实践目标
通过一个预测新冠病毒发病率的题目，学习深度学习的训练过程，注意常见的库的使用、基于pytorch进行深度学习模型训练的流程。
先回顾下上篇提到的深度学习模型训练的流程，数据加载-->训练-->校验-->测试
代码结构
准备工作
- 准备数据集
- Kaggle下载数据:Kaggle: ml2022spring-hw1百度云下载数据: 云盘(提取码：ml22)
下载好后，放到项目根目录，声明数据集path，后面需要用
tr_path = 'covid.train.csv'tt_path = 'covid.test.csv'- import 常用库
- torch：张量、数据集积累、自动梯度、计算图、GPU加速、深度学习模块数据处理：numpy、csv绘图工具：Matplotlib，直观的呈现模型训练的效果
# PyTorch# 导入 PyTorch 核心库，后续所有张量操作、自动求导、模型定义都基于它import torch # torch.nn是 PyTorch 中定义各种神经网络层、损失函数等组件的模块。通常我们用 nn 作为别名，后面定义模型时会频繁使用，比如 nn.Linear、nn.Conv2d、nn.ReLU 等import torch.nn as nn# Dataset：抽象类，用于封装数据集；子类需要实现 __len__() 和 __getitem__()，以便按索引取样本# DataLoader：数据加载器，可以将任意 Dataset 封装成可迭代对象，自动支持多线程并行加载、批量采样（batch）、打乱顺序（shuffle）等from torch.utils.data import Dataset, DataLoaderprint(torch.__version__)print(torch.cuda.is_available())# For data processingimport numpy as npimport csvimport os# For plotting# Matplotlib 是最常用的绘图库，pyplot 提供类似 MATLAB 的绘图接口，用于可视化训练过程中的损失曲线、预测结果、图像样本等import matplotlib.pyplot as plt# 直接导入 figure() 函数，用来创建一个新的图表窗口或画布，方便接下来调用 fig = figure(figsize=(x, y)) 来设置画布大小from matplotlib.pyplot import figure- 准备好辅助函数
- 设备获取绘制学习曲线、预测值和真实值对比
def get_device(): return 'cuda' if torch.cuda.is_available() else 'cpu'def plot_learning_curve(loss_record, title=''): ''' 绘制学习曲线 ''' # 1. 确定横坐标范围：训练总步数 total_steps = len(loss_record['train']) x_1 = range(total_steps) # 2. 新建画布 figure(figsize=(6,4)) # 3. 绘制训练损失曲线 plt.plot(x_1, loss_record['train'], c='tab:red',label='train') # 4. 如果有验证集损失，则绘制验证曲线 if len(loss_record['dev']) != 0 : # 4.1 计算验证曲线在横坐标上的抽样点 # 因为 train 损失记录可能比 dev 多很多，因此做等间隔采样 x_2 = x_1[::len(loss_record['train']) // len(loss_record['dev'])] plt.plot(x_2, loss_record['dev'], c='tab:cyan', label='dev') # 5. 设置纵坐标范围 plt.ylim(0.0, 20.) # 6. 添加坐标轴标签和标题 plt.xlabel('Training steps') plt.ylabel('MSE loss') plt.title('Learning curve of {}'.format(title)) # 7. 显示图例并渲染 plt.legend() plt.show()def plot_pred(dv_set, model, device, lim=35., preds=None, targets=None): '''Plot prediction of your DNN''' if preds is None or targets is None: model.eval() preds, targets = [], [] for x, y in dv_set: x, y = x.to(device), y.to(device) with torch.no_grad(): pred = model(x) preds.append(pred.detach().cpu()) targets.append(y.detach().cpu()) print("preds shape : {}".format(len(preds))) preds = torch.cat(preds, dim=0).numpy() targets = torch.cat(targets, dim=0).numpy() print("preds after cat shape : {}".format(preds.shape)) figure(figsize=(5, 5)) plt.scatter(targets, preds, c='r', alpha=0.5) plt.plot([-0.2, lim], [-0.2, lim], c='b') plt.xlim(-0.2, lim) plt.ylim(-0.2, lim) plt.xlabel('ground truth value') plt.ylabel('predicted value') plt.title('Ground Truth v.s. Prediction') plt.show()定义数据结构&网络模型
- 定义数据结构
继承自torch中的基类Dataset，实现三个接口：
注意train(测试)、dev(校验)、test(测试)数据集处理的差异
