增加说明

.gitignore

@@ -302,3 +302,5 @@ Temporary Items
 runs/
 *.pt
 *.cache
+.vscode/
+*.json

README.md

@@ -2,7 +2,34 @@
 
 毕业设计：基于YOLO和图像融合技术的无人机检测系统及安全性研究
 
-Linux 运行训练
+Linux 运行联邦训练
 ```bash
-nohup python -u yolov8_fed.py >> runtime.log 2>&1 &
+cd federated_learning
+```
+
+```bash
+nohup python -u yolov8_fed.py > runtime.log 2>&1 &
+```
+
+Linux 运行集中训练
+```bash
+cd yolov8
+```
+
+```bash
+nohup python -u yolov8_train.py > runtime.log 2>&1 &
+```
+
+实时监控日志文件
+```bash
+tail -f runtime.log
+```
+
+运行图像融合配准代码
+```bash
+cd image_fusion
+```
+
+```bash
+python Image_Registration_test.py
 ```

federated_learning/yolov8.yaml

@@ -0,0 +1,49 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLOv8 object detection model with P3/8 - P5/32 outputs
+# Model docs: https://docs.ultralytics.com/models/yolov8
+# Task docs: https://docs.ultralytics.com/tasks/detect
+
+# Parameters
+nc: 1 # number of classes
+scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.33, 0.25, 1024] # YOLOv8n summary: 129 layers, 3157200 parameters, 3157184 gradients, 8.9 GFLOPS
+  s: [0.33, 0.50, 1024] # YOLOv8s summary: 129 layers, 11166560 parameters, 11166544 gradients, 28.8 GFLOPS
+  m: [0.67, 0.75, 768] # YOLOv8m summary: 169 layers, 25902640 parameters, 25902624 gradients, 79.3 GFLOPS
+  l: [1.00, 1.00, 512] # YOLOv8l summary: 209 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPS
+  x: [1.00, 1.25, 512] # YOLOv8x summary: 209 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPS
+
+# YOLOv8.0n backbone
+backbone:
+  # [from, repeats, module, args]
+  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
+  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
+  - [-1, 3, C2f, [128, True]]
+  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
+  - [-1, 6, C2f, [256, True]]
+  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
+  - [-1, 6, C2f, [512, True]]
+  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
+  - [-1, 3, C2f, [1024, True]]
+  - [-1, 1, SPPF, [1024, 5]] # 9
+
+# YOLOv8.0n head
+head:
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
+  - [-1, 3, C2f, [512]] # 12
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
+  - [-1, 3, C2f, [256]] # 15 (P3/8-small)
+
+  - [-1, 1, Conv, [256, 3, 2]]
+  - [[-1, 12], 1, Concat, [1]] # cat head P4
+  - [-1, 3, C2f, [512]] # 18 (P4/16-medium)
+
+  - [-1, 1, Conv, [512, 3, 2]]
+  - [[-1, 9], 1, Concat, [1]] # cat head P5
+  - [-1, 3, C2f, [1024]] # 21 (P5/32-large)
+
+  - [[15, 18, 21], 1, Detect, [nc]] # Detect(P3, P4, P5)

federated_learning/yolov8_fed.py

@@ -1,6 +1,9 @@
 import glob
 import os
 from pathlib import Path
+import json
+from pydoc import cli
+from threading import local
 
