389 lines
13 KiB
Python
Executable File
389 lines
13 KiB
Python
Executable File
import math
|
|
import os
|
|
import time
|
|
import copy
|
|
import psutil
|
|
import torch
|
|
from utils.logger import get_logger
|
|
from utils.loss_function import all_metrics
|
|
from tqdm import tqdm
|
|
|
|
|
|
class TrainingStats:
|
|
"""记录训练过程中的统计信息"""
|
|
|
|
def __init__(self, device):
|
|
self.device = device
|
|
self.reset()
|
|
|
|
def reset(self):
|
|
"""重置所有统计数据"""
|
|
self.gpu_mem_usage_list = []
|
|
self.cpu_mem_usage_list = []
|
|
self.train_time_list = []
|
|
self.infer_time_list = []
|
|
self.total_iters = 0
|
|
self.start_time = None
|
|
self.end_time = None
|
|
|
|
def start_training(self):
|
|
"""记录训练开始时间"""
|
|
self.start_time = time.time()
|
|
|
|
def end_training(self):
|
|
"""记录训练结束时间"""
|
|
self.end_time = time.time()
|
|
|
|
def record_step_time(self, duration, mode):
|
|
"""记录单步耗时和总迭代次数"""
|
|
if mode == "train":
|
|
self.train_time_list.append(duration)
|
|
else:
|
|
self.infer_time_list.append(duration)
|
|
self.total_iters += 1
|
|
|
|
def record_memory_usage(self):
|
|
"""记录当前 GPU 和 CPU 内存占用"""
|
|
process = psutil.Process(os.getpid())
|
|
cpu_mem = process.memory_info().rss / (1024**2)
|
|
|
|
if torch.cuda.is_available():
|
|
gpu_mem = torch.cuda.max_memory_allocated(device=self.device) / (1024**2)
|
|
torch.cuda.reset_peak_memory_stats(device=self.device)
|
|
else:
|
|
gpu_mem = 0.0
|
|
|
|
self.cpu_mem_usage_list.append(cpu_mem)
|
|
self.gpu_mem_usage_list.append(gpu_mem)
|
|
|
|
def _calculate_average(self, values_list):
|
|
"""安全计算平均值，避免除零错误"""
|
|
return sum(values_list) / len(values_list) if values_list else 0
|
|
|
|
def report(self, logger):
|
|
"""在训练结束时输出汇总统计"""
|
|
if not self.start_time or not self.end_time:
|
|
logger.warning("TrainingStats: start/end time not recorded properly.")
|
|
return
|
|
|
|
total_time = self.end_time - self.start_time
|
|
avg_gpu_mem = self._calculate_average(self.gpu_mem_usage_list)
|
|
avg_cpu_mem = self._calculate_average(self.cpu_mem_usage_list)
|
|
avg_train_time = self._calculate_average(self.train_time_list)
|
|
avg_infer_time = self._calculate_average(self.infer_time_list)
|
|
iters_per_sec = self.total_iters / total_time if total_time > 0 else 0
|
|
|
|
logger.info("===== Training Summary =====")
|
|
logger.info(f"Total training time: {total_time:.2f} s")
|
|
logger.info(f"Total iterations: {self.total_iters}")
|
|
logger.info(f"Average iterations per second: {iters_per_sec:.2f}")
|
|
logger.info(f"Average GPU Memory Usage: {avg_gpu_mem:.2f} MB")
|
|
logger.info(f"Average CPU Memory Usage: {avg_cpu_mem:.2f} MB")
|
|
if avg_train_time:
|
|
logger.info(f"Average training step time: {avg_train_time * 1000:.2f} ms")
|
|
if avg_infer_time:
|
|
logger.info(f"Average inference step time: {avg_infer_time * 1000:.2f} ms")
|
|
|
|
|
|
class Trainer:
|
|
"""模型训练器，负责整个训练流程的管理"""
|
|
|
|
def __init__(
|
|
self,
|
|
model,
|
|
loss,
|
|
optimizer,
|
|
train_loader,
|
|
val_loader,
|
|
test_loader,
|
|
scaler,
|
|
args,
|
|
lr_scheduler=None,
|
|
):
|
|
# 设备和基本参数
|
|
self.device = args["basic"]["device"]
|
|