class COVID19Dataset(Dataset): ''' Dataset for loading and preprocessing the COVID19 dataset''' def __init__(self, path, mode='train', target_only=False): self.mode = mode # Read data into numpy arrays with open(path, 'r') as fp: data = list(csv.reader(fp)) data = np.array(data[1:])[:,1:].astype(float) if not target_only: feats = list(range(93)) else: # TODO: Using 40 states & 2 tested_positive features (indices = 57 & 75) feats = [40, 41, 42, 43, 57, 58, 59, 60, 61, 75, 76, 77, 78, 79] # sklean mutual info if mode == 'test': # Testing data # data: 893 * 93 (40 states + day 1 (18) + day 2 (18) + day 3 (17)) data = data[:, feats] self.data = torch.FloatTensor(data) else: # Training data (train/dev sets) # data: 2700 * 94 (40 states + day 1 (18) + day 2 (18) + day 3 (18)) target = data[:, -1] data = data[:, feats] self.mean = torch.FloatTensor(data).mean(dim=0, keepdim=True) self.std = torch.FloatTensor(data).std(dim=0, keepdim=True) # Splitting training data into train & dev sets if mode == 'train': indices = [i for i in range(len(data)) if i % 10 != 0] elif mode == 'dev': indices = [i for i in range(len(data)) if i % 5 == 0] # Convert data into PyTorch tensors self.data = torch.FloatTensor(data[indices]) self.target = torch.FloatTensor(target[indices]) self.dim = self.data.shape[1] print('Finished reading the {} set of COVID19 Dataset ({} samples found, each dim = {})' .format(mode, len(self.data), self.dim)) def __getitem__(self, index): # Returns one sample at a time if self.mode in ['train', 'dev']: # For training return self.data[index], self.target[index] else: # For testing (no target) return self.data[index] def __len__(self): # Returns the size of the dataset return len(self.data) def normalization(self, mean=None, std=None): # Normalize each dimension to follow the Gaussian distribution # The mean and standard variance of training data will reused to normalize testing data. if self.mode == 'train' or self.mode == 'dev': mean = self.mean std = self.std self.data = (self.data-mean) / std else: self.data = (self.data-mean) / std return mean, std- 定义数据集加载的函数
使用torch提供的DataLoader，直接把数据灌进去即可，DataLoader能自动处理好batch、shuffle等工作
def prep_dataloader(path, mode, batch_size, n_jobs=0, target_only=False, mean=None, std=None): ''' Generates a dataset, then is put into a dataloader. ''' dataset = COVID19Dataset(path, mode=mode, target_only=target_only) # Construct dataset mean, std = dataset.normalization(mean, std) dataloader = DataLoader( dataset, batch_size, shuffle=(mode == 'train'), drop_last = False, num_workers = n_jobs, pin_memory=True) return dataloader, mean, std- 定义网络模型
网络模型的代码其实很少
模型层级的串联有两种写法，我们这里用左边的Sequential方式，更简洁
模型层级，是由大到小，这是固定的套路，这样的先扩散后收敛的形式，能让神经网络拟合出更丰富的特征空间。
注意！上一篇讲过，在torch中Linear是右乘。
