2017.10.26代码调试记录

昨天的结果中，最好的效果是3层全连接：Train Accuracy:  1.0，Test Accuracy:  0.765

---

4层卷积网络

隐藏层：2层卷积层+1层全连接层，输出层：Softmax分类器

K = 4 #hindden1的卷积核数目

L = 8 #隐藏层2的卷积核数

N = 200 #全连接层的节点数

使用了学习率衰减：

max_lr = 0.003

min_lr = 0.0001

decay_speed = 2000.0

未使用Droupout。pkeep:1.0

iter = 800

All cost time:  1519.3290100097656

Train Accuracy:  1.0

Test Accuracy:  0.805

使用卷机网络耗时明显增加

核数目稍微增多

K = 5 #hindden1的卷积核数目

L = 10

pkeep:1.0

终于，loss曲线有了很大的改观

核数目减少

K = 3 #hindden1的卷积核数目

L = 5

All cost time:  1459.0826830863953

Train Accuracy:  1.0

Test Accuracy:  0.84

K = 2 #hindden1的卷积核数目

L = 2

All cost time:  1211.8775815963745

Train Accuracy:  1.0

Test Accuracy:  0.905

第一次上90%！

将全连接层的节点数减少

K = 2 #hindden1的卷积核数目

L = 2 #隐藏层2的卷积核数

N = 100

All cost time:  1124.0736384391785

Train Accuracy:  1.0

Test Accuracy:  0.76

N=500

All cost time:  1369.4422793388367

Train Accuracy:  1.0

Test Accuracy:  0.765

N=180

All cost time:  1144.1208844184875

Train Accuracy:  1.0

Test Accuracy:  0.95

终于，这个95%还算不错了