import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import tqdm class OGD: ## online gradient descent def __init__(self, model: nn.Module,lr=0.01,device='cuda:0'): self.model=model self.lr=lr self.optimizer=optim.SGD(model.parameters(),lr=lr) self.device=device def online(self,test_loader:torch.utils.data.DataLoader): preds=[] labels=[] for data in tqdm.tqdm(test_loader): for i in range(5): self.optimizer.zero_grad() x,y=data x=x.to(self.device) y=y.to(self.device) output=self.model(x) if i==0: preds.append(output.detach().cpu().numpy()) loss=F.mse_loss(output,y) loss.backward() self.optimizer.step() labels.append(y.detach().cpu().numpy()) return preds,labels import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from collections import deque import random import tqdm class ER: """Experience Replay with buffer to store historical data batches.""" def __init__(self, model: nn.Module, lr=0.01, buffer_size=100, device='cuda:0'): self.model = model self.lr = lr self.optimizer = optim.SGD(model.parameters(), lr=lr) self.device = device self.buffer = deque(maxlen=buffer_size) # 固定大小的经验缓冲区 def online(self, test_loader: torch.utils.data.DataLoader): """Process test data with experience replay.""" preds = [] labels = [] for data in tqdm.tqdm(test_loader): self.optimizer.zero_grad() # 处理当前批次数据 x, y = data x = x.to(self.device) y = y.to(self.device) # 将当前批次存入缓冲区（自动淘汰旧数据） self.buffer.append((x.detach().clone(), y.detach().clone())) # 从缓冲区随机采样旧数据（至少1个批次） replay_x, replay_y = [], [] if len(self.buffer) > 0: num_replay = min(1, len(self.buffer)) # 采样1个旧批次 replay_samples = random.sample(self.buffer, num_replay) replay_x, replay_y = zip(*replay_samples) replay_x = torch.cat(replay_x, dim=0) replay_y = torch.cat(replay_y, dim=0) # 合并当前数据与回放数据 if len(replay_x) > 0: combined_x = torch.cat([x, replay_x], dim=0) combined_y = torch.cat([y, replay_y], dim=0) else: combined_x, combined_y = x, y # 前向传播与损失计算 output = self.model(combined_x) loss = F.mse_loss(output, combined_y) # 反向传播与参数更新 loss.backward() self.optimizer.step() # 记录预测结果 preds.append(output[0].detach().cpu().numpy()) labels.append(y.detach().cpu().numpy()) return preds, labels class OGD_output: ## online gradient descent for output def __init__(self, model: nn.Module,num_node,lr=0.01,device='cuda:0'): self.model=model self.lr=lr self.delta=nn.Parameter(torch.zeros((1,1,num_node,2),device=device),requires_grad=True) self.optimizer=optim.SGD([self.delta],lr=lr) self.device=device def online(self,test_loader:torch.utils.data.DataLoader): preds=[] labels=[] for data in tqdm.tqdm(test_loader): self.optimizer.zero_grad() x,y=data x=x.to(self.device) y=y.to(self.device) output=self.model(x)+self.delta loss=abs(output-y).mean() loss.backward() self.optimizer.step() preds.append(output.detach().cpu().numpy()) labels.append(y.detach().cpu().numpy()) return preds,labels if __name__ == "__main__": model='GWNET' train_month=12 import yaml import os import yaml import torch import numpy as np from sklearn.linear_model import Ridge from collections import defaultdict, deque import tqdm import numpy as np from torch.utils.data import DataLoader from datetime import datetime from models import STGCN,AGCRN,DCRNN,GWNET,MTGNN,causal_model from tools.data_tools import load_data, get_datasets, expand_adjacency_matrix, load_paper_adj from Dataset import STDataset import csv from sklearn.metrics import mean_absolute_error, mean_squared_error import logging print(torch.version.cuda) model_dict={'STGCN':STGCN,'AGCRN':AGCRN,'DCRNN':DCRNN,'GWNET':GWNET,'MTGNN':MTGNN} with open(r'models\nyctaxi_config.yaml', 'r') as config_file: config = yaml.safe_load(config_file) config['model'],config['train_months']=model,train_month # Configure logging # log_dir = config['log_dir'] # You can specify the directory for log files # logging.basicConfig(filename=log_file_path, level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s') # logging.info('{}_training_month_{}_.log'.format(config['model'],config['train_months'])) # Check if a GPU is available, otherwise use CPU device = config['device'] # Get datasets and scaler train_dataset, val_dataset, test_dataset, scaler,valid_gird = get_datasets( config) logging.info("Configuration:") for key, value in config.items(): logging.info(f"{key}: {value}") # Create data loaders train_loader = DataLoader(train_dataset, batch_size=config['batch_size'], shuffle=True) val_loader = DataLoader(val_dataset, batch_size=1, shuffle=False) test_loader = DataLoader(test_dataset,1, shuffle=False) adj_mx = load_paper_adj(config, valid_gird) config['adj_mx_file'] = os.path.join(config.get('resolved_dataset_dir', 'dataset'), 'adj_mx.npy') config['num_nodes']=len(valid_gird) # 初始化基础模型（假设已有STGCN模型） base_stgcn = causal_model.CausalModel(model_dict[config['model']].Model,config,adj_mx).to(device) base_stgcn.eval() base_stgcn.load_state_dict(torch.load(f'saved_models\{model}_nyctaxi_final_model.pth')) ogd=OGD(base_stgcn,lr=0.01,device='cuda:0') preds, labels = ogd.online(test_loader) preds = np.concatenate(preds) labels = np.concatenate(labels) preds = scaler.inverse_transform(preds).reshape(-1) labels = scaler.inverse_transform(labels).reshape(-1) mae = mean_absolute_error(labels, preds) mse = mean_squared_error(labels, preds) rmse = np.sqrt(mse) print(f"MAE: {mae:.4f}, MSE: {mse:.4f}, RMSE: {rmse:.4f}")