【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)
3 改进程序
以下统一设定1000epoch，改进角度包括
（1）特征选择
- 皮尔逊相关性 斯皮尔曼相关性
（2）模型改进
- DNN
- FCDNN
- DenseNet
- ResNet
（3）优化器
- SGD
- Aadm
- Adadelta
（4）余弦学习率
# 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 *
# 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}""")
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.
# feat_idx = range(0,117)
correlation_matrix = np.corrcoef(raw_x_train, rowvar=False)
corr_with_target = np.abs(correlation_matrix[-1, :-1])
feat_idx = list(np.argsort(corr_with_target)[::-1][:100]) # 选择与目标变量相关性最高的五个特征索引
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)
class Raw_Model(nn.Module):
def __init__(self, input_dim):
super(Raw_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
import torch
import torch.nn as nn
import torch.nn.functional as F
class FCNN_Model(nn.Module):
def __init__(self, input_dim):
super(FCNN_Model, self).__init__()
# 修改模型结构
self.layers = nn.Sequential(
nn.Linear(input_dim, 64),
nn.BatchNorm1d(64),
nn.LeakyReLU(0.01), # 使用LeakyReLU
nn.Dropout(0.3),
nn.Linear(64, 32),
nn.BatchNorm1d(32),
nn.LeakyReLU(0.01), # 使用LeakyReLU
nn.Dropout(0.3),
nn.Linear(32, 16),
nn.BatchNorm1d(16),
nn.LeakyReLU(0.01), # 使用LeakyReLU
nn.Dropout(0.3),
nn.Linear(16, 1)
)
def forward(self, x):
x = self.layers(x)
x = x.squeeze(1) # (B, 1) -> (B)
return x
import torch
import torch.nn as nn
import torch.nn.functional as F
# 定义基础的残差块
class ResidualBlock(nn.Module):
def __init__(self, input_dim):
super(ResidualBlock, self).__init__()
self.fc1 = nn.Linear(input_dim, input_dim)
self.relu = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(input_dim, input_dim)
def forward(self, x):
residual = x
out = self.fc1(x)
out = self.relu(out)
out = self.fc2(out)
out += residual # 这里添加了跳越连接
out = self.relu(out)
return out
# 将模型定义为一个等于ResNet的回归模型
class RegressionResNet(nn.Module):
def __init__(self, input_dim, num_blocks=2):
super(RegressionResNet, self).__init__()
# 输入层
self.input_fc = nn.Linear(input_dim, input_dim)
# 创建残差块堆叠
self.res_blocks = nn.Sequential(
*[ResidualBlock(input_dim) for _ in range(num_blocks)]
)
# 输出层
self.output_fc = nn.Linear(input_dim, 1)
def forward(self, x):
x = F.relu(self.input_fc(x))
x = self.res_blocks(x)
x = self.output_fc(x)
x = x.squeeze(1) # (B, 1) -> (B)
return x
import torch
import torch.nn as nn
import torch.nn.functional as F
class DenseLayer(nn.Module):
def __init__(self, in_channels, growth_rate):
super(DenseLayer, self).__init__()
# A single Dense Layer within a Dense Block
self.dense_layer = nn.Sequential(
nn.BatchNorm1d(in_channels),
nn.ReLU(inplace=True),
nn.Linear(in_channels, growth_rate),
nn.Dropout(0.2) # Dropout for regularization
)
def forward(self, x):
new_features = self.dense_layer(x)
# Concatenating the input features with the new features
return torch.cat([x, new_features], 1)
class DenseBlock(nn.Module):
def __init__(self, num_layers, in_channels, growth_rate):
super(DenseBlock, self).__init__()
self.block = nn.Sequential()
for i in range(num_layers):
layer = DenseLayer(in_channels + i * growth_rate, growth_rate)
self.block.add_module(f"dense_layer_{i + 1}", layer)
def forward(self, x):
return self.block(x)
class TransitionLayer(nn.Module):
def __init__(self, in_channels, out_channels):
super(TransitionLayer, self).__init__()
# This layer reduces the number of features (compression)
self.transition = nn.Sequential(
nn.BatchNorm1d(in_channels),
nn.ReLU(inplace=True),
nn.Linear(in_channels, out_channels),
nn.Dropout(0.2) # Dropout for regularization
)
def forward(self, x):
return self.transition(x)
class My_DenseNet_Model(nn.Module):
def __init__(self, input_dim, num_classes=1, growth_rate=12, block_config=(6, 12, 24), compression=0.5):
super(My_DenseNet_Model, self).__init__()
# Initial convolution layer
self.init_features = nn.Sequential(
nn.Linear(input_dim, growth_rate * 2),
nn.ReLU(inplace=True)
)
# DenseBlocks and TransitionLayers
num_features = growth_rate * 2 # Initial number of features
self.features = nn.Sequential()
for i, num_layers in enumerate(block_config):
block = DenseBlock(num_layers=num_layers, in_channels=num_features, growth_rate=growth_rate)
self.features.add_module(f"denseblock_{i + 1}", block)
num_features += num_layers * growth_rate
if i != len(block_config) - 1: # Do not add Transition Layer after the last block
trans = TransitionLayer(in_channels=num_features, out_channels=int(num_features * compression))
self.features.add_module(f"transition_{i + 1}", trans)
num_features = int(num_features * compression)
# Final batch normalization
self.features.add_module('norm5', nn.BatchNorm1d(num_features))
# Linear layer for regression
self.classifier = nn.Linear(num_features, num_classes)
def forward(self, x):
x = self.init_features(x)
x = self.features(x)
x = F.relu(x, inplace=True)
x = F.avg_pool1d(x, kernel_size=1).view(x.size(0), -1)
x = self.classifier(x)
return x
from torch.optim.lr_scheduler import CosineAnnealingLR
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
optimizer = torch.optim.SGD(model.parameters(), lr=config['learning_rate'], momentum=0.9)
# optimizer = torch.optim.Adam(model.parameters(), lr=config['learning_rate'])
# optimizer = torch.optim.Adadelta(model.parameters(), lr=config['learning_rate'], rho=0.9, eps=1e-06, weight_decay=0)
# 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'))
# 添加余弦退火调度器
scheduler = CosineAnnealingLR(optimizer, T_max=config['n_epochs'], eta_min=0)
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
# 更新学习率
scheduler.step()
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = Raw_Model(input_dim=x_train.shape[1]).to(device)
# model = RegressionResNet(input_dim=x_train.shape[1],num_blocks=10).to(device)
# model = My_DenseNet_Model(input_dim=x_train.shape[1]).to(device)
# model = FCNN_Model(input_dim=x_train.shape[1]).to(device)
trainer(train_loader, valid_loader, model, config, device)
# model = RegressionResNet(input_dim=x_train.shape[1],num_blocks=10).to(device)
# model = My_DenseNet_Model(input_dim=x_train.shape[1]).to(device)
model = Raw_Model(input_dim=x_train.shape[1]).to(device)
# model = FCNN_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)
