5. Convolutional neural networks (CNNs)
1. Fully connected layer(dense layer)
all the neurons are connected
2. Convolution layer
就是过滤器在前一层滑动,每一个位置就对应一个值(点乘)。过滤器的深度就等于前一层的深度。
output of convolution layer 叫做 activation map。注意:一个filter产生的输出就叫做一个activation map
一开始的卷积层学一些lower feature,靠后的卷积层学一些higher feature
过滤之后的长度计算公式: ((N - F) / stride + 1)
此处有几个超参数:
- stride(1 2)
- size of filter(3 5 7 …)
- number of filter(power of 2)
- amount of padding(sizeOfFilter // 2)
3. Pooling layer
makes the representation smaller and more manageable.
Pooling Operates over each activation map independently.
Pooling主要两种
- max
- avg
Pooling层的参数主要就是
- size of filter
- stride
在pooling层,size of filter = stride