228] Iteration 1853000, loss = 0.216011
I0914 09:35:34.420799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:35:34.420799 6584 sgd_solver.cpp:106] Iteration 1853000, lr = 0.000174374
I0914 09:35:40.890799 6584 solver.cpp:228] Iteration 1853100, loss = 0.318283
I0914 09:35:40.890799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:35:40.890799 6584 sgd_solver.cpp:106] Iteration 1853100, lr = 0.000174367
I0914 09:35:47.363800 6584 solver.cpp:228] Iteration 1853200, loss = 0.332405
I0914 09:35:47.363800 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:35:47.363800 6584 sgd_solver.cpp:106] Iteration 1853200, lr = 0.000174361
I0914 09:35:53.817800 6584 solver.cpp:228] Iteration 1853300, loss = 0.436057
I0914 09:35:53.817800 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:35:53.817800 6584 sgd_solver.cpp:106] Iteration 1853300, lr = 0.000174355
I0914 09:36:00.295799 6584 solver.cpp:228] Iteration 1853400, loss = 0.667568
I0914 09:36:00.295799 6584 solver.cpp:244] Train net output #0: accuracy = 0.78125
I0914 09:36:00.295799 6584 sgd_solver.cpp:106] Iteration 1853400, lr = 0.000174348
I0914 09:36:06.755800 6584 solver.cpp:228] Iteration 1853500, loss = 0.609652
I0914 09:36:06.755800 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:36:06.755800 6584 sgd_solver.cpp:106] Iteration 1853500, lr = 0.000174342
I0914 09:36:13.220799 6584 solver.cpp:228] Iteration 1853600, loss = 0.749501
I0914 09:36:13.220799 6584 solver.cpp:244] Train net output #0: accuracy = 0.75
I0914 09:36:13.220799 6584 sgd_solver.cpp:106] Iteration 1853600, lr = 0.000174335
I0914 09:36:19.695799 6584 solver.cpp:228] Iteration 1853700, loss = 0.424909
I0914 09:36:19.695799 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:36:19.695799 6584 sgd_solver.cpp:106] Iteration 1853700, lr = 0.000174329
I0914 09:36:26.165799 6584 solver.cpp:228] Iteration 1853800, loss = 0.249959
I0914 09:36:26.165799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:36:26.165799 6584 sgd_solver.cpp:106] Iteration 1853800, lr = 0.000174323
I0914 09:36:32.630800 6584 solver.cpp:228] Iteration 1853900, loss = 0.721237
I0914 09:36:32.630800 6584 solver.cpp:244] Train net output #0: accuracy = 0.8125
I0914 09:36:32.630800 6584 sgd_solver.cpp:106] Iteration 1853900, lr = 0.000174316
I0914 09:36:39.029799 6584 solver.cpp:337] Iteration 1854000, Testing net (#0)
I0914 09:36:39.582799 6584 solver.cpp:391] Test loss: 3.77665
I0914 09:36:39.582799 6584 solver.cpp:404] Test net output #0: accuracy = 0.475
I0914 09:36:39.608799 6584 solver.cpp:228] Iteration 1854000, loss = 0.263935
I0914 09:36:39.608799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:36:39.608799 6584 sgd_solver.cpp:106] Iteration 1854000, lr = 0.00017431
I0914 09:36:46.059799 6584 solver.cpp:228] Iteration 1854100, loss = 0.150226
I0914 09:36:46.059799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:36:46.059799 6584 sgd_solver.cpp:106] Iteration 1854100, lr = 0.000174304
I0914 09:36:52.532799 6584 solver.cpp:228] Iteration 1854200, loss = 0.254497
I0914 09:36:52.532799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:36:52.532799 6584 sgd_solver.cpp:106] Iteration 1854200, lr = 0.000174297
I0914 09:36:58.983799 6584 solver.cpp:228] Iteration 1854300, loss = 0.240074
I0914 09:36:58.983799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:36:58.983799 6584 sgd_solver.cpp:106] Iteration 1854300, lr = 0.000174291
I0914 09:37:05.448799 6584 solver.cpp:228] Iteration 1854400, loss = 0.18402
