yalecyu / crnn.caffe

crnn.caffe
Other
76 stars 51 forks source link

accuracy is 0 #26

Open liuyiyiyiyi opened 5 years ago

liuyiyiyiyi commented 5 years ago

我生成了5w的样本,4w做训练,1w做测试,accuracy为0,可能的原因是什么呢,还有就是你的accuracy和ctc loss分别是什么呢

liuyiyiyiyi commented 5 years ago

I1222 07:07:39.157032 26829 sgd_solver.cpp:105] Iteration 19500, lr = 1e-05 I1222 07:11:35.553428 26829 solver.cpp:331] Iteration 19600, Testing net (#0) I1222 07:12:46.248581 26829 solver.cpp:398] Test net output #0: accuracy = 8e-05 I1222 07:12:46.248845 26829 solver.cpp:398] Test net output #1: ctc_loss = 11.8095 ( 1 = 11.8095 loss) I1222 07:12:48.490643 26829 solver.cpp:219] Iteration 19600 (0.323269 iter/s, 309.34s/100 iters), loss = -nan I1222 07:12:48.490779 26829 solver.cpp:238] Train net output #0: accuracy = 0 I1222 07:12:48.490815 26829 solver.cpp:238] Train net output #1: ctc_loss = 11.9346 ( 1 = 11.9346 loss) I1222 07:12:48.490839 26829 sgd_solver.cpp:105] Iteration 19600, lr = 1e-05 I1222 07:16:47.911034 26829 solver.cpp:219] Iteration 19700 (0.417668 iter/s, 239.425s/100 iters), loss = -nan I1222 07:16:47.911434 26829 solver.cpp:238] Train net output #0: accuracy = 0 I1222 07:16:47.911478 26829 solver.cpp:238] Train net output #1: ctc_loss = 11.7821 ( 1 = 11.7821 loss) I1222 07:16:47.911489 26829 sgd_solver.cpp:105] Iteration 19700, lr = 1e-05 I1222 07:20:46.812930 26829 solver.cpp:331] Iteration 19800, Testing net (#0) I1222 07:21:54.835827 26829 solver.cpp:398] Test net output #0: accuracy = 0.00012 I1222 07:21:54.839115 26829 solver.cpp:398] Test net output #1: ctc_loss = 11.8588 ( 1 = 11.8588 loss) I1222 07:21:57.172338 26829 solver.cpp:219] Iteration 19800 (0.323346 iter/s, 309.266s/100 iters), loss = -nan I1222 07:21:57.172468 26829 solver.cpp:238] Train net output #0: accuracy = 0.002 I1222 07:21:57.172489 26829 solver.cpp:238] Train net output #1: ctc_loss = 11.8786 ( 1 = 11.8786 loss) I1222 07:21:57.172499 26829 sgd_solver.cpp:105] Iteration 19800, lr = 1e-05 I1222 07:25:57.027230 26829 solver.cpp:219] Iteration 19900 (0.416913 iter/s, 239.858s/100 iters), loss = -nan I1222 07:25:57.027503 26829 solver.cpp:238] Train net output #0: accuracy = 0 I1222 07:25:57.027542 26829 solver.cpp:238] Train net output #1: ctc_loss = 11.7884 ( 1 = 11.7884 loss) I1222 07:25:57.027555 26829 sgd_solver.cpp:105] Iteration 19900, lr = 1e-05 I1222 07:29:54.699640 26829 solver.cpp:448] Snapshotting to binary proto file ./examples/crnn/model/crnn_captcha_iter_20000.caffemodel I1222 07:29:54.869462 26829 sgd_solver.cpp:273] Snapshotting solver state to binary proto file ./examples/crnn/model/crnn_captcha_iter_20000.solverstate I1222 07:29:56.457839 26829 solver.cpp:311] Iteration 20000, loss = -nan I1222 07:29:56.457885 26829 solver.cpp:331] Iteration 20000, Testing net (#0) I1222 07:31:05.673779 26829 solver.cpp:398] Test net output #0: accuracy = 8e-05 I1222 07:31:05.676453 26829 solver.cpp:398] Test net output #1: ctc_loss = 11.8025 (* 1 = 11.8025 loss)

yalecyu commented 5 years ago

学习率会不会太低?你的Loss没有下降下去?

liuyiyiyiyi commented 5 years ago

好的,我修改lr试一下

liuyiyiyiyi commented 5 years ago

你好,accuracy为0.99,但是测试结果均为- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

yalecyu commented 5 years ago

@liuyiyiyiyi 我已经更新了代码,是训练和测试数据数据分布不一致造成的。训练归一化了输入