 import yaml
 from ultralytics import YOLO
@@ -15,121 +18,235 @@ def federated_avg(global_model, client_weights):
     total_samples = sum(n for _, n in client_weights)
     if total_samples == 0:
         raise ValueError("Total number of samples must be positive.")
-    
+
+    # DEBUG: global_dict
+    # print(global_model)
+
     # 获取YOLO底层PyTorch模型参数
     global_dict = global_model.model.state_dict()
     # 提取所有客户端的 state_dict 和对应样本数
     state_dicts, sample_counts = zip(*client_weights)
-    
-    for key in global_dict:
-        # 对每一层参数取平均
-        # if global_dict[key].data.dtype == torch.float32:
-        #     global_dict[key].data = torch.stack(
-        #         [w[key].float() for w in client_weights], 0
-        #     ).mean(0)
-        
-        # 加权平均
-        if global_dict[key].dtype == torch.float32:  # 只聚合浮点型参数
-            # 跳过 BatchNorm 层的统计量
-            if any(x in key for x in ['running_mean', 'running_var', 'num_batches_tracked']):
-                continue
-            # 按照样本数加权求和
-            weighted_tensors = [sd[key].float() * (n / total_samples)
-                                for sd, n in zip(state_dicts, sample_counts)]
-            global_dict[key] = torch.stack(weighted_tensors, dim=0).sum(dim=0)
-    
-    # 解决模型参数不匹配问题
-    try:
-        # 加载回YOLO模型
-        global_model.model.load_state_dict(global_dict)
-    except RuntimeError as e:
-        print('Ignoring "' + str(e) + '"')
-    
-    # 添加调试输出
-    print("\n=== 参数聚合检查 ===")
-    
-    # 选取一个典型参数层
-    # sample_key = list(global_dict.keys())[10]
-    # original = global_dict[sample_key].data.mean().item()
-    # aggregated = torch.stack([w[sample_key] for w in client_weights]).mean().item()
-    # print(f"参数层 '{sample_key}' 变化: {original:.4f} → {aggregated:.4f}")
-    # print(f"客户端参数差异: {[w[sample_key].mean().item() for w in client_weights]}")
-    
-    # 随机选取一个非统计量层进行对比
-    sample_key = next(k for k in global_dict if 'running_' not in k)
-    aggregated_mean = global_dict[sample_key].mean().item()
-    client_means = [sd[sample_key].float().mean().item() for sd in state_dicts]
-    print(f"layer: '{sample_key}' Mean after aggregation: {aggregated_mean:.6f}")
-    print(f"The average value of the layer for each client: {client_means}")
-    
+
+    # 克隆参数并脱离计算图
+    global_dict_copy = {
+        k: v.clone().detach().requires_grad_(False) for k, v in global_dict.items()
+    }
+
+    # 聚合可训练且存在的参数
+    for key in global_dict_copy:
+        # if global_dict_copy[key].dtype != torch.float32:
+        #     continue
+        # if any(
+        #     x in key for x in ["running_mean", "running_var", "num_batches_tracked"]
+        # ):
+        #     continue
+        # 检查所有客户端是否包含当前键
+        all_clients_have_key = all(key in sd for sd in state_dicts)
+        if all_clients_have_key:
+            # 计算每个客户端的加权张量
+            # weighted_tensors = [
+            #     client_state[key].float() * (sample_count / total_samples)
+            #     for client_state, sample_count in zip(state_dicts, sample_counts)
+            # ]
+            weighted_tensors = []
+            for client_state, sample_count in zip(state_dicts, sample_counts):
+                weight = sample_count / total_samples  # 计算权重
+                weighted_tensor = client_state[key].float() * weight  # 加权张量
+                weighted_tensors.append(weighted_tensor)
+            # 聚合加权张量并更新全局参数
+            global_dict_copy[key] = torch.stack(weighted_tensors, dim=0).sum(dim=0)
+
+        # else:
+        #     print(f"错误: 键 {key} 在部分客户端缺失，已保留全局参数")
+        # 终止训练或记录日志
+        # raise KeyError(f"键 {key} 缺失")
+
+    # 加载回YOLO模型
+    global_model.model.load_state_dict(global_dict_copy, strict=True)
+
+    # global_model.model.train()
+    # with torch.no_grad():
+    #     global_model.model.load_state_dict(global_dict_copy, strict=True)
+
+    # 定义多个关键层
+    MONITOR_KEYS = [
+        "model.0.conv.weight",
+        "model.1.conv.weight",
+        "model.3.conv.weight",
+        "model.5.conv.weight",
+        "model.7.conv.weight",
+        "model.9.cv1.conv.weight",
+        "model.12.cv1.conv.weight",
+        "model.15.cv1.conv.weight",
+        "model.18.cv1.conv.weight",
+        "model.21.cv1.conv.weight",
+        "model.22.dfl.conv.weight",
+    ]
+
+    with open("aggregation_check.txt", "a") as f:
+        f.write("\n=== 参数聚合检查 ===\n")
+    for key in MONITOR_KEYS:
+        # if key not in global_dict:
+        #     continue
+        # if not all(key in sd for sd in state_dicts):
+        #     continue
+
+        # 计算聚合后均值
+        aggregated_mean = global_dict[key].mean().item()
+
+        # 计算各客户端均值
+        client_means = [sd[key].float().mean().item() for sd in state_dicts]
+        with open("aggregation_check.txt", "a") as f:
+            f.write(f"层 '{key}' 聚合后均值: {aggregated_mean:.6f}\n")
+            f.write(f"各客户端该层均值差异: {[f'{cm:.6f}' for cm in client_means]}\n")
+            f.write(f"客户端最大差异: {max(client_means) - min(client_means):.6f}\n\n")
+
     return global_model
 