class NeuralNet(nn.Module): ''' A simple fully-connected deep neural network ''' def __init__(self, input_dim): super(NeuralNet, self).__init__() # Define your neural network here # TODO: How to modify this model to achieve better performance? self.net = nn.Sequential( nn.Linear(input_dim, 64), nn.ReLU(), nn.Linear(64, 16), nn.ReLU(), nn.Linear(16, 8), nn.ReLU(), nn.Linear(8,4), nn.ReLU(), nn.Linear(4,1) ) # Mean squared error loss self.criterion = nn.MSELoss(reduction='mean') def forward(self, x): '''Given input of size(batch_size * input_dime), compute output of the network''' return self.net(x).squeeze(1) def cal_loss(self, pred, target, l1_lambda): ''' Calculate loss ''' loss = self.criterion(pred, target) # L1 regularization l1_reg = torch.tensor(0.).to(get_device()) for param in self.parameters(): l1_reg += torch.sum(torch.abs(param)) loss += l1_lambda * l1_reg return losstrain
train环节需要注意几点：
- 损失函数
- Mean Squared Error (for regression tasks)criterion = nn.MSELoss()Cross Entropy (for classification tasks)criterion = nn.CrossEntropyLoss()
- 参数优化
损失函数驱动模型参数朝哪个方向优化，参数优化方法决定怎么“快速”的将参数降到最优值。
可以粗糙的理解为：前者是指南针，后者是交通工具。合适的优化方法能大幅降低train的迭代次数。
最常见的就是随机梯度下降(Stochastic Gradient Descent (SGD))
torch.optim.SGD(model.parameters(), lr, momentum = 0)3. 整个训练流程为：
train代码：
def train(tr_set, dv_set, model, config, device): ''' DNN training ''' n_epochs = config['n_epochs'] # Maximum number of epochs # Setup optimizer optimizer = getattr(torch.optim, config['optimizer'])( model.parameters(), **config['optim_hparas']) min_mse = 1000. loss__record = {'train':[], 'dev':[]} early_stop_cnt = 0 epoch = 0 while epoch < n_epochs: model.train() # Set your model to train mode. for x, y in tr_set: # Iterate through the dataloader. optimizer.zero_grad() # Gradients stored in the parameters in the previous step should be cleared out first. x, y = x.to(device), y.to(device) # Move your data to device. pred = model(x) # Forward pass (compute output) mse_loss = model.cal_loss(pred, y, config['l1_lambda']) # Compute loss. mse_loss.backward() # Compute gradient(backpropagation). optimizer.step() # Update model with optimizer. loss__record['train'].append(mse_loss.detach().cpu().item()) # After each epoch, test your model on the validation (development) set. dev_mse = dev(dv_set, model, device) if dev_mse < min_mse: # Save model if your model improved min_mse = dev_mse print('Saving model (epoch = {:4d}, loss = {:.4f})'.format(epoch + 1, min_mse)) torch.save(model.state_dict(), config['save_path']) # Save model to specified path early_stop_cnt = 0 else: early_stop_cnt += 1 epoch += 1 loss__record['dev'].append(dev_mse) if early_stop_cnt > config['early_stop']: # Stop training if your model stops improving for "config['early_stop']" epochs. break print('Finished training after {} epochs'.format(epoch)) return min_mse, loss__record校验代码：
注意！校验时，要关闭自动梯度，减少不要的内存和计算开销
def dev(dv_set, model, device): model.eval() # Set your model to evaluation mode. total_loss = 0.0 for x, y in dv_set: # Iterate through the dataloader. x, y = x.to(device), y.to(device) # Move your data to device. with torch.no_grad(): pred = model(x) # Forward pass (compute output). mse_loss = model.cal_loss(pred, y, config['l1_lambda']) # Compute loss. total_loss += mse_loss.detach().cpu().item() * len(x) # accumulate loss total_loss = total_loss / len(dv_set.dataset) # Compute averaged loss. return total_loss- 超参数定义