I0914 09:37:05.448799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:37:05.448799 6584 sgd_solver.cpp:106] Iteration 1854400, lr = 0.000174285
I0914 09:37:11.923799 6584 solver.cpp:228] Iteration 1854500, loss = 0.585598
I0914 09:37:11.923799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:37:11.923799 6584 sgd_solver.cpp:106] Iteration 1854500, lr = 0.000174278
I0914 09:37:18.374799 6584 solver.cpp:228] Iteration 1854600, loss = 0.412825
I0914 09:37:18.374799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:37:18.374799 6584 sgd_solver.cpp:106] Iteration 1854600, lr = 0.000174272
I0914 09:37:24.826799 6584 solver.cpp:228] Iteration 1854700, loss = 0.324418
I0914 09:37:24.826799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:37:24.826799 6584 sgd_solver.cpp:106] Iteration 1854700, lr = 0.000174265
I0914 09:37:31.305799 6584 solver.cpp:228] Iteration 1854800, loss = 0.616231
I0914 09:37:31.305799 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:37:31.305799 6584 sgd_solver.cpp:106] Iteration 1854800, lr = 0.000174259
I0914 09:37:37.772799 6584 solver.cpp:228] Iteration 1854900, loss = 0.485702
I0914 09:37:37.773799 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:37:37.773799 6584 sgd_solver.cpp:106] Iteration 1854900, lr = 0.000174253
I0914 09:37:44.170799 6584 solver.cpp:337] Iteration 1855000, Testing net (#0)
I0914 09:37:44.726799 6584 solver.cpp:391] Test loss: 4.11952
I0914 09:37:44.726799 6584 solver.cpp:404] Test net output #0: accuracy = 0.475
I0914 09:37:44.751799 6584 solver.cpp:228] Iteration 1855000, loss = 0.106197
I0914 09:37:44.751799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:37:44.752799 6584 sgd_solver.cpp:106] Iteration 1855000, lr = 0.000174246
I0914 09:37:51.207799 6584 solver.cpp:228] Iteration 1855100, loss = 0.103739
I0914 09:37:51.207799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:37:51.207799 6584 sgd_solver.cpp:106] Iteration 1855100, lr = 0.00017424
I0914 09:37:57.676800 6584 solver.cpp:228] Iteration 1855200, loss = 0.678847
I0914 09:37:57.676800 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:37:57.676800 6584 sgd_solver.cpp:106] Iteration 1855200, lr = 0.000174234
I0914 09:38:04.143800 6584 solver.cpp:228] Iteration 1855300, loss = 0.241907
I0914 09:38:04.143800 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:38:04.143800 6584 sgd_solver.cpp:106] Iteration 1855300, lr = 0.000174227
I0914 09:38:10.615799 6584 solver.cpp:228] Iteration 1855400, loss = 0.528937
I0914 09:38:10.615799 6584 solver.cpp:244] Train net output #0: accuracy = 0.8125
I0914 09:38:10.615799 6584 sgd_solver.cpp:106] Iteration 1855400, lr = 0.000174221
I0914 09:38:17.080799 6584 solver.cpp:228] Iteration 1855500, loss = 0.739377
I0914 09:38:17.081799 6584 solver.cpp:244] Train net output #0: accuracy = 0.75
I0914 09:38:17.081799 6584 sgd_solver.cpp:106] Iteration 1855500, lr = 0.000174215
I0914 09:38:23.549799 6584 solver.cpp:228] Iteration 1855600, loss = 0.544462
I0914 09:38:23.549799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:38:23.549799 6584 sgd_solver.cpp:106] Iteration 1855600, lr = 0.000174208
I0914 09:38:30.013799 6584 solver.cpp:228] Iteration 1855700, loss = 0.38861
I0914 09:38:30.013799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:38:30.013799 6584 sgd_solver.cpp:106] Iteration 1855700, lr = 0.000174202
I0914 09:38:36.480799 6584 solver.cpp:228] Iteration 1855800, loss = 0.261983
