bak

14.训练模型-姿态估计.py

@@ -1,7 +1,7 @@
 from ultralytics import YOLO
 import torch
 # 加载预训练模型
-# model = YOLO('models/yolov8n-pose.pt')
+model = YOLO('models/yolov8n-pose.pt')
 # model = YOLO('models/yolo11x-pose.pt')
 # model = YOLO('models/yolov8n-pose.pt')
 # model = YOLO('runs/pose/train2/weights/last.pt')
@@ -22,38 +22,36 @@ import torch
 # save_period: 每隔多少轮保存一次模型
 # save_dir: 训练结果保存路径
 # weights: 预训练模型路径
-# model.info()
-# model.train(data='./dataset1/train.yaml', epochs=300, imgsz=640, batch=32, device='cpu')
+model.info()
+model.train(data='./dataset1/train.yaml', epochs=300, imgsz=640, batch=32, device=0)
 
 
 
-
-# 检查MPS可用性
-print(f"MPS available: {torch.backends.mps.is_available()}")
-print(f"MPS built: {torch.backends.mps.is_built()}")
-
-model = YOLO('models/yolov8n-pose.pt')
-
-# MPS优化配置
-model.train(
-    data='./dataset1/train.yaml',
-    epochs=300,
-    imgsz=320,  # M2 Pro上建议减小尺寸
-    batch=8,    # 根据内存调整
-    device='mps',  # 使用Apple Metal Performance Shaders
-    workers=2,  # MPS下建议2个worker
-    patience=50,
-    lr0=0.01,
-    lrf=0.01,
-    momentum=0.9,  # MPS上动量稍小
-    weight_decay=0.0005,
-    warmup_epochs=5.0,
-    box=7.5,
-    pose=1.0,   # 增加姿态损失权重
-    kobj=1.5,
-    save=True,
-    exist_ok=True,
-    verbose=True,
-    amp=False   # MPS上关闭自动混合精度
-)
-print("训练完成")
+#
+# # 检查MPS可用性
+# print(f"MPS available: {torch.backends.mps.is_available()}")
+# print(f"MPS built: {torch.backends.mps.is_built()}")
+#
+# model = YOLO('models/yolov8n-pose.pt')
+#
+# # MPS优化配置
+# model.train( data='./dataset1/train.yaml', epochs=300,
+#     imgsz=320,  # M2 Pro上建议减小尺寸
+#     batch=8,    # 根据内存调整
+#     device='mps',  # 使用Apple Metal Performance Shaders
+#     workers=2,  # MPS下建议2个worker
+#     patience=50,
+#     lr0=0.01,
+#     lrf=0.01,
+#     momentum=0.9,  # MPS上动量稍小
+#     weight_decay=0.0005,
+#     warmup_epochs=5.0,
+#     box=7.5,
+#     pose=1.0,   # 增加姿态损失权重
+#     kobj=1.5,
+#     save=True,
+#     exist_ok=True,
+#     verbose=True,
+#     amp=False   # MPS上关闭自动混合精度
+# )
+# print("训练完成")