图片引于它出网络参数及连接计算
Reducing overfitting
4.1 Data Augmentation
- image translations(平移)and horizontal reflections (水平翻转)
- extract randomly 224 * 224 patches(截图)from 256 * 256 images
in this case, data size grow by 2048 (=2*(256 - 224) * (256 - 224) )倍 - perform PCA on the set of RGB, 并对主成分进行标准差为0.1的高斯扰动, increase noise and reduce error rate
PCA(降维)详解 - at test time, extract five 224 * 224 patches (one from center and other four from corner) as well as horizontal reflection, hence 共提取10次
4.2 Dropout - dropout rate is 0.5
Hyperparameters
- 批量大小 mini batchsize = 128
- 权重衰减 weight decay =0.0005
- 学习率 learning rate = 0.01 衰减率为0.1
- 轮数 epoches = 90
- initialization of weights : zero-mean Gaussian distribustion with standard deviation 0.01 ( 均值为0 方差为1 的高斯分布)
- initialization of neuron biases is constant 1 in 2,4,5 convolutional layers , the remaining layers is constant 0
知识补充
steps 迭代次数 - 学习完所有数据所需要的次数
epoch 轮数 - 所有数据学习一遍为一个epoch
Feature of network
- Using ReLU
- Dropout
- max-pool layer
- LRN layer(but not widely use)
- some tircks about data augmentation