I0914 09:38:36.480799 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:38:36.480799 6584 sgd_solver.cpp:106] Iteration 1855800, lr = 0.000174195
I0914 09:38:42.947799 6584 solver.cpp:228] Iteration 1855900, loss = 0.697443
I0914 09:38:42.947799 6584 solver.cpp:244] Train net output #0: accuracy = 0.8125
I0914 09:38:42.947799 6584 sgd_solver.cpp:106] Iteration 1855900, lr = 0.000174189
I0914 09:38:49.350800 6584 solver.cpp:337] Iteration 1856000, Testing net (#0)
I0914 09:38:49.901799 6584 solver.cpp:391] Test loss: 4.44912
I0914 09:38:49.901799 6584 solver.cpp:404] Test net output #0: accuracy = 0.4525
I0914 09:38:49.928799 6584 solver.cpp:228] Iteration 1856000, loss = 0.167307
I0914 09:38:49.928799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:38:49.928799 6584 sgd_solver.cpp:106] Iteration 1856000, lr = 0.000174183
I0914 09:38:56.401799 6584 solver.cpp:228] Iteration 1856100, loss = 0.439052
I0914 09:38:56.401799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:38:56.401799 6584 sgd_solver.cpp:106] Iteration 1856100, lr = 0.000174176
I0914 09:39:02.875799 6584 solver.cpp:228] Iteration 1856200, loss = 0.285193
I0914 09:39:02.875799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:39:02.875799 6584 sgd_solver.cpp:106] Iteration 1856200, lr = 0.00017417
I0914 09:39:09.344799 6584 solver.cpp:228] Iteration 1856300, loss = 0.359473
I0914 09:39:09.344799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:39:09.344799 6584 sgd_solver.cpp:106] Iteration 1856300, lr = 0.000174164
I0914 09:39:15.805799 6584 solver.cpp:228] Iteration 1856400, loss = 0.800733
I0914 09:39:15.805799 6584 solver.cpp:244] Train net output #0: accuracy = 0.8125
I0914 09:39:15.805799 6584 sgd_solver.cpp:106] Iteration 1856400, lr = 0.000174157
I0914 09:39:22.270799 6584 solver.cpp:228] Iteration 1856500, loss = 0.302216
I0914 09:39:22.270799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:39:22.270799 6584 sgd_solver.cpp:106] Iteration 1856500, lr = 0.000174151
I0914 09:39:28.735800 6584 solver.cpp:228] Iteration 1856600, loss = 0.154521
I0914 09:39:28.735800 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:39:28.735800 6584 sgd_solver.cpp:106] Iteration 1856600, lr = 0.000174145
I0914 09:39:35.197799 6584 solver.cpp:228] Iteration 1856700, loss = 0.298379
I0914 09:39:35.197799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:39:35.197799 6584 sgd_solver.cpp:106] Iteration 1856700, lr = 0.000174138
I0914 09:39:41.658799 6584 solver.cpp:228] Iteration 1856800, loss = 0.166663
I0914 09:39:41.658799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:39:41.658799 6584 sgd_solver.cpp:106] Iteration 1856800, lr = 0.000174132
I0914 09:39:48.146800 6584 solver.cpp:228] Iteration 1856900, loss = 0.557385
I0914 09:39:48.146800 6584 solver.cpp:244] Train net output #0: accuracy = 0.78125
I0914 09:39:48.146800 6584 sgd_solver.cpp:106] Iteration 1856900, lr = 0.000174126
I0914 09:39:54.572799 6584 solver.cpp:337] Iteration 1857000, Testing net (#0)
I0914 09:39:55.127799 6584 solver.cpp:391] Test loss: 3.7937
I0914 09:39:55.127799 6584 solver.cpp:404] Test net output #0: accuracy = 0.4975
I0914 09:39:55.153800 6584 solver.cpp:228] Iteration 1857000, loss = 0.199672
I0914 09:39:55.153800 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:39:55.153800 6584 sgd_solver.cpp:106] Iteration 1857000, lr = 0.000174119
I0914 09:40:01.610800 6584 solver.cpp:228] Iteration 1857100, loss = 0.240449
I0914 09:40:01.610800 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:40:01.610800 6584 sgd_solver.cpp:106] Iteration 1857100, lr = 0.000174113