train_args = args["train"]
|
|
|
|
# 模型和训练相关组件
|
|
self.model = model
|
|
self.loss = loss
|
|
self.optimizer = optimizer
|
|
self.lr_scheduler = lr_scheduler
|
|
|
|
# 数据加载器
|
|
self.train_loader = train_loader
|
|
self.val_loader = val_loader
|
|
self.test_loader = test_loader
|
|
|
|
# 数据处理工具
|
|
self.scaler = scaler
|
|
self.args = train_args
|
|
|
|
# 统计信息
|
|
self.train_per_epoch = len(train_loader)
|
|
self.val_per_epoch = len(val_loader) if val_loader else 0
|
|
|
|
# 初始化路径、日志和统计
|
|
self._initialize_paths(train_args)
|
|
self._initialize_logger(train_args)
|
|
self._initialize_stats()
|
|
|
|
def _initialize_paths(self, args):
|
|
"""初始化模型保存路径"""
|
|
self.best_path = os.path.join(args["log_dir"], "best_model.pth")
|
|
self.best_test_path = os.path.join(args["log_dir"], "best_test_model.pth")
|
|
self.loss_figure_path = os.path.join(args["log_dir"], "loss.png")
|
|
|
|
def _initialize_logger(self, args):
|
|
"""初始化日志记录器"""
|
|
if not os.path.isdir(args["log_dir"]) and not args["debug"]:
|
|
os.makedirs(args["log_dir"], exist_ok=True)
|
|
self.logger = get_logger(
|
|
args["log_dir"], name=self.model.__class__.__name__, debug=args["debug"]
|
|
)
|
|
self.logger.info(f"Experiment log path in: {args['log_dir']}")
|
|
|
|
def _initialize_stats(self):
|
|
"""初始化统计信息记录器"""
|
|
self.stats = TrainingStats(device=self.device)
|
|
|
|
def _run_epoch(self, epoch, dataloader, mode):
|
|
"""运行一个训练/验证/测试epoch"""
|
|
# 设置模型模式和是否进行优化
|
|
if mode == "train":
|
|
self.model.train()
|
|
optimizer_step = True
|
|
else:
|
|
self.model.eval()
|
|
optimizer_step = False
|
|
|
|
# 初始化变量
|
|
total_loss = 0
|
|
epoch_time = time.time()
|
|
y_pred, y_true = [], []
|
|
|
|
# 训练/验证循环
|
|
with torch.set_grad_enabled(optimizer_step):
|
|
progress_bar = tqdm(
|
|
enumerate(dataloader),
|
|
total=len(dataloader),
|
|
desc=f"{mode.capitalize()} Epoch {epoch}"
|
|
)
|
|
|
|
for _, (data, target) in progress_bar:
|
|
# 记录步骤开始时间
|
|
start_time = time.time()
|
|
|
|
# 前向传播
|
|
label = target[..., : self.args["output_dim"]]
|
|
output = self.model(data).to(self.device)
|
|
loss = self.loss(output, label)
|
|
|
|
# 反归一化
|
|
d_output = self.scaler.inverse_transform(output)
|
|
d_label = self.scaler.inverse_transform(label)
|
|
|
|
# 反向传播和优化（仅在训练模式）
|
|
if optimizer_step and self.optimizer is not None:
|
|
self.optimizer.zero_grad()
|
|
loss.backward()
|
|
|
|
# 梯度裁剪（如果需要）
|
|
if self.args["grad_norm"]:
|
|
torch.nn.utils.clip_grad_norm_(
|
|
self.model.parameters(), self.args["max_grad_norm"]
|
|
)
|
|
self.optimizer.step()
|
|
|
|
# 反归一化的loss
|
|
d_loss = self.loss(d_output, d_label)
|
|
|
|
# 记录步骤时间和内存使用
|
|
step_time = time.time() - start_time
|
|
self.stats.record_step_time(step_time, mode)
|
|
|
|
# 累积损失和预测结果
|
|
total_loss += d_loss.item()
|
|
y_pred.append(d_output.detach().cpu())
|
|
y_true.append(d_label.detach().cpu())
|
|
|
|
# 更新进度条
|
|
progress_bar.set_postfix(loss=d_loss.item())
|
|
|
|
# 合并所有批次的预测结果
|
|
y_pred = torch.cat(y_pred, dim=0)
|
|
y_true = torch.cat(y_true, dim=0)
|
|
|
|
# 计算平均损失
|
|
avg_loss = total_loss / len(dataloader)
|
|
|
|
# 计算并记录指标
|
|
mae, rmse, mape = all_metrics(
|
|
