文字识别系统LeNet-5
下面,有必要来解释下上面这个用于文字识别的LeNet-5深层卷积网络。
1. 输入图像是32x32的大小,局部滑动窗的大小是5x5的,由于不考虑对图像的边界进行拓展,则滑动窗将有28x28个不同的位置,也就是C1层的大小是28x28。这里设定有6个不同的C1层,每一个C1层内的权值是相同的。
2. S2层是一个下采样层。简单的说,由4个点下采样为1个点,也就是4个数的加权平均。但在LeNet-5系统,下采样层比较复杂,因为这4个加权系数也需要学习得到,这显然增加了模型的复杂度。在斯坦福关于深度学习的教程中,这个过程叫做Pool。
源码:Convolutional Neural Networks (LeNet)
可能用到的源码:logistic_sgd,mlp
结果:
...
epoch 180, minibatch 100/100, validation error 0.920000 %
training @ iter = 18000
epoch 181, minibatch 100/100, validation error 0.920000 %
training @ iter = 18100
epoch 182, minibatch 100/100, validation error 0.920000 %
training @ iter = 18200
epoch 183, minibatch 100/100, validation error 0.910000 %
epoch 183, minibatch 100/100, test error of best model 0.920000 %
training @ iter = 18300
epoch 184, minibatch 100/100, validation error 0.910000 %
training @ iter = 18400
epoch 185, minibatch 100/100, validation error 0.910000 %
training @ iter = 18500
epoch 186, minibatch 100/100, validation error 0.910000 %
training @ iter = 18600
epoch 187, minibatch 100/100, validation error 0.910000 %
training @ iter = 18700
epoch 188, minibatch 100/100, validation error 0.910000 %
training @ iter = 18800
epoch 189, minibatch 100/100, validation error 0.910000 %
training @ iter = 18900
epoch 190, minibatch 100/100, validation error 0.910000 %
training @ iter = 19000
epoch 191, minibatch 100/100, validation error 0.910000 %
training @ iter = 19100
epoch 192, minibatch 100/100, validation error 0.910000 %
training @ iter = 19200
epoch 193, minibatch 100/100, validation error 0.910000 %
training @ iter = 19300
epoch 194, minibatch 100/100, validation error 0.910000 %
training @ iter = 19400
epoch 195, minibatch 100/100, validation error 0.910000 %
training @ iter = 19500
epoch 196, minibatch 100/100, validation error 0.910000 %
training @ iter = 19600
epoch 197, minibatch 100/100, validation error 0.910000 %
training @ iter = 19700
epoch 198, minibatch 100/100, validation error 0.910000 %
training @ iter = 19800
epoch 199, minibatch 100/100, validation error 0.910000 %
training @ iter = 19900
epoch 200, minibatch 100/100, validation error 0.910000 %
Optimization complete.
Best validation score of 0.910000 % obtained at iteration 18300, with test performance 0.920000 %
The code for file convolutional_mlp.py ran for 527.38m