I0914 09:40:08.078799 6584 solver.cpp:228] Iteration 1857200, loss = 0.421281
I0914 09:40:08.078799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:40:08.078799 6584 sgd_solver.cpp:106] Iteration 1857200, lr = 0.000174107
I0914 09:40:14.538800 6584 solver.cpp:228] Iteration 1857300, loss = 0.238799
I0914 09:40:14.538800 6584 solver.cpp:244] Train net output #0: accuracy = 0.9375
I0914 09:40:14.538800 6584 sgd_solver.cpp:106] Iteration 1857300, lr = 0.0001741
I0914 09:40:21.005800 6584 solver.cpp:228] Iteration 1857400, loss = 0.228657
I0914 09:40:21.005800 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:40:21.005800 6584 sgd_solver.cpp:106] Iteration 1857400, lr = 0.000174094
I0914 09:40:27.474799 6584 solver.cpp:228] Iteration 1857500, loss = 0.113201
I0914 09:40:27.474799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:40:27.474799 6584 sgd_solver.cpp:106] Iteration 1857500, lr = 0.000174088
I0914 09:40:33.143999 6584 solver.cpp:228] Iteration 1857600, loss = 0.909531
I0914 09:40:33.143999 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:40:33.143999 6584 sgd_solver.cpp:106] Iteration 1857600, lr = 0.000174081
I0914 09:40:39.610999 6584 solver.cpp:228] Iteration 1857700, loss = 0.380397
I0914 09:40:39.610999 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:40:39.610999 6584 sgd_solver.cpp:106] Iteration 1857700, lr = 0.000174075
I0914 09:40:46.124399 6584 solver.cpp:228] Iteration 1857800, loss = 0.316336
I0914 09:40:46.124399 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:40:46.124399 6584 sgd_solver.cpp:106] Iteration 1857800, lr = 0.000174068
I0914 09:40:52.606400 6584 solver.cpp:228] Iteration 1857900, loss = 0.309551
I0914 09:40:52.606400 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:40:52.606400 6584 sgd_solver.cpp:106] Iteration 1857900, lr = 0.000174062
I0914 09:40:59.008399 6584 solver.cpp:337] Iteration 1858000, Testing net (#0)
I0914 09:40:59.561399 6584 solver.cpp:391] Test loss: 3.97371
I0914 09:40:59.561399 6584 solver.cpp:404] Test net output #0: accuracy = 0.4875
I0914 09:40:59.584399 6584 solver.cpp:228] Iteration 1858000, loss = 0.382728
I0914 09:40:59.584399 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:40:59.584399 6584 sgd_solver.cpp:106] Iteration 1858000, lr = 0.000174056
I0914 09:41:06.047399 6584 solver.cpp:228] Iteration 1858100, loss = 0.548754
I0914 09:41:06.047399 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:41:06.047399 6584 sgd_solver.cpp:106] Iteration 1858100, lr = 0.000174049
I0914 09:41:12.507400 6584 solver.cpp:228] Iteration 1858200, loss = 0.478362
I0914 09:41:12.507400 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:41:12.507400 6584 sgd_solver.cpp:106] Iteration 1858200, lr = 0.000174043
I0914 09:41:18.969399 6584 solver.cpp:228] Iteration 1858300, loss = 0.332062
I0914 09:41:18.969399 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:41:18.969399 6584 sgd_solver.cpp:106] Iteration 1858300, lr = 0.000174037
I0914 09:41:25.435400 6584 solver.cpp:228] Iteration 1858400, loss = 0.587138
I0914 09:41:25.436399 6584 solver.cpp:244] Train net output #0: accuracy = 0.8125
I0914 09:41:25.436399 6584 sgd_solver.cpp:106] Iteration 1858400, lr = 0.00017403
I0914 09:41:31.901399 6584 solver.cpp:228] Iteration 1858500, loss = 0.498611
I0914 09:41:31.901399 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:41:31.901399 6584 sgd_solver.cpp:106] Iteration 1858500, lr = 0.000174024
