【人工智能课程】计算机科学博士作业一

【人工智能课程】计算机科学博士作业一

1 任务要求

• 模型拟合：用深度神经网络拟合一个回归模型。从各种角度对其改进，评价指标为MSE。
• 掌握技巧：
  • 熟悉并掌握深度学习模型训练的基本技巧。
  • 提高PyTorch的使用熟练度。
  • 掌握改进深度学习的方法。

数据集下载：

• Kaggle下载数据:
  https://www.kaggle.com/competitions/ml2022spring-hw1

• 百度云下载数据: https://pan.baidu.com/s/1ahGxV7dO2JQMRCYbmDQyVg (提取码：ml22)

这是一个非时间序列的回归任务，预测公共场所获取的人群数据，预测会发生COVID-19阳性的人数。改进角度，参考博客：http://t.csdnimg.cn/fUAzT

2 baseline 代码

我将老师给的代码重构了结构，便于组员之间协作编程，无需修改的代码都放到了utils.py中。只需要修改特征选择、神经网络、模型训练部分的代码就可以。

2.1 导入包

# 数值、矩阵操作
import math
# 数据读取与写入make_dot
import pandas as pd
import os
import csv
# 学习曲线绘制
from torch.utils.tensorboard import SummaryWriter
from utils import *

2.2 数据读取

# 设置随机种子便于复现
same_seed(config['seed'])

# 训练集大小(train_data size) : 2699 x 118 (id + 37 states + 16 features x 5 days) 
# 测试集大小(test_data size）: 1078 x 117 (没有label (last day's positive rate))
pd.set_option('display.max_column', 200) # 设置显示数据的列数
train_df, test_df = pd.read_csv('./covid.train.csv'), pd.read_csv('./covid.test.csv')
display(train_df.head(3)) # 显示前三行的样本
train_data, test_data = train_df.values, test_df.values
del train_df, test_df # 删除数据减少内存占用
train_data, valid_data = train_valid_split(train_data, config['valid_ratio'], config['seed'])

# 打印数据的大小
print(f"""train_data size: {train_data.shape} 
valid_data size: {valid_data.shape} 
test_data size: {test_data.shape}""")

2.3 特征选择

def select_feat(train_data, valid_data, test_data, select_all=True):
    '''
    特征选择
    选择较好的特征用来拟合回归模型
    '''
    y_train, y_valid = train_data[:,-1], valid_data[:,-1]
    raw_x_train, raw_x_valid, raw_x_test = train_data[:,:-1], valid_data[:,:-1], test_data

    if select_all:
        feat_idx = list(range(raw_x_train.shape[1]))
    else:
        feat_idx = [0,1,2,3,4] # TODO: 选择需要的特征 ，这部分可以自己调研一些特征选择的方法并完善.

    return raw_x_train[:,feat_idx], raw_x_valid[:,feat_idx], raw_x_test[:,feat_idx], y_train, y_valid

# 特征选择
x_train, x_valid, x_test, y_train, y_valid = select_feat(train_data, valid_data, test_data, config['select_all'])

# 打印出特征数量.
print(f'number of features: {x_train.shape[1]}')

train_dataset, valid_dataset, test_dataset = COVID19Dataset(x_train, y_train), \
                                            COVID19Dataset(x_valid, y_valid), \
                                            COVID19Dataset(x_test)

# 使用Pytorch中Dataloader类按照Batch将数据集加载
train_loader = DataLoader(train_dataset, batch_size=config['batch_size'], shuffle=True, pin_memory=True)
valid_loader = DataLoader(valid_dataset, batch_size=config['batch_size'], shuffle=True, pin_memory=True)
test_loader = DataLoader(test_dataset, batch_size=config['batch_size'], shuffle=False, pin_memory=True)

2.4 神经网络

class My_Model(nn.Module):
    def __init__(self, input_dim):
        super(My_Model, self).__init__()
        # TODO: 修改模型结构, 注意矩阵的维度（dimensions） 
        self.layers = nn.Sequential(
            nn.Linear(input_dim, 16),
            nn.ReLU(),
            nn.Linear(16, 8),
            nn.ReLU(),
            nn.Linear(8, 1)
        )

    def forward(self, x):
        x = self.layers(x)
        x = x.squeeze(1) # (B, 1) -> (B)
        return x

2.5 模型训练

def trainer(train_loader, valid_loader, model, config, device):

    criterion = nn.MSELoss(reduction='mean') # 损失函数的定义
    # 定义优化器
    # TODO: 可以查看学习更多的优化器 https://pytorch.org/docs/stable/optim.html 
    # TODO: L2 正则( 可以使用optimizer(weight decay...) )或者 自己实现L2正则.
    optimizer = torch.optim.SGD(model.parameters(), lr=config['learning_rate'], momentum=0.9) 
    
    # tensorboard 的记录器
    # 将 train loss 保存到 "tensorboard/train" 文件夹
    train_writer = SummaryWriter(log_dir=os.path.join('tensorboard', 'train'))
    # 将 valid loss 保存到 "tensorboard/valid" 文件夹
    valid_writer = SummaryWriter(log_dir=os.path.join('tensorboard', 'valid'))


    if not os.path.isdir('./models'):
        # 创建文件夹-用于存储模型
        os.mkdir('./models')

    n_epochs, best_loss, step, early_stop_count = config['n_epochs'], math.inf, 0, 0

    for epoch in range(n_epochs):
        model.train() # 训练模式
        loss_record = []
        # tqdm可以帮助我们显示训练的进度  
        train_pbar = tqdm(train_loader, position=0, leave=True)
        # 设置进度条的左边 ： 显示第几个Epoch了
        train_pbar.set_description(f'Epoch [{epoch+1}/{n_epochs}]')
        for x, y in train_pbar:
            optimizer.zero_grad()               # 将梯度置0.
            x, y = x.to(device), y.to(device)   # 将数据一到相应的存储位置(CPU/GPU)
            pred = model(x)             
            loss = criterion(pred, y)
            loss.backward()                     # 反向传播 计算梯度.
            optimizer.step()                    # 更新网络参数
            step += 1
            loss_record.append(loss.detach().item())
            
