对于神经网络,层的类型比较的多,合理的设计这些层的类型,对于任务准确度的提升都有很大的价值。
Vision layer, Common Layer:
InnerProduct, Convolution, Pooling, DropOut
Activation/Neron Layer:
ReLU, Sigmoid, TanH, AbsVal, Power, BNLL, Scale
Normalization Layer:
LRN, BatchNorm
Utility Layer:
Split, Concat, Slicing,Eltwise
Loss Layer:
Softmax, SoftmaxWithLoss, EuclideanLoss, HingeLoss, SigmoidCrossEntropyLoss, Accuracy
对于卷积和池化操作所对引得图片的大小有一定的规律:
conv_h_o = (h_i + 2*pad_h - kernel_h) / stride_h + 1
pool_h_o = ceil((h_i + 2*pad_h - kernel_h) / stride_h) + 1
数据输入 --- 卷积 -- 池化--- 输出