I0914 09:41:38.362399 6584 solver.cpp:228] Iteration 1858600, loss = 0.496264
I0914 09:41:38.362399 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:41:38.362399 6584 sgd_solver.cpp:106] Iteration 1858600, lr = 0.000174018
I0914 09:41:44.857800 6584 solver.cpp:228] Iteration 1858700, loss = 0.509912
I0914 09:41:44.857800 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:41:44.857800 6584 sgd_solver.cpp:106] Iteration 1858700, lr = 0.000174011
I0914 09:41:51.321799 6584 solver.cpp:228] Iteration 1858800, loss = 0.180461
I0914 09:41:51.321799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:41:51.321799 6584 sgd_solver.cpp:106] Iteration 1858800, lr = 0.000174005
I0914 09:41:57.799799 6584 solver.cpp:228] Iteration 1858900, loss = 0.65346
I0914 09:41:57.799799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:41:57.800799 6584 sgd_solver.cpp:106] Iteration 1858900, lr = 0.000173999
I0914 09:42:04.218799 6584 solver.cpp:337] Iteration 1859000, Testing net (#0)
I0914 09:42:04.775799 6584 solver.cpp:391] Test loss: 3.73394
I0914 09:42:04.776799 6584 solver.cpp:404] Test net output #0: accuracy = 0.485
I0914 09:42:04.802799 6584 solver.cpp:228] Iteration 1859000, loss = 0.743077
I0914 09:42:04.802799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:42:04.802799 6584 sgd_solver.cpp:106] Iteration 1859000, lr = 0.000173992
I0914 09:42:11.266799 6584 solver.cpp:228] Iteration 1859100, loss = 0.386253
I0914 09:42:11.266799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:42:11.266799 6584 sgd_solver.cpp:106] Iteration 1859100, lr = 0.000173986
I0914 09:42:17.729799 6584 solver.cpp:228] Iteration 1859200, loss = 0.853737
I0914 09:42:17.729799 6584 solver.cpp:244] Train net output #0: accuracy = 0.78125
I0914 09:42:17.729799 6584 sgd_solver.cpp:106] Iteration 1859200, lr = 0.00017398
I0914 09:42:24.178799 6584 solver.cpp:228] Iteration 1859300, loss = 0.547904
I0914 09:42:24.178799 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:42:24.178799 6584 sgd_solver.cpp:106] Iteration 1859300, lr = 0.000173973
I0914 09:42:30.638799 6584 solver.cpp:228] Iteration 1859400, loss = 0.492802
I0914 09:42:30.638799 6584 solver.cpp:244] Train net output #0: accuracy = 0.875
I0914 09:42:30.638799 6584 sgd_solver.cpp:106] Iteration 1859400, lr = 0.000173967
I0914 09:42:37.101799 6584 solver.cpp:228] Iteration 1859500, loss = 0.27371
I0914 09:42:37.101799 6584 solver.cpp:244] Train net output #0: accuracy = 0.96875
I0914 09:42:37.101799 6584 sgd_solver.cpp:106] Iteration 1859500, lr = 0.000173961
I0914 09:42:43.605799 6584 solver.cpp:228] Iteration 1859600, loss = 0.171972
I0914 09:42:43.605799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:42:43.605799 6584 sgd_solver.cpp:106] Iteration 1859600, lr = 0.000173954
I0914 09:42:50.068799 6584 solver.cpp:228] Iteration 1859700, loss = 0.533729
I0914 09:42:50.068799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:42:50.068799 6584 sgd_solver.cpp:106] Iteration 1859700, lr = 0.000173948
I0914 09:42:56.544800 6584 solver.cpp:228] Iteration 1859800, loss = 0.354224
I0914 09:42:56.544800 6584 solver.cpp:244] Train net output #0: accuracy = 0.84375
I0914 09:42:56.544800 6584 sgd_solver.cpp:106] Iteration 1859800, lr = 0.000173942
I0914 09:43:03.003799 6584 solver.cpp:228] Iteration 1859900, loss = 0.262504
I0914 09:43:03.003799 6584 solver.cpp:244] Train net output #0: accuracy = 0.90625
I0914 09:43:03.003799 6584 sgd_solver.cpp:106] Iteration 1859900, lr = 0.000173935

