0
点赞
收藏
分享

微信扫一扫

数据预处理 ToTensor和Normalize

kiliwalk 2022-04-30 阅读 65

使用机器学习和深度学习对数据进行训练前,需要对数据进行预处理,本文记录下,与数据预处理相关过程。

数据预处理两个操作:

transform.ToTensor(),
transform.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))

关于对这两个的操作理解均来自文章参考

1、transform.ToTensor() 功能

  1. 转变一个PIL图片或者np.array转变成单精度的tensor
  2. 将输入的数据shape W,H,C ——> C,H,W
  3. 数据归一化,由[0,255] -> [0,1]
    注意:自定义图片数组,数据类型一定要转为‘uint8’,不然transforms.ToTensor()不会归一化。
x = np.array(x,dtype='uint8')
y = transforms.Totensor()(x)

示例实现:

import torch
import numpy as np
from torchvision import transforms
import cv2
#自定义图片数组,数据类型一定要转为‘uint8’,不然transforms.ToTensor()不会归一化
data = np.array([
                [[1,1,1],[1,1,1],[1,1,1],[1,1,1],[1,1,1]],
                [[2,2,2],[2,2,2],[2,2,2],[2,2,2],[2,2,2]],
                [[3,3,3],[3,3,3],[3,3,3],[3,3,3],[3,3,3]],
                [[4,4,4],[4,4,4],[4,4,4],[4,4,4],[4,4,4]],
                [[5,5,5],[5,5,5],[5,5,5],[5,5,5],[5,5,5]]
        ],dtype='uint8')
print(data)
print(data.shape)   #(5,5,3)
data = transforms.ToTensor()(data)
print(data)
print(data.shape)	#(3,5,5)

输出:
tensor([[[0.0039, 0.0039, 0.0039, 0.0039, 0.0039],
         [0.0078, 0.0078, 0.0078, 0.0078, 0.0078],
         [0.0118, 0.0118, 0.0118, 0.0118, 0.0118],
         [0.0157, 0.0157, 0.0157, 0.0157, 0.0157],
         [0.0196, 0.0196, 0.0196, 0.0196, 0.0196]],

        [[0.0039, 0.0039, 0.0039, 0.0039, 0.0039],
         [0.0078, 0.0078, 0.0078, 0.0078, 0.0078],
         [0.0118, 0.0118, 0.0118, 0.0118, 0.0118],
         [0.0157, 0.0157, 0.0157, 0.0157, 0.0157],
         [0.0196, 0.0196, 0.0196, 0.0196, 0.0196]],

        [[0.0039, 0.0039, 0.0039, 0.0039, 0.0039],
         [0.0078, 0.0078, 0.0078, 0.0078, 0.0078],
         [0.0118, 0.0118, 0.0118, 0.0118, 0.0118],
         [0.0157, 0.0157, 0.0157, 0.0157, 0.0157],
         [0.0196, 0.0196, 0.0196, 0.0196, 0.0196]]])

2、transforms.Normalize() 功能
计算同一维度数据的平均值和标准差,将该维度的每个值减去平均值再除以标准差。使每一维数据符合标准正态分布,即均值为0,标准差为1,并非[-1,1],使模型更容易收敛。
x = (x - mean) / std
计算一批数据的mean和std:

import torch
import numpy as np
from torchvision import transforms

# 这里以上述创建的单数据为例子
data = np.array([
                [[1,1,1],[1,1,1],[1,1,1],[1,1,1],[1,1,1]],
                [[2,2,2],[2,2,2],[2,2,2],[2,2,2],[2,2,2]],
                [[3,3,3],[3,3,3],[3,3,3],[3,3,3],[3,3,3]],
                [[4,4,4],[4,4,4],[4,4,4],[4,4,4],[4,4,4]],
                [[5,5,5],[5,5,5],[5,5,5],[5,5,5],[5,5,5]]
        ],dtype='uint8')

#将数据转为C,W,H,并归一化到[0,1]
data = transforms.ToTensor()(data)
# 需要对数据进行扩维,增加batch维度
data = torch.unsqueeze(data,0)

nb_samples = 0.
#创建3维的空列表
channel_mean = torch.zeros(3)
channel_std = torch.zeros(3)
print(data.shape)
N, C, H, W = data.shape[:4]
data = data.view(N, C, -1)     #将w,h维度的数据展平,为batch,channel,data,然后对三个维度上的数分别求和和标准差
print(data.shape)
#展平后,w,h属于第二维度,对他们求平均,sum(0)为将同一纬度的数据累加
channel_mean += data.mean(2).sum(0)  
#展平后,w,h属于第二维度,对他们求标准差,sum(0)为将同一纬度的数据累加
channel_std += data.std(2).sum(0)
#获取所有batch的数据,这里为1
nb_samples += N
#获取同一batch的均值和标准差
channel_mean /= nb_samples
channel_std /= nb_samples
print(channel_mean, channel_std)

得到均值和标准差之后自己实现标准化:

data = np.array([
                [[1,1,1],[1,1,1],[1,1,1],[1,1,1],[1,1,1]],
                [[2,2,2],[2,2,2],[2,2,2],[2,2,2],[2,2,2]],
                [[3,3,3],[3,3,3],[3,3,3],[3,3,3],[3,3,3]],
                [[4,4,4],[4,4,4],[4,4,4],[4,4,4],[4,4,4]],
                [[5,5,5],[5,5,5],[5,5,5],[5,5,5],[5,5,5]]
        ],dtype='uint8')
data = transforms.ToTensor()(data)
for i in range(3):
    data[i,:,:] = (data[i,:,:] - channel_mean[i]) / channel_std[i]
print(data)

输出:
tensor([[[-1.3856, -1.3856, -1.3856, -1.3856, -1.3856],
         [-0.6928, -0.6928, -0.6928, -0.6928, -0.6928],
         [ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000],
         [ 0.6928,  0.6928,  0.6928,  0.6928,  0.6928],
         [ 1.3856,  1.3856,  1.3856,  1.3856,  1.3856]],

        [[-1.3856, -1.3856, -1.3856, -1.3856, -1.3856],
         [-0.6928, -0.6928, -0.6928, -0.6928, -0.6928],
         [ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000],
         [ 0.6928,  0.6928,  0.6928,  0.6928,  0.6928],
         [ 1.3856,  1.3856,  1.3856,  1.3856,  1.3856]],

        [[-1.3856, -1.3856, -1.3856, -1.3856, -1.3856],
         [-0.6928, -0.6928, -0.6928, -0.6928, -0.6928],
         [ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000],
         [ 0.6928,  0.6928,  0.6928,  0.6928,  0.6928],
         [ 1.3856,  1.3856,  1.3856,  1.3856,  1.3856]]])

官方方法实现:

data = np.array([
                [[1,1,1],[1,1,1],[1,1,1],[1,1,1],[1,1,1]],
                [[2,2,2],[2,2,2],[2,2,2],[2,2,2],[2,2,2]],
                [[3,3,3],[3,3,3],[3,3,3],[3,3,3],[3,3,3]],
                [[4,4,4],[4,4,4],[4,4,4],[4,4,4],[4,4,4]],
                [[5,5,5],[5,5,5],[5,5,5],[5,5,5],[5,5,5]]
        ],dtype='uint8')
data = transforms.ToTensor()(data)
data = transforms.Normalize(channel_mean, channel_std)(data)
print(data)
举报

相关推荐

0 条评论