关于模型结构最后两层的BN层

[bolun365]

    convRelu(5)
    cnn.add_module('pooling{0}'.format(3),
                   nn.MaxPool2d((2, 2), (2, 1), (0, 1)))  # 512x2x16
    convRelu(6, True)  # 512x1x16

发现代码的BN设置与原论文不同, 貌似这样效果确实更好些, 请问这样做的原因是?

[Zhang-O]

It is true. After I compared the code and the original paper , I found the difference lied in the location and numbers of bn layers.