y_pred, y_true, self.args["mae_thresh"], self.args["mape_thresh"]
|
|
)
|
|
self.logger.info(
|
|
f"Epoch #{epoch:02d}: {mode.capitalize():<5} MAE:{mae:5.2f} | RMSE:{rmse:5.2f} | MAPE:{mape:7.4f} | Time: {time.time() - epoch_time:.2f} s"
|
|
)
|
|
|
|
# 记录内存使用情况
|
|
self.stats.record_memory_usage()
|
|
|
|
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):
|
|
"""执行完整的训练流程"""
|
|
# 初始化最佳模型和损失记录
|
|
best_model, best_test_model = None, None
|
|
best_loss, best_test_loss = float("inf"), float("inf")
|
|
not_improved_count = 0
|
|
|
|
# 开始训练
|
|
self.stats.start_training()
|
|
self.logger.info("Training process started")
|
|
|
|
# 训练循环
|
|
for epoch in range(1, self.args["epochs"] + 1):
|
|
# 训练、验证和测试一个epoch
|
|
train_epoch_loss = self.train_epoch(epoch)
|
|
val_epoch_loss = self.val_epoch(epoch)
|
|
test_epoch_loss = self.test_epoch(epoch)
|
|
|
|
# 检查梯度爆炸
|
|
if train_epoch_loss > 1e6:
|
|
self.logger.warning("Gradient explosion detected. Ending...")
|
|
break
|
|
|
|
# 更新最佳验证模型
|
|
if val_epoch_loss < best_loss:
|
|
best_loss = val_epoch_loss
|
|
not_improved_count = 0
|
|
best_model = copy.deepcopy(self.model.state_dict())
|
|
self.logger.info("Best validation model saved!")
|
|
else:
|
|
not_improved_count += 1
|
|
|
|
# 检查早停条件
|
|
if self._should_early_stop(not_improved_count):
|
|
break
|
|
|
|
# 更新最佳测试模型
|
|
if test_epoch_loss < best_test_loss:
|
|
best_test_loss = test_epoch_loss
|
|
best_test_model = copy.deepcopy(self.model.state_dict())
|
|
|
|
# 保存最佳模型
|
|
if not self.args["debug"]:
|
|
self._save_best_models(best_model, best_test_model)
|
|
|
|
# 结束训练并输出统计信息
|
|
self.stats.end_training()
|
|
self.stats.report(self.logger)
|
|
|
|
# 最终评估
|
|
self._finalize_training(best_model, best_test_model)
|
|
|
|
# 输出模型参数量
|
|
self._log_model_params()
|
|
|
|
def _should_early_stop(self, not_improved_count):
|
|
"""检查是否满足早停条件"""
|
|
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."
|
|
)
|
|
return True
|
|
return False
|
|
|
|
def _save_best_models(self, best_model, best_test_model):
|
|
"""保存最佳模型到文件"""
|
|
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}"
|
|
)
|
|
|
|
def _log_model_params(self):
|
|
"""输出模型可训练参数数量"""
|
|
try:
|
|
total_params = sum(
|
|
p.numel() for p in self.model.parameters() if p.requires_grad
|
|
)
|
|
self.logger.info(f"Trainable params: {total_params}")
|
|
except Exception:
|
|
pass
|
|
|
|
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)
|
|
|
|
@staticmethod
|
|
def test(model, args, data_loader, scaler, logger, path=None):
|
|
"""对模型进行评估并输出性能指标"""
|
|
# 加载模型检查点（如果提供了路径）
|
|
if path:
|
|
checkpoint = torch.load(path)
|
|
model.load_state_dict(checkpoint["state_dict"])
|
|
model.to(args["basic"]["device"])
|
|
|
|
# 设置为评估模式
|
|
model.eval()
|
|
|
|
# 收集预测和真实标签
|
|
y_pred, y_true = [], []
|
|
|
|
# 不计算梯度的情况下进行预测
|
|
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)
|
|
|
|
# 合并所有批次的预测结果
|
|
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)
|
|
|
|
# 计算并记录每个时间步的指标
|
|
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}"
|
|
)
|
|
|
|
# 计算并记录平均指标
|
|
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))
|