请问确定标点位置的code_face脚本在哪？您一般是用什么方法确定标点位置bbox.txt的呢？

Tony大神 我随意拍了些照片没有经过处理 扔到demo.py尝试对比 发现精度特别差 ，我想问下 是不是要对输入的人脸图片经过些处理

Hi：
I want to use classification 100 people,like ImageNet result,one people one class to realize face recogition?(I am not sure use this idea to realize face recogiton,I just want to convert this model to coreML format ,so I can identify person on ios ？)
I use fine tune mode，dataset are 100 people，test choose 20% of that，after 300 iterations,I got new model with arruray 1,and I use python classification code to test people in data set.but i got the wrong result.here is the code for classification .

-- coding: UTF-8 --

import caffe
import numpy as np
import cv2
def test(my_project_root, deploy_proto):
caffe_model = my_project_root + 'examples/digits_light_cnn/my.caffemodel' #caffe_model文件的位置
img = my_project_root + 'examples/digits_light_cnn/face_1.jpg' #随机找的一张待测图片
labels_filename = my_project_root + 'examples/digits_light_cnn/labels.txt' #类别名称文件,将数字标签转换回类别名称
caffe.set_mode_cpu()
net = caffe.Net(deploy_proto, caffe_model, caffe.TEST) #加载model和deploy
#图片预处理设置
print net.blobs['data'].data.shape
input = cv2.imread(img,0)
input = cv2.resize(input,(128,128),interpolation=cv2.INTER_CUBIC) #we just need to resize the face to (128,128)
img_blobinp = input[np.newaxis, np.newaxis, :, :]/255.0 #divide 255.0 ,make input is between 0-1
net.blobs['data'].reshape(*img_blobinp.shape)
net.blobs['data'].data[...] = img_blobinp
net.blobs['data'].data.shape
out = net.forward() #执行测试
labels = np.loadtxt(labels_filename, str, delimiter='\t') #读取类别名称文件
prob = net.blobs['fc2'].data[0].flatten() #取出最后一层（Softmax）属于某个类别的概率值
order = prob.argsort()[-1] #将概率值排序，取出最大值所在的序号
print '图片数字为:',labels[order] #将该序号转换成对应的类别名称，并打印
if name == 'main':
my_project_root = "/home/huang/caffe/caffe/" #my-caffe-project目录
deploy_proto = my_project_root + "examples/digits_light_cnn/deploy.prototxt" #保存deploy.prototxt文件的位置
test(my_project_root, deploy_proto)

tony大神，请问下有没有提供TestNet.exe的源码我想学习一下

您好，我用http://ethereon.github.io/netscope/#/gist/c64671dc9359cc82c8baa02b1e1e5ec5， 后，显示Can't infer network data shapes. Can't infer output shape of the 'slice1' layer of type 'Slice'. Unsupported layer type: 'Slice'. 请问您那能绘制完整的网络结构吗

您好。
我看了accurate/目录下的lfw-roc.py和绘制曲线的m文件。
我不是很清楚您在roc.txt中第2第3列记录的 负样本拒识率 和 正样本通过率 是不是 FNR 和 TPR ？
绘制ROC曲线需要的是TPR和FPR。
我想请教一下，TPR和FPR如何用代码实现？
换句话说就是您提供的那张ROC曲线是如何得到的？
谢谢！

with open(roc_file, 'w') as f:
            for i in th:
                TP = 0
                TN = 0
                FP = 0
                FN = 0
                for m in posi_simi_metric:
                    if float(m) > float(i):
                        TP = TP + 1
                    else:
                        FN = FN + 1
                for n in nega_simi_metric:
                    if float(n) < float(i):
                        TN = TN + 1
                    else:
                        FP = FP + 1
                f.write(str(i) + ' ' + str((TP) / (0.000001 + len(posi_simi_metric))) +
                        ' ' + str((TN) / (0.000001 + len(nega_simi_metric))) +
                        ' ' + str((FP) / (0.000001 + len(posi_simi_metric))) +
                        ' ' + str((FN) / (0.000001 + len(nega_simi_metric))) + ' ' + '\n')

