import math import os import sys import time import copy import torch.nn.functional as F import torch from torch import nn from tqdm import tqdm from lib.logger import get_logger from lib.loss_function import all_metrics from model.STMLP.STMLP import STMLP class Trainer: def __init__(self, model, loss, optimizer, train_loader, val_loader, test_loader, scaler, args, lr_scheduler=None): self.model = model self.loss = loss self.optimizer = optimizer self.train_loader = train_loader self.val_loader = val_loader self.test_loader = test_loader self.scaler = scaler self.args = args['train'] self.lr_scheduler = lr_scheduler self.train_per_epoch = len(train_loader) self.val_per_epoch = len(val_loader) if val_loader else 0 # Paths for saving models and logs self.best_path = os.path.join(self.args['log_dir'], 'best_model.pth') self.best_test_path = os.path.join(self.args['log_dir'], 'best_test_model.pth') self.loss_figure_path = os.path.join(self.args['log_dir'], 'loss.png') self.pretrain_dir = f'./pre-train/{args["model"]["type"]}/{args["data"]["type"]}' self.pretrain_path = os.path.join(self.pretrain_dir, 'best_model.pth') self.pretrain_best_path = os.path.join(self.pretrain_dir, 'best_test_model.pth') # Initialize logger if not os.path.isdir(self.args['log_dir']) and not self.args['debug']: os.makedirs(self.args['log_dir'], exist_ok=True) if not os.path.isdir(self.pretrain_dir) and not self.args['debug']: os.makedirs(self.pretrain_dir, exist_ok=True) self.logger = get_logger(self.args['log_dir'], name=self.model.__class__.__name__, debug=self.args['debug']) self.logger.info(f"Experiment log path in: {self.args['log_dir']}") if self.args['teacher_stu']: self.tmodel = self.loadTeacher(args) else: self.logger.info(f"当前使用预训练模式，预训练后请移动教师模型到" f"./pre-train/{args['model']['type']}/{args['data']['type']}/best_model.pth" f"然后在config中配置train.teacher_stu模式为True开启蒸馏模式") def _run_epoch(self, epoch, dataloader, mode): # self.tmodel.eval() if mode == 'train': self.model.train() optimizer_step = True else: self.model.eval() optimizer_step = False total_loss = 0 epoch_time = time.time() with torch.set_grad_enabled(optimizer_step): with tqdm(total=len(dataloader), desc=f'{mode.capitalize()} Epoch {epoch}') as pbar: for batch_idx, (data, target) in enumerate(dataloader): if self.args['teacher_stu']: label = target[..., :self.args['output_dim']] output, out_, _ = self.model(data) gout, tout, sout = self.tmodel(data) if self.args['real_value']: output = self.scaler.inverse_transform(output) loss1 = self.loss(output, label) scl = self.loss_cls(out_, sout) kl_loss = nn.KLDivLoss(reduction="batchmean", log_target=True).cuda() gout = F.log_softmax(gout, dim=-1).cuda() mlp_emb_ = F.log_softmax(output, dim=-1).cuda() tkloss = kl_loss(mlp_emb_.cuda().float(), gout.cuda().float()) loss = loss1 + 10 * tkloss + 1 * scl else: label = target[..., :self.args['output_dim']] output, out_, _ = self.model(data) if self.args['real_value']: output = self.scaler.inverse_transform(output) loss = self.loss(output, label) if optimizer_step and self.optimizer is not None: self.optimizer.zero_grad() loss.backward() if self.args['grad_norm']: torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args['max_grad_norm']) self.optimizer.step() total_loss += loss.item() if mode == 'train' and (batch_idx + 1) % self.args['log_step'] == 0: self.logger.info( f'Train Epoch {epoch}: {batch_idx + 1}/{len(dataloader)} Loss: {loss.item():.6f}') # 更新 tqdm 的进度 pbar.update(1) pbar.set_postfix(loss=loss.item()) avg_loss = total_loss / len(dataloader) self.logger.info( f'{mode.capitalize()} Epoch {epoch}: average Loss: {avg_loss:.6f}, time: {time.time() - epoch_time:.2f} s') return avg_loss def train_epoch(self, epoch): return self._run_epoch(epoch, self.train_loader, 'train') def val_epoch(self, epoch): return self._run_epoch(epoch, self.val_loader or self.test_loader, 'val') def test_epoch(self, epoch): return self._run_epoch(epoch, self.test_loader, 'test') def train(self): best_model, best_test_model = None, None best_loss, best_test_loss = float('inf'), float('inf') not_improved_count = 0 self.logger.info("Training process started") for epoch in range(1, self.args['epochs'] + 1): train_epoch_loss = self.train_epoch(epoch) val_epoch_loss = self.val_epoch(epoch) test_epoch_loss = self.test_epoch(epoch) if train_epoch_loss > 1e6: self.logger.warning('Gradient explosion detected. Ending...') break if val_epoch_loss < best_loss: best_loss = val_epoch_loss not_improved_count = 0 best_model = copy.deepcopy(self.model.state_dict()) torch.save(best_model, self.best_path) torch.save(best_model, self.pretrain_path) self.logger.info('Best validation model saved!') else: not_improved_count += 1 if self.args['early_stop'] and not_improved_count == self.args['early_stop_patience']: self.logger.info( f"Validation performance didn't improve for {self.args['early_stop_patience']} epochs. Training stops.") break if test_epoch_loss < best_test_loss: best_test_loss = test_epoch_loss best_test_model = copy.deepcopy(self.model.state_dict()) torch.save(best_test_model, self.best_test_path) torch.save(best_model, self.pretrain_best_path) if not self.args['debug']: torch.save(best_model, self.best_path) torch.save(best_test_model, self.best_test_path) self.logger.info(f"Best models saved at {self.best_path} and {self.best_test_path}") self._finalize_training(best_model, best_test_model) def _finalize_training(self, best_model, best_test_model): self.model.load_state_dict(best_model) self.logger.info("Testing on best validation model") self.test(self.model, self.args, self.test_loader, self.scaler, self.logger) self.model.load_state_dict(best_test_model) self.logger.info("Testing on best test model") self.test(self.model, self.args, self.test_loader, self.scaler, self.logger) def loadTeacher(self, args): model_path = f'./pre-train/{args["model"]["type"]}/{args["data"]["type"]}/best_model.pth' try: # 尝试加载教师模型权重 state_dict = torch.load(model_path) self.logger.info(f"成功加载教师模型权重: {model_path}") # 初始化并返回教师模型 args['model']['model_type'] = 'teacher' tmodel = STMLP(args['model']) tmodel = tmodel.to(args['device']) tmodel.load_state_dict(state_dict, strict=False) return tmodel except FileNotFoundError: # 如果找不到权重文件，记录日志并修改 args self.logger.error( f"未找到教师模型权重文件: {model_path}。切换到预训练模式训练老师权重。\n" f"在预训练完成后，再次启动模型则为蒸馏模式") self.args['teacher_stu'] = False return None def loss_cls(self, x1, x2): temperature = 0.05 x1 = F.normalize(x1, p=2, dim=-1) x2 = F.normalize(x2, p=2, dim=-1) weight = F.cosine_similarity(x1, x2, dim=-1) batch_size = x1.size()[0] # neg score out = torch.cat([x1, x2], dim=0) neg = torch.exp(torch.matmul(out, out.transpose(2, 3).contiguous()) / temperature) pos = torch.exp(torch.sum(x1 * x2, dim=-1) * weight / temperature) # pos = torch.exp(torch.sum(x1 * x2, dim=-1) / temperature) pos = torch.cat([pos, pos], dim=0).sum(dim=1) Ng = neg.sum(dim=-1).sum(dim=1) loss = (- torch.log(pos / (pos + Ng))).mean() return loss @staticmethod def test(model, args, data_loader, scaler, logger, path=None): if path: checkpoint = torch.load(path) model.load_state_dict(checkpoint['state_dict']) model.to(args['device']) model.eval() y_pred, y_true = [], [] with torch.no_grad(): for data, target in data_loader: label = target[..., :args['output_dim']] output, _, _ = model(data) y_pred.append(output) y_true.append(label) if args['real_value']: y_pred = scaler.inverse_transform(torch.cat(y_pred, dim=0)) else: y_pred = torch.cat(y_pred, dim=0) y_true = torch.cat(y_true, dim=0) # 你在这里需要把y_pred和y_true保存下来 # torch.save(y_pred, "./test/PEMS07/y_pred_D.pt") # [3566,12,170,1] # torch.save(y_true, "./test/PEMSD8/y_true.pt") # [3566,12,170,1] for t in range(y_true.shape[1]): mae, rmse, mape = all_metrics(y_pred[:, t, ...], y_true[:, t, ...], args['mae_thresh'], args['mape_thresh']) logger.info(f"Horizon {t + 1:02d}, MAE: {mae:.4f}, RMSE: {rmse:.4f}, MAPE: {mape:.4f}") mae, rmse, mape = all_metrics(y_pred, y_true, args['mae_thresh'], args['mape_thresh']) logger.info(f"Average Horizon, MAE: {mae:.4f}, RMSE: {rmse:.4f}, MAPE: {mape:.4f}") @staticmethod def _compute_sampling_threshold(global_step, k): return k / (k + math.exp(global_step / k))