Java实现深度神经网络：BatchNorm2d层

使用Java实现

/**
 * 归一化 对一张图的归一化
 * 
 * @author SXC 2020年9月11日 下午2:00:15
 */
public class BatchNorm2d {
	double weight = 0.5653, bias = 0;//默认为weight =1,bias=0

	public BatchNorm2d(int num_features) {//输入图片通道数，也为特征数，此处无用

	}

	public double mean(double[] x) {
		double meanvalue = 0;
		for (int i = 0; i < x.length; i++) {
			meanvalue += x[i];
		}
		return meanvalue / x.length;
	}

	public double var(double[] x) {
		double meanvalue = mean(x);
		double varvalue = 0;
		for (int i = 0; i < x.length; i++) {
			varvalue += (x[i] - meanvalue) * (x[i] - meanvalue);
		}
		return varvalue / x.length;
	}

	public void setBias(double bias) {
		this.bias = bias;
	}

	public void setWeight(double weight) {
		this.weight = weight;
	}

	public double getBias() {
		return bias;
	}

	public double getWeight() {
		return weight;
	}

	public double[] out(double[] x) {
		double[] out = new double[x.length];
		for (int i = 0; i < x.length; i++) {
			out[i] = (x[i] - mean(x)) / Math.pow(var(x) + 1e-5, 0.5) * weight + bias;
		}
		return out;
	}
}

测试代码

public class t {

	public static void main(String[] args) {
		BatchNorm2d BatchNorm2d=new BatchNorm2d(2);
		double[] x= {-2.4308, -1.0281, -1.1322,  0.9819,
        -0.4069,  0.7973,  1.6296,  1.6797,
         0.2802, -0.8285,  2.0101,  0.1286};
		double[] k=BatchNorm2d.out(x);
		for (int i = 0; i < k.length; i++) {
			System.out.println(k[i]);
		}
	}
}

运行结果

-1.1235822212153521
-0.5105423878128046
-0.5560385357112898
0.36791349380059857
-0.2390514456942742
0.28723540636486006
0.6509860681891563
0.6728819088204158
0.06124061214481498
-0.4233086594695034
0.8172808258776638
-0.005015065294285479

Pytorch实现1

import torch

import torch.nn as nn

m = nn.BatchNorm2d(2,affine=True) #权重w和偏重将被使用
input = torch.randn(1,2,3,4)
output = m(input)

print("输入图片：")
print(input)
print("归一化权重：")
print(m.weight)
print("归一化的偏重：")
print(m.bias)

print("归一化的输出：")
print(output)
print("输出的尺度：")
print(output.size())

# i = torch.randn(1,1,2)
print("输入的第一个维度：")
print(input[0][0])
firstDimenMean = torch.Tensor.mean(input[0][0])
firstDimenVar= torch.Tensor.var(input[0][0],False) #Bessel's Correction贝塞尔校正不会被使用

print(m.eps)
print("输入的第一个维度平均值：")
print(firstDimenMean)
print("输入的第一个维度方差：")
print(firstDimenVar)

bacthnormone = \
    ((input[0][0][0][0] - firstDimenMean)/(torch.pow(firstDimenVar+m.eps,0.5) ))\
               * m.weight[0] + m.bias[0]
print(bacthnormone)

参考自https://blog.youkuaiyun.com/tmk_01/article/details/80679549?utm_medium=distribute.pc_relevant.none-task-blog-title-3&spm=1001.2101.3001.4242
运行结果

输入图片：
tensor([[[[-2.4308, -1.0281, -1.1322,  0.9819],
          [-0.4069,  0.7973,  1.6296,  1.6797],
          [ 0.2802, -0.8285,  2.0101,  0.1286]],


         [[-0.5740,  0.1970, -0.7209, -0.7231],
          [-0.1489,  0.4993,  0.4159,  1.4238],
          [ 0.0334, -0.6333,  0.1308, -0.2180]]]])
归一化权重：
Parameter containing:
tensor([ 0.5653,  0.0322])
归一化的偏重：
Parameter containing:
tensor([ 0.,  0.])
归一化的输出：
tensor([[[[-1.1237, -0.5106, -0.5561,  0.3679],
          [-0.2391,  0.2873,  0.6510,  0.6729],
          [ 0.0612, -0.4233,  0.8173, -0.0050]],


         [[-0.0293,  0.0120, -0.0372, -0.0373],
          [-0.0066,  0.0282,  0.0237,  0.0777],
          [ 0.0032, -0.0325,  0.0084, -0.0103]]]])
输出的尺度：
torch.Size([1, 2, 3, 4])
输入的第一个维度：
tensor([[-2.4308, -1.0281, -1.1322,  0.9819],
        [-0.4069,  0.7973,  1.6296,  1.6797],
        [ 0.2802, -0.8285,  2.0101,  0.1286]])
1e-05
输入的第一个维度平均值：
tensor(0.1401)
输入的第一个维度方差：
tensor(1.6730)
tensor(-1.1237)

输出结果相同，但这里未弄清weight,bias的取值怎么来，只实现了BatchNorm2d (1)，num_features未使用到，后期需修改。

Pytorch实现2

n=nn.BatchNorm2d(num_features=1)
w=torch.tensor([0.5653],dtype=torch.float32)
print(n.weight)
n.weight=torch.nn.Parameter(w)
k=torch.tensor([[[[-2.4308, -1.0281, -1.1322,  0.9819],
          [-0.4069,  0.7973,  1.6296,  1.6797],
          [ 0.2802, -0.8285,  2.0101,  0.1286]]]])
print(n.weight)
y = n(k)
print(y)

运行结果

Parameter containing:
tensor([1.], requires_grad=True)
Parameter containing:
tensor([0.5653], requires_grad=True)
tensor([[[[-1.1236, -0.5105, -0.5560,  0.3679],
          [-0.2391,  0.2872,  0.6510,  0.6729],
          [ 0.0612, -0.4233,  0.8173, -0.0050]]]],
       grad_fn=<NativeBatchNormBackward>)

Process finished with exit code 0

原BatchNorm2d函数的weight的默认值为1，bias默认值为0。