[人工智能]Pytorch实战：基于pytorch预测文章阅读量

介绍

这里用简单的网络来预测CSDN的阅读量。

我希望训练后的模型，给它前7天的阅读量，让它预测出第八天的阅读量。

阅读量的数据（很少）采用CSDN提供的excel文件：

这里采用xrld工具包来读取文件。

def get_readings():
    book = xlrd.open_workbook('./data/article_readings.xlsx')
    sheet1 = book.sheet_by_name('sheet1')  # excel中一个文件有多个表格，因为数据量小，所以就一个表格
    cols = sheet1.col_values(1)  # 获取第一列的数据，也就是阅读量
    readings = []
    for i in range(8, len(cols)):  # 要8天的数据，7天作为输入，第8天作为损失函数的比较参数 
        pre_days = []
        for j in range(i - 7, i):
            pre_days.append(int(cols[j]))
        pre_days.append(int(cols[i]))
        readings.append(pre_days)
    return readings  # 每一列都是这样的：[34, 10, 62, 50, 7, 10, 7, 48]

训练的xlsx文件：

阿里云盘

链接：https://www.aliyundrive.com/s/eQcRpbuWakp

百度网盘

链接：https://pan.baidu.com/s/1Icz5QOtFWe58pX37XhcPsw
提取码：91dt

效果

loss的下降趋势：

环境

pytorch=1.10.1

xlrd=1.2.0

步骤

1.准备数据

其中目录结构为：
所有代码都在train.py中

def get_readings():
    book = xlrd.open_workbook('./data/article_readings.xlsx')
    sheet1 = book.sheet_by_name('sheet1')
    cols = sheet1.col_values(1)
    readings = []
    for i in range(8, len(cols)):
        pre_days = []
        for j in range(i - 7, i):
            pre_days.append(int(cols[j]))
        pre_days.append(int(cols[i]))
        readings.append(pre_days)
    return readings

2.定义网络

class MyNet(nn.Module):
    def __init__(self):
        super(MyNet, self).__init__()
        self.fc = nn.Sequential(
            nn.Linear(7, 16, bias=True),
            nn.ReLU(),
            nn.Linear(16, 32, bias=True),
            nn.ReLU(),
            nn.Linear(32, 1, bias=True)
        )
        self.optimizer = torch.optim.Adam(self.fc.parameters(), lr=LR)  # LR是在外部定义的，也可以作为参数传入网络
        self.loss_fn = nn.MSELoss()

    def forward(self, input):
        return self.fc(input)

3.定义训练方法

def train(model: MyNet, readings: list[[int, int]]):
    loss_list = []  # matplotlib作图用，记录每一次的损失值 
    for reading in readings:
        predict = torch.tensor(reading[:7], dtype=torch.float)  # reading[:7] 索引0~6的是前7天的数据，因为是列表，所以不用resize
        predict = model(predict)

        target = reading[7]
        target = torch.tensor(target, dtype=torch.float)
        target = torch.unsqueeze(target, 0)  # 在target的第一维增加一个维度，原本的size是(1, )，增加后变为(1, 1)

        loss = model.loss_fn(target, predict)  # 计算损失值，损失函数用的是均方误差
        loss_list.append(loss.item())

        model.optimizer.zero_grad()
        loss.backward()
        model.optimizer.step()  # 更新参数

    return loss_list

4.预测

def predict(model: MyNet, pre_reading) -> int:
    pre_reading = torch.tensor(pre_reading, dtype=torch.float)
    pre_reading
    out = model(pre_reading).item()
    return int(out)

train.py

__main__中epoch是要自己设定的，如果你有大量的数据可以随机取出，像文章中数据太少，大了很容易过拟合。

import random

import torch.optim
import xlrd
import torch.nn as nn
import matplotlib.pyplot as plt

LR = 0.01


def get_readings():
    book = xlrd.open_workbook('./data/article_readings.xlsx')
    sheet1 = book.sheet_by_name('sheet1')
    cols = sheet1.col_values(1)
    readings = []
    for i in range(8, len(cols)):
        pre_days = []
        for j in range(i - 7, i):
            pre_days.append(int(cols[j]))
        pre_days.append(int(cols[i]))
        readings.append(pre_days)
    return readings


class MyNet(nn.Module):
    def __init__(self):
        super(MyNet, self).__init__()
        self.fc = nn.Sequential(
            nn.Linear(7, 16, bias=True),
            nn.ReLU(),
            nn.Linear(16, 32, bias=True),
            nn.ReLU(),
            nn.Linear(32, 1, bias=True)
        )
        self.optimizer = torch.optim.Adam(self.fc.parameters(), lr=LR)
        self.loss_fn = nn.MSELoss()

    def forward(self, input):
        return self.fc(input)


def train(model: MyNet, readings: list[[int, int]]):
    loss_list = []
    for reading in readings:
        predict = torch.tensor(reading[:7], dtype=torch.float)
        # predict = torch.unsqueeze(predict, 0)
        predict = model(predict)

        target = reading[7]
        target = torch.tensor(target, dtype=torch.float)
        target = torch.unsqueeze(target, 0)

        loss = model.loss_fn(target, predict)
        loss_list.append(loss.item())

        model.optimizer.zero_grad()
        loss.backward()
        model.optimizer.step()

    return loss_list


def predict(model: MyNet, pre_reading) -> int:
    pre_reading = torch.tensor(pre_reading, dtype=torch.float)
    pre_reading
    out = model(pre_reading).item()
    return int(out)


if __name__ == "__main__":
    model = MyNet()
    readings = get_readings()
    print(readings)
    pre_reading = readings[random.randint(0, len(readings))]
    real_reading = pre_reading[-1]
    loss_list_y = []
    epoch = 200  # 注意，这里的epoch是偏大的，因为数据量太小了，所以必然是过拟合的 
    for i in range(epoch):
        loss_list_y.extend(train(model, readings))
    print('train done!')
    loss_list_x = [i + 1 for i in range(len(loss_list_y))]
    plt.plot(loss_list_x, loss_list_y)
    plt.show()

    print(f'predict:{predict(model, pre_reading[:7])}, real:{real_reading}')