PyTorch入门笔记

原教程

加载数据集

表格数据集CSV

此处使用NumPy来导入，除此之外还可以使用csv和Pandas导入
数据集链接

import csv
import numpy as np
wine_path = "../../data/chapter3/winequality-white.csv" # 路径
wineq_numpy = np.loadtxt(wine_path, dtype=np.float32, delimiter=";",
     skiprows=1)
wineq_numpy #输出演示

# 检查
col_list = next(csv.reader(open(wine_path), delimiter=';')) # 取文件第一行
wineq_numpy.shape, col_list # 检查是否读取全部数据

# 将numpy数组转化为torch张量
import torch
wineq = torch.from_numpy(wineq_numpy)
wineq.shape, wineq.type()

# 拆分data和targtet
data = wineq[:, :-1] # 除最后一列外所有列，索引用法见原教程2.5
data, data.shape
target = wineq[:, -1] # 最后一列
target, target.shape

# 标签转化为整数，一般标签有实际比较意义如评分时使用，适用于本数据集
target = wineq[:, -1].long()

# 标签独热编码，一般标签无实际意义，仅作分类时使用
target_onehot = torch.zeros(target.shape[0], 10)
target_onehot.scatter_(1, target.unsqueeze(1), 1.0) 
# 函数下划线结尾指就地修改张量，不返回新的张量
# unsqueeze(1)方法可提供一个额外的维度，具体见原文

# 获取每一列均值和标准差
data_mean = torch.mean(data, dim=0)
data_mean
data_var = torch.var(data, dim=0)
data_var

# 归一化
data_normalized = (data - data_mean) / torch.sqrt(data_var)
data_normalized

np.loadtxt()用法
next()用法
原文中有阈值预测的示例，此处不做介绍
文件式代码

'''
输出归一化属性数据和标签
'''

import csv
import torch as t
import numpy as np


DataPath = r'DataBase/winequality-white.csv'
IniDat_numpy = np.loadtxt(DataPath, dtype=np.float32, delimiter=';', skiprows=1)
col_list = next(csv.reader(open(DataPath), delimiter=';'))  # 每一列属性名称
IniDat = t.from_numpy(IniDat_numpy)
data = IniDat[:, :-1]  # 取最后一列外的张量做属性数据
target = IniDat[:, -1]  # 取最后一列做标签
target = target.long()  # 整数化标签
# target = t.zeros(target.shape[0], 10).scatter(1, target.unsqueeze(1), 1.0)  # 对标签进行one-hot编码
data_mean = t.mean(data, dim=0)  # 求均值
data_var = t.var(data, dim=0)  # 求方差
data_normalized = (data - data_mean) / t.sqrt(data_var)  # 归一化
print(data_normalized, '\n', target)

文本数据

文本数据：傲慢与偏见

图像数据