我这样修改对吗？？？

那个testNet.exe 我这无法运行

I have my own experiments with the data and apply to your model(*.prototxt).
Composition of the difference is that dataset. (caffeType).
I made a lmdb using the caffe tool(using ~/caffe/build/tools/convert_imageset.bin).
first! made two file(train.txt, val.txt).
train.txt is format
[imagepath][blank][label]
example..
./0/0_20.jpg 0
./0/0_21.jpg 0
./0/0_22.jpg 0
./0/0_23.jpg 0
./0/0_24.jpg 0
./0/0_25.jpg 0
./0/0_26.jpg 0
./0/0_27.jpg 0
./0/0_28.jpg 0
...
./4900/4900_29.jpg 4900

val.txt is same.

and. image is grayscale and 144x144 size

finally i make train_lmdb/val_lmdb using tools'convert_imageset.bin

and.
I was created to the learning model by using train_lmdb/val_lmdb(adding LCNN_train_test.prototxt).
Using a model constructed softmax saw the results(about new image query).
But the results are very disappointed.

Originally, softmax ( fc2) results are bad is it?
or measurement distance(fc1) do you have to do(verification problem)?
or composition of the data wrong?

please, need your opinion.
I've had a few times did it, but it didn't get good results.

thank you.

When i used the trained model. it appears

F0914 14:49:38.433799 1816 net.cpp:773] Cannot copy param 0 weights from layer 'conv1'; shape mismatch. Source param shape is 96 3 5 5 (7200); target param shape is 96 1 5 5 (2400). To learn this layer's parameters from scratch rather than copying from a saved net, rename the layer.
*** Check failure stack trace: ***

博主，我刚接触这东西，现在不会运行，代码里面code_point这个路径里没有processCASIA这个文件夹请问如何解决，还想请教下博主，你的代码怎么实现检测两张图片是否同一个人QAQ

您好，我想问下，我按照您上面的步骤整理好数据，然后用您的solver和train_net进行训练，但是500万次后LFW测试准确率和您公布的相差比较大，请问您了解会是什么原因？

已经把自己的图片转成了灰度图但是在用您的模型fine-tune时，载入模型还是出现了这个错误：
Source param s hape is 96 1 5 5 (2400); target param shape is 96 3 5 5 (7200)
请问这是怎么回事呢

用作者提供的model进行微调，用100个人已经经过工具进行人脸对齐处理，迭代大概300次后精度为1，但是测试分类时一个都对不上，请问是哪里出问题了。
测试代码：

-- coding: UTF-8 --

import caffe
import numpy as np
import cv2
def test(my_project_root, deploy_proto):
caffe_model = my_project_root + 'examples/digits_light_cnn/my.caffemodel' #caffe_model文件的位置
img = my_project_root + 'examples/digits_light_cnn/face_1.jpg' #随机找的一张待测图片
labels_filename = my_project_root + 'examples/digits_light_cnn/labels.txt' #类别名称文件,将数字标签转换回类别名称
caffe.set_mode_cpu()
net = caffe.Net(deploy_proto, caffe_model, caffe.TEST) #加载model和deploy

#图片预处理设置
print net.blobs['data'].data.shape
input = cv2.imread(img,0)
input = cv2.resize(input,(128,128),interpolation=cv2.INTER_CUBIC)   #we just need to resize the face to (128,128) 
img_blobinp = input[np.newaxis, np.newaxis, :, :]/255.0    #divide 255.0 ,make input is between 0-1
net.blobs['data'].reshape(*img_blobinp.shape)
net.blobs['data'].data[...] = img_blobinp
net.blobs['data'].data.shape

out = net.forward()                                                    #执行测试

labels = np.loadtxt(labels_filename, str, delimiter='\t')           #读取类别名称文件
prob = net.blobs['fc2'].data[0].flatten()                             #取出最后一层（Softmax）属于某个类别的概率值
order = prob.argsort()[-1]                                          #将概率值排序，取出最大值所在的序号
print '图片数字为:',labels[order]                                   #将该序号转换成对应的类别名称，并打印

if name == 'main':
my_project_root = "/home/huang/caffe/caffe/" #my-caffe-project目录
deploy_proto = my_project_root + "examples/digits_light_cnn/deploy.prototxt" #保存deploy.prototxt文件的位置
test(my_project_root, deploy_proto)