就是一些常亮定义，这是学习项目，所以将一些seed定义成确定值，确保随机函数的一致性，便于观察模型的特点和结论复现。
device = get_device()os.makedirs('models', exist_ok=True)target_only = Trueseed = 459np.random.seed(seed)delta = np.random.normal(loc=0, scale = 0.000001)# TODO: How to tune these hyper-paramerters to improve your model's performance?config = { 'n_epochs': 3000, # maximum number of epochs 'batch_size': 270, # mini-batch size for dataloader 'optimizer': 'Adam', # optimization algorithm (optimizer in torch.optim) 'optim_hparas': { # hyper-parameters for the optimizer (depends on which optimizer you are using) 'lr': 0.001, # learning rate of SGD }, 'l1_lambda': 0.001 + delta, 'early_stop': 200, # early stopping epochs (the number epochs since your model's last improvement) 'save_path': 'models/model.pth' # your model will be saved here}# 把随机种子（seed）设置为一个固定的整数 42069，使得后续所有基于此种子的随机操作都能“从同一起点”开始。myseed = 42069# 强制 cuDNN 在可选的实现中只使用确定性算法（deterministic），避免某些卷积、池化等操作因为底层优化而产生的细微随机性。torch.backends.cudnn.deterministic = True# 关闭 cuDNN 的自动调优（benchmark）。如果启用，cuDNN 会根据多种算法在第一次运行时做性能测试，然后选最快的算法，这个选择过程本身可能引入非确定性。关闭后，每次都会用上面 deterministic 指定的那套实现torch.backends.cudnn.benchmark = False# 为 NumPy 的全局随机数生成器设置种子，保证后续所有 np.random.*（如 randn、choice 等）都能复现相同的随机序列。np.random.seed(myseed)# 为 PyTorch CPU 上的随机数生成器（RNG）设定种子，影响诸如 torch.randn、torch.randperm、dropout、数据打乱等操作在 CPU 上的随机性。torch.manual_seed(myseed)if torch.cuda.is_available(): # 如果检测到至少有一块可用 GPU，就为所有 GPU 设备上的 PyTorch RNG 也都设定相同的种子，保证在多卡训练时，每张卡上的随机操作也都可复现。 torch.cuda.manual_seed_all(myseed)数据加载&训练
- 加载数据
tr_set, mean, std = prep_dataloader(tr_path, 'train', config['batch_size'], target_only=target_only)dv_set, _, _ = prep_dataloader(tr_path, 'dev', config['batch_size'], target_only=target_only, mean=mean, std=std)tt_set, _, _ = prep_dataloader(tt_path, 'test', config['batch_size'], target_only=target_only, mean=mean, std=std)# 日志# Finished reading the train set of COVID19 Dataset (2429 samples found, each dim = 14)# Finished reading the dev set of COVID19 Dataset (540 samples found, each dim = 14)# Finished reading the test set of COVID19 Dataset (1078 samples found, each dim = 14)- train触发
model = NeuralNet(tr_set.dataset.dim).to(device)model_loss,model_loss_record = train(tr_set, dv_set, model, config, device)# 耗时：2m 4.0s- 绘制学习曲线，观察收敛趋势
plot_learning_curve(model_loss_record, title='deep model')- 模型测试
用测试数据集测试模型
del model # 输入的batch可能不一样，删除模型，重新建一个model = NeuralNet(tr_set.dataset.dim).to(device)ckpt = torch.load(config['save_path'], map_location='cpu') # Load your best modelmodel.load_state_dict(ckpt)if len(dv_set) > 0: plot_pred(dv_set, model, device) # Show prediction on the validation set可以观察到，predicted value和ground truth value组成的(x,y)值很靠近斜对角线，说明模型预测的值还是比较准的。
总结
第一个正式的案例，把深度学习过程涉及到的知识点都带上了。麻雀虽小五脏俱全，要理解透彻需要花一单时间。
可以看到，真正的模型定义部分的代码是很少的，大量的代码都是关于数据处理，数据呈现的。
参考：
- 李宏毅机器学习github.com/pai4451/ML2…