            # 训练完一个batch的数据，将loss 显示在进度条的右边
            train_pbar.set_postfix({'loss': loss.detach().item()})

        mean_train_loss = sum(loss_record)/len(loss_record)
        
        model.eval() # 将模型设置成 evaluation 模式.
        loss_record = []
        for x, y in valid_loader:
            x, y = x.to(device), y.to(device)
            with torch.no_grad():
                pred = model(x)
                loss = criterion(pred, y)

            loss_record.append(loss.item())
            
        mean_valid_loss = sum(loss_record)/len(loss_record)
        print(f'Epoch [{epoch+1}/{n_epochs}]: Train loss: {mean_train_loss:.4f}, Valid loss: {mean_valid_loss:.4f}')
        # 每个epoch,在tensorboard 中记录验证的损失（后面可以展示出来）
        # 将训练损失和验证损失写入TensorBoard
        train_writer.add_scalar('Train-Valid Loss', mean_train_loss, step)
        valid_writer.add_scalar('Train-Valid Loss', mean_valid_loss, step)

        if mean_valid_loss < best_loss:
            best_loss = mean_valid_loss
            torch.save(model.state_dict(), config['save_path']) # 模型保存
            print('Saving model with loss {:.3f}...'.format(best_loss))
            early_stop_count = 0
        else: 
            early_stop_count += 1

        if early_stop_count >= config['early_stop']:
            print('\nModel is not improving, so we halt the training session.')
            return
        
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = My_Model(input_dim=x_train.shape[1]).to(device) # 将模型和训练数据放在相同的存储位置(CPU/GPU)
trainer(train_loader, valid_loader, model, config, device)

2.6 模型可视化

%reload_ext tensorboard
%tensorboard --logdir=tensorboard
#执行完后这两行代码，在浏览器打开：http://localhost:6006/

打开后，将smoothing调为0，就不会有四条曲线了。如果不改为0，就会自动加入一条平滑后的曲线在图中，影响观察。

2.7 模型评价

model = My_Model(input_dim=x_train.shape[1]).to(device)
model.load_state_dict(torch.load(config['save_path']))
MSE = predict_MSE(valid_loader, model, device) 

print("MSE:",MSE) 

只跑了10epoch的MSE
MSE: 30.798155

2.8 新建一个utils.py文件

把以下代码放进去utils.py文件中，放到和以上代码文件同一级的目录

import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader, random_split
import numpy as np
from tqdm import tqdm



config = {
    'seed': 5201314,      # 随机种子，可以自己填写. :)
    'select_all': True,   # 是否选择全部的特征
    'valid_ratio': 0.2,   # 验证集大小(validation_size) = 训练集大小(train_size) * 验证数据占比(valid_ratio)
    'n_epochs': 10,     # 数据遍历训练次数
    'batch_size': 256,
    'learning_rate': 1e-5,
    'early_stop': 400,    # 如果early_stop轮损失没有下降就停止训练.
    'save_path': './models/model.ckpt'  # 模型存储的位置
}

def same_seed(seed):
    '''
    设置随机种子(便于复现)
    '''
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)
    print(f'Set Seed = {seed}')


def train_valid_split(data_set, valid_ratio, seed):
    '''
    数据集拆分成训练集（training set）和 验证集（validation set）
    '''
    valid_set_size = int(valid_ratio * len(data_set))
    train_set_size = len(data_set) - valid_set_size
    train_set, valid_set = random_split(data_set, [train_set_size, valid_set_size], generator=torch.Generator().manual_seed(seed))
    return np.array(train_set), np.array(valid_set)


def predict(test_loader, model, device):
    model.eval() # 设置成eval模式.
    preds = []
    for x in tqdm(test_loader):
        x = x.to(device)
        with torch.no_grad():
            pred = model(x)
            preds.append(pred.detach().cpu())
    preds = torch.cat(preds, dim=0).numpy()
    return preds


def predict_MSE(valid_loader, model, device):
    model.eval() # 设置成eval模式.
    preds = []
    labels = []
    for x,y in tqdm(valid_loader):
        x = x.to(device)
        with torch.no_grad():
            pred = model(x)
            preds.append(pred.detach().cpu())
            labels.append(y)
    preds = torch.cat(preds, dim=0).numpy()
    labels = torch.cat(labels, dim=0).numpy()
    # 计算MSE
    mse = np.mean((preds - labels) ** 2)
    return mse



class COVID19Dataset(Dataset):
    '''
    x: np.ndarray  特征矩阵.
    y: np.ndarray  目标标签, 如果为None,则是预测的数据集
    '''
    def __init__(self, x, y=None):
        if y is None:
            self.y = y
        else:
            self.y = torch.FloatTensor(y)
        self.x = torch.FloatTensor(x)

    def __getitem__(self, idx):
        if self.y is None:
            return self.x[idx]
        return self.x[idx], self.y[idx]

    def __len__(self):
        return len(self.x)