diff --git a/src/mmap_drl_nav/main.py b/src/mmap_drl_nav/main.py index 3756dc661d..93020a92f7 100644 --- a/src/mmap_drl_nav/main.py +++ b/src/mmap_drl_nav/main.py @@ -1,59 +1,135 @@ import torch import torch.nn as nn from torch.optim import Adam -from utils.dataloader import get_dataloader -from models.perception_module import PerceptionModule -from models.attention_module import CrossDomainAttention -from models.decision_module import DecisionModule -from models.sagm import SelfAssessmentGradientModel +from utils.dataloader import get_dataloader # 导入自定义的数据加载工具 +from models.perception_module import PerceptionModule # 导入感知模块 +from models.attention_module import CrossDomainAttention # 导入跨域注意力模块 +from models.decision_module import DecisionModule # 导入决策模块 +from models.sagm import SelfAssessmentGradientModel # 导入自评估梯度模型 + class IntegratedSystem: + """ + 集成决策系统类 + 整合感知、注意力、决策和自评估模块,形成端到端的智能决策系统 + """ + def __init__(self, device='cpu', state_dim=128, action_dim=2): - self.device = device + """ + 初始化集成系统 + Args: + device (str): 模型运行设备,默认CPU,可设为'cuda'使用GPU + state_dim (int): 状态特征维度,默认128 + action_dim (int): 动作输出维度,默认2 + """ + self.device = device # 保存设备信息 + # 初始化感知模块并移至指定设备 self.perception = PerceptionModule().to(self.device) + # 初始化跨域注意力模块(6个注意力块)并移至指定设备 self.attention = CrossDomainAttention(num_blocks=6).to(self.device) + # 初始化决策模块并移至指定设备 self.decision = DecisionModule().to(self.device) - self.sagm = SelfAssessmentGradientModel(input_dim=state_dim + action_dim).to(self.device) + # 初始化自评估梯度模型(输入维度为状态+动作维度)并移至指定设备 + self.sagm = SelfAssessmentGradientModel(input_dim=state_dim + action_dim).to(self.device) def forward(self, image, lidar_data, imu_data, action): + """ + 前向传播函数:整合各模块完成一次完整的前向计算 + Args: + image: 图像输入数据 + lidar_data: 激光雷达输入数据 + imu_data: 惯性测量单元输入数据 + action: 动作输入数据 + Returns: + policy: 策略输出 + value: 价值输出 + sagm_q_value: 自评估Q值输出 + """ + # 感知模块处理多源数据,输出场景信息、分割结果、里程计、障碍物和边界信息 scene_info, segmentation, odometry, obstacles, boundary = self.perception(imu_data, image, lidar_data) + # 跨域注意力模块融合感知模块输出的多维度特征 fused_features = self.attention(scene_info, segmentation, odometry, obstacles, boundary) - + + # 决策模块基于融合特征输出策略和价值估计 policy, value = self.decision(fused_features) - - sagm_q_value = self.sagm(fused_features, action) + + # 自评估梯度模型基于融合特征和动作计算Q值 + sagm_q_value = self.sagm(fused_features, action) return policy, value, sagm_q_value + def train_model(model, dataloader, optimizer, device, num_epochs=10): - model.train() - for epoch in range(num_epochs): - running_loss = 0.0 + """ + 模型训练函数 + Args: + model: 待训练的集成系统模型 + dataloader: 训练数据加载器 + optimizer: 优化器 + device: 训练设备 + num_epochs (int): 训练轮数,默认10 + """ + model.train() # 将模型设为训练模式(启用dropout、batchnorm等训练特性) + for epoch in range(num_epochs): # 遍历每个训练轮次 + running_loss = 0.0 # 初始化累计损失 + # 遍历数据加载器中的每个批次 for i, (image, lidar_data, imu_data, target_action) in enumerate(dataloader): - image, lidar_data, imu_data, target_action = image.to(device), lidar_data.to(device), imu_data.to(device), target_action.to(device) - optimizer.zero_grad() + # 将所有数据移至指定设备(CPU/GPU) + image, lidar_data, imu_data, target_action = image.to(device), lidar_data.to(device), imu_data.to( + device), target_action.to(device) + + optimizer.zero_grad() # 清空梯度缓存 + + # 前向传播:输入数据得到模型输出 policy_output, value_output, sagm_q_value = model(image, lidar_data, imu_data, target_action) - + + # 计算总损失:策略损失 + 价值损失 + 自评估Q值损失(均使用MSE损失) + # 价值输出和Q值输出与目标动作的求和结果做回归 loss = (nn.MSELoss()(policy_output, target_action) + nn.MSELoss()(value_output, target_action.sum(dim=1, keepdim=True)) + nn.MSELoss()(sagm_q_value, target_action.sum(dim=1, keepdim=True))) - loss.backward() - optimizer.step() - running_loss += loss.item() + + loss.backward() # 反向传播计算梯度 + optimizer.step() # 优化器更新模型参数 + + running_loss += loss.item() # 累计损失值 + + # 每10个批次打印一次训练信息 if i % 10 == 9: - print(f'Epoch [{epoch+1}/{num_epochs}], Batch [{i+1}/{len(dataloader)}], Loss: {running_loss / 10:.4f}') - running_loss = 0.0 - print('Training complete') + print( + f'Epoch [{epoch + 1}/{num_epochs}], Batch [{i + 1}/{len(dataloader)}], Loss: {running_loss / 10:.4f}') + running_loss = 0.0 # 重置累计损失 + print('Training complete') # 训练完成提示 + def test_model(model, dataloader, device): - model.eval() - total_loss = 0.0 + """ + 模型测试函数 + Args: + model: 已训练的集成系统模型 + dataloader: 测试数据加载器 + device: 测试设备 + """ + model.eval() # 将模型设为评估模式(禁用dropout、batchnorm等训练特性) + total_loss = 0.0 # 初始化总损失 + + # 禁用梯度计算(加速推理,节省内存) with torch.no_grad(): + # 遍历测试数据集 for image, lidar_data, imu_data, target_action in dataloader: - image, lidar_data, imu_data, target_action = image.to(device), lidar_data.to(device), imu_data.to(device), target_action.to(device) + # 将所有数据移至指定设备 + image, lidar_data, imu_data, target_action = image.to(device), lidar_data.to(device), imu_data.to( + device), target_action.to(device) + + # 前向传播得到模型输出 policy_output, value_output, sagm_q_value = model(image, lidar_data, imu_data, target_action) + + # 计算批次损失(与训练损失计算方式一致) loss = (nn.MSELoss()(policy_output, target_action) + nn.MSELoss()(value_output, target_action.sum(dim=1, keepdim=True)) + nn.MSELoss()(sagm_q_value, target_action.sum(dim=1, keepdim=True))) - total_loss += loss.item() + + total_loss += loss.item() # 累计测试损失 + + # 计算平均损失 avg_loss = total_loss / len(dataloader) - print(f'Test Average Loss: {avg_loss:.4f}') + print(f'Test Average Loss: {avg_loss:.4f}') # 打印测试平均损失 \ No newline at end of file