 
 # ------------ 修改训练流程 ------------
 def federated_train(num_rounds, clients_data):
+    # ========== 初始化指标记录 ==========
+    metrics = {
+        "round": [],
+        "val_mAP": [],  # 每轮验证集mAP
+        # "train_loss": [],  # 每轮平均训练损失
+        "client_mAPs": [],  # 各客户端本地模型在验证集上的mAP
+        "communication_cost": [],  # 每轮通信开销（MB）
+    }
     # 初始化全局模型
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    global_model = YOLO("../yolov8n.pt").to(device)
-    # 设置类别数
-    global_model.model.nc = 1
-    
+    global_model = (
+        YOLO("/home/image1325/DATA/Graduation-Project/federated_learning/yolov8n.yaml")
+        .load("/home/image1325/DATA/Graduation-Project/federated_learning/yolov8n.pt")
+        .to(device)
+    )
+    global_model.model.model[-1].nc = 1  # 设置检测类别数为1
+    # global_model.model.train.ema.enabled = False
+
+    # 克隆全局模型
+    local_model = copy.deepcopy(global_model)
+
     for _ in range(num_rounds):
         client_weights = []
-        
+        # 各客户端的训练损失
+        # client_losses = []
+
+        # DEBUG: 检查全局模型参数
+        # global_dict = global_model.model.state_dict()
+        # print(global_dict.keys())
+
         # 每个客户端本地训练
         for data_path in clients_data:
             # 统计本地训练样本数
-            with open(data_path, 'r') as f:
+            with open(data_path, "r") as f:
                 config = yaml.safe_load(f)
             #  Resolve img_dir relative to the YAML file's location
             yaml_dir = os.path.dirname(data_path)
-            img_dir = os.path.join(yaml_dir, config.get('train', data_path))  # 从配置文件中获取图像目录
-            
+            img_dir = os.path.join(
+                yaml_dir, config.get("train", data_path)
+            )  # 从配置文件中获取图像目录
+
             # print(f"Image directory: {img_dir}")
-            num_samples = (len(glob.glob(os.path.join(img_dir, '*.jpg'))) +
-                           len(glob.glob(os.path.join(img_dir, '*.png'))))
+            num_samples = (
+                len(glob.glob(os.path.join(img_dir, "*.jpg")))
+                + len(glob.glob(os.path.join(img_dir, "*.png")))
+                + len(glob.glob(os.path.join(img_dir, "*.jpeg")))
+            )
             # print(f"Number of images: {num_samples}")
-            
-            # 克隆全局模型
-            local_model = copy.deepcopy(global_model)
-            
+
+            local_model.model.load_state_dict(
+                global_model.model.state_dict(), strict=True
+            )
+
             # 本地训练（保持你的原有参数设置）
             local_model.train(
+                name=f"train{_ + 1}",  # 当前轮次
                 data=data_path,
-                epochs=16,  # 每轮本地训练1个epoch
-                save_period=16,
-                imgsz=640,  # 图像大小
+                # model=local_model,
+                epochs=16,  # 每轮本地训练多少个epoch
+                # save_period=16,
+                imgsz=768,  # 图像大小
                 verbose=False,  # 关闭冗余输出
-                batch=-1
+                batch=-1,  # 批大小
+                workers=6,  # 工作线程数
             )
-            
+
+            # 记录客户端训练损失
+            # client_loss = results.results_dict['train_loss']
+            # client_losses.append(client_loss)
+
             # 收集模型参数及样本数
-            client_weights.append((copy.deepcopy(local_model.model.state_dict()), num_samples))
-        
+            client_weights.append((local_model.model.state_dict(), num_samples))
+
         # 聚合参数更新全局模型
         global_model = federated_avg(global_model, client_weights)
-        print(f"Round {_ + 1}/{num_rounds} completed.")
-    return global_model
+
+        # DEBUG: 检查全局模型参数
+        # keys = global_model.model.state_dict().keys()
+
+        # ========== 评估全局模型 ==========
+        # 复制全局模型以避免在评估时修改参数
+        val_model = copy.deepcopy(global_model)
+        # 评估全局模型在验证集上的性能
+        with torch.no_grad():
+            val_results = val_model.val(
+                data="/mnt/DATA/uav_dataset_old/UAVdataset/fed_data.yaml",  # 指定验证集配置文件
+                imgsz=768,  # 图像大小
+                batch=16,  # 批大小
+                verbose=False,  # 关闭冗余输出
+            )
+        # 丢弃评估模型
+        del val_model
+
+        # DEBUG: 检查全局模型参数
+        # if keys != global_model.model.state_dict().keys():
+        #     print("模型参数不一致！")
+
+        val_mAP = val_results.box.map  # 获取mAP@0.5
+
+        # 计算平均训练损失
+        # avg_train_loss = sum(client_losses) / len(client_losses)
+
+        # 计算通信开销（假设传输全部模型参数）
+        model_size = sum(p.numel() * 4 for p in global_model.model.parameters()) / (
+            1024**2
+        )  # MB
+
+        # 记录到指标容器
+        metrics["round"].append(_ + 1)
+        metrics["val_mAP"].append(val_mAP)
+        # metrics['train_loss'].append(avg_train_loss)
+        metrics["communication_cost"].append(model_size)
+        # 打印当前轮次结果
+        with open("aggregation_check.txt", "a") as f:
+            f.write(f"\n[Round {_ + 1}/{num_rounds}]\n")
+            f.write(f"Validation mAP@0.5: {val_mAP:.4f}\n")
+            # f.write(f"Average Train Loss: {avg_train_loss:.4f}")
+            f.write(f"Communication Cost: {model_size:.2f} MB\n\n")
+
+    return global_model, metrics
 
