DNN 的奇怪分类输出答案

【问题标题】：Strange classification output with DNNDNN 的奇怪分类输出
【发布时间】：2017-06-05 21:54:29
【问题描述】：

我尝试使用微调网络 CaffeNet 对图像进行分类。我按照 Caffe 的教程将训练文件中的输出数量更改为 3，同时我关闭了前两个卷积层的学习。出于某种原因，当我将分类器与经过训练的模型一起使用时，测试集中每个图像的所有类都得到 0.3。

number of classes: 3
train set size: 6570 images (80%)
test set size: 1645 images (20%)

求解器：

net: "train.prototxt"
test_iter: 100
test_interval: 1000
base_lr: 0.0001
lr_policy: "step"
gamma: 0.1
stepsize: 20000
display: 200
max_iter: 60000
momentum: 0.9
weight_decay: 0.0005
snapshot: 10000
snapshot_prefix: "snapshot"
solver_mode: GPU

我如何进行训练：

caffe train -solver solver.prototxt -weights bvlc_reference_caffenet.caffemodel

一些输出：

I0531 00:35:52.622647  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:36:02.699782  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:36:03.900009  5528 solver.cpp:218] Iteration 3600 (10.1266 iter/s, 19.7499s/200 iters), loss = 0.679402
I0531 00:36:03.900009  5528 solver.cpp:237]     Train net output #0: loss = 0.679402 (* 1 = 0.679402 loss)
I0531 00:36:03.900009  5528 sgd_solver.cpp:105] Iteration 3600, lr = 0.0001

I0531 00:41:20.139937  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:41:30.934025  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:41:34.199774  5528 solver.cpp:218] Iteration 6800 (9.66881 iter/s, 20.6851s/200 iters), loss = 0.451174
I0531 00:41:34.199774  5528 solver.cpp:237]     Train net output #0: loss = 0.451174 (* 1 = 0.451174 loss)
I0531 00:41:34.199774  5528 sgd_solver.cpp:105] Iteration 6800, lr = 0.0001

I0531 00:41:41.794001  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:41:52.743448  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:41:55.126147  5528 solver.cpp:330] Iteration 7000, Testing net (#0)
I0531 00:41:55.891929  3704 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:41:58.393698  3704 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:41:58.862452  5528 solver.cpp:397]     Test net output #0: accuracy = 0.6952
I0531 00:41:58.862452  5528 solver.cpp:397]     Test net output #1: loss = 0.873388 (* 1 = 0.873388 loss)

I0531 00:43:08.320360  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:43:18.514559  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:43:18.920881  5528 solver.cpp:218] Iteration 7800 (10.0073 iter/s, 19.9854s/200 iters), loss = 0.196175
I0531 00:43:18.920881  5528 solver.cpp:237]     Train net output #0: loss = 0.196175 (* 1 = 0.196175 loss)
I0531 00:43:18.920881  5528 sgd_solver.cpp:105] Iteration 7800, lr = 0.0001
I0531 00:43:28.660408  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:43:38.561293  5528 solver.cpp:330] Iteration 8000, Testing net (#0)
I0531 00:43:40.405230  3704 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:43:42.077230  5528 solver.cpp:397]     Test net output #0: accuracy = 0.7004
I0531 00:43:42.077230  5528 solver.cpp:397]     Test net output #1: loss = 0.991567 (* 1 = 0.991567 loss)

I0531 00:45:22.426592  3704 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:45:24.761165  3704 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:45:25.329238  5528 solver.cpp:397]     Test net output #0: accuracy = 0.6856
I0531 00:45:25.329238  5528 solver.cpp:397]     Test net output #1: loss = 1.08582 (* 1 = 1.08582 loss)
I0531 00:45:25.394567  5528 solver.cpp:218] Iteration 9000 (8.39955 iter/s, 23.8108s/200 iters), loss = 0.107816
I0531 00:45:25.394567  5528 solver.cpp:237]     Train net output #0: loss = 0.107816 (* 1 = 0.107816 loss)

I0531 00:46:49.099460  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:46:59.269830  2944 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:47:03.997443  5528 solver.cpp:447] Snapshotting to binary proto file snapshot_iter_10000.caffemodel
I0531 00:47:05.185039  5528 sgd_solver.cpp:273] Snapshotting solver state to binary proto file snapshot_iter_10000.solverstate
I0531 00:47:05.403774  5528 solver.cpp:330] Iteration 10000, Testing net (#0)
I0531 00:47:07.122831  3704 data_layer.cpp:73] Restarting data prefetching from start.
I0531 00:47:08.870923  5528 solver.cpp:397]     Test net output #0: accuracy = 0.7012
I0531 00:47:08.870923  5528 solver.cpp:397]     Test net output #1: loss = 1.18649 (* 1 = 1.18649 loss)
I0531 00:47:08.964664  5528 solver.cpp:218] Iteration 10000 (8.12416 iter/s, 24.6179s/200 iters), loss = 0.0347012
I0531 00:47:08.964664  5528 solver.cpp:237]     Train net output #0: loss = 0.0347012 (* 1 = 0.0347012 loss)
I0531 00:47:08.964664  5528 sgd_solver.cpp:105] Iteration 10000, lr = 0.0001

我如何运行分类：

classification deploy.prototxt snapshot_iter_10000.caffemodel labels.txt ..\test

一些输出：

"0.jpg",0.333333,0.333333,0.333333
"1.jpg",0.333333,0.333333,0.333333
"10.jpg",0.333333,0.333333,0.333333
"100.jpg",0.333333,0.333333,0.333333
"101.jpg",0.333333,0.333333,0.333333
"102.jpg",0.333333,0.333333,0.333333,
"103.jpg",0.333333,0.333333,0.333333

由于某种原因，准确率达到 70%，我得到的结果与 50% 相同 - 每个班级都有 0.3。

【问题讨论】：

标签： machine-learning neural-network deep-learning caffe

【解决方案1】：

您的分类输出没有什么奇怪的，只是您需要正确解释它。 0.333for 3 个类的准确度仅仅意味着你的网络没有学习任何东西——它是随机猜测。对于N 类，随机猜测将为您提供1/N 的准确度。所以在你的情况下，它是 1/3，即 0.333。

现在，设置超参数没有这样的标准规则，但考虑到损失的巨大变化，我建议你将基础学习率降低到 0.00001。

【讨论】：

我没有得到你的最后一行For some reason with 70% accuracy I'm getting same result as with 50% - every class has 0.3. 你在谈论训练准确性吗？请详细说明。
是的，我指的是训练准确度。您可以看到，对于第 9000 次迭代，我有 70% 的准确率。
这只是意味着您的网络过度拟合。您需要在网络中添加 dropout 层。