I'm training Light cnn now,my solver is:
base_lr: 0.001
momentum: 0.9
weight_decay: 0.0005
lr_policy: "step"
stepsize:500000
gamma:0.457305051927326

display: 100
max_iter: 4000000
snapshot: 40000
snapshot_prefix: "DeepFace_set003_net"

solver_mode: GPU

debug_info: false

clip_gradients: 150

I'm training light cnn with the clean list, after screening to get 79056 categories, about 4,920,000 images.But run 400,000 iteration loss is also 11.2?,Is this normal?do you have some advice?

I have another question in the layer of fc2,you have no param in this layer.

param {
lr_mult: 10
decay_mult: 1
}
param {
lr_mult: 20
decay_mult: 0
}

对人脸旋转剪裁是如何做的呢？是工程中的哪个文件呢？

bbox.txt中自由第一张图片的标点信息，我想得到webface中所有人的人脸标点信息怎么做。 tony,我用其他的crop工具 在人脸检测 的过程中49万张人脸检测加crop完只剩42万张了，请教下大神是怎么解决的

能不能提供下lfw测试部分的代码，形成完整的闭环？复现得到的准确率

Hi,thanks for your codes. When I run the addLabeltopic.py,I encounter the following problems and don't know how to solver this error,can you help me? thank you very much!
Traceback (most recent call last):
File "./tools/addLabeltopic.py", line 14, in
pic=os.listdir(folder)
OSError: [Errno 2] No such file or directory: '/home/dl/webface\0031162'

您好，我有两个小问题想请教您。

1. 运行lfw.py时的lfw-deepfunneled是否需要提前裁剪旋转处理好？
2. 运行时在26行：assert(len(pairs) == 6000) 报错AssertionError
   错误信息如下：

INFO:root:Namespace(l2=False, lfw_feature='/home/leo/Documents/light_cnn/LCNN_TRAIN-master/lfw-deepfunneled', pairs='/home/leo/Documents/light_cnn/LCNN_TRAIN-master/pairs.txt', predict_file='/home/leo/Documents/light_cnn/LCNN_TRAIN-master/predict.txt', suffix='data')
Loading embeddings done
INFO:root:begin generate the predict.txt.
...Reading pairs.
Traceback (most recent call last):
File "lfw.py", line 154, in
main()
File "lfw.py", line 150, in main
get_predict_file(args)
File "lfw.py", line 100, in get_predict_file
pairs = load_pairs(args.pairs)
File "lfw.py", line 26, in load_pairs
assert(len(pairs) == 6000)
AssertionError

您好，请问数据采取了哪些预处理的方式？，是否归一化？输入图像的范围是0-255， 0-1， 还是-1-1，谢谢。

请问LCNN.caffemodel是您已经训好的吧？那均值文件能否也提供一下？

The dataset contains many face images that is very difficult to be detected. Did you use your own face detection tool or some open-source?

您好！
我在运行lfw.py的时候。
报错：IOError: [Errno 2] No such file or directory: '/media/data/liuhan/light_cnn/LCNN_TRAIN-master/accurate/lfw-deepfunneled-aligned/Abel_Pacheco/Abel_Pacheco_0001.data'
这是不是说明参数'--lfw-feature'这个文件夹里存的是图片的特征数据？
该怎么获得.data文件？
谢谢！

您好，我是在您训练好的模型上微调自己的数据。我先用了AR人脸数据库进行微调，发现效果很好，准确率挺高。但我用自己采集的数据去微调，结果不收敛，效果很差，loss，accuracy来回震荡。这是我的数据。为了简单的测试，我训练集2个人，每个人100张。虽然数据少，但我觉得应该收敛才对。您觉得是哪里的问题，跟采集设备有没有关系呢？还是参数设置的问题？

RT