 
-# ------------ 使用示例 ------------
 if __name__ == "__main__":
     # 联邦训练配置
     clients_config = [
-        "/root/autodl-tmp/dataset/train1/train1.yaml",  # 客户端1数据路径
-        "/root/autodl-tmp/dataset/train2/train2.yaml"  # 客户端2数据路径
+        "/mnt/DATA/uav_fed/train1/train1.yaml",  # 客户端1数据路径
+        "/mnt/DATA/uav_fed/train2/train2.yaml",  # 客户端2数据路径
     ]
-    
+
+    # 使用本地数据集进行测试
+    # clients_config = [
+    #     "/home/image1325/DATA/Graduation-Project/dataset/train1/train1.yaml",
+    #     "/home/image1325/DATA/Graduation-Project/dataset/train2/train2.yaml",
+    # ]
+
     # 运行联邦训练
-    final_model = federated_train(num_rounds=10, clients_data=clients_config)
-    
+    final_model, metrics = federated_train(num_rounds=10, clients_data=clients_config)
+
     # 保存最终模型
     final_model.save("yolov8n_federated.pt")
     # final_model.export(format="onnx")  # 导出为ONNX格式
-    
-    # 检查1：确认模型保存
-    # assert Path("yolov8n_federated.onnx").exists(), "模型导出失败"
-    
-    # 检查2：验证预测功能
-    # results = final_model.predict("../dataset/val/images/VS_P65.jpg", save=True)
-    # assert len(results[0].boxes) > 0, "预测结果异常"
+
+    with open("metrics.json", "w") as f:
+        json.dump(metrics, f, indent=4)