增加学习率调度方式

configs/compare.yml

@@ -15,8 +15,8 @@ base:
 
 # 模型配置
 models:
-  backbone: 'resnet18'
-  channel_ratio: 0.75
+  backbone: 'resnet34'
+  channel_ratio: 1.0
 
 # 训练参数
 training:
@@ -29,11 +29,14 @@ training:
   lr_step: 10                         # 学习率调整间隔（epoch）
   lr_decay: 0.98                      # 学习率衰减率
   weight_decay: 0.0005                  # 权重衰减
-  scheduler: "cosine_annealing"       # 学习率调度器（可选：cosine_annealing/step/none）
+  scheduler: "cosine"                 # 学习率调度器（可选：cosine/cosine_warm/step/None）
   num_workers: 32                     # 数据加载线程数
-  checkpoints: "./checkpoints/resnet18_test/" # 模型保存目录
+  checkpoints: "./checkpoints/resnet34_20250612_scale=1.0/" # 模型保存目录
   restore: false
   restore_model: "resnet18_test/epoch_600.pth" # 模型恢复路径
+  cosine_t_0: 10                       # 初始周期长度
+  cosine_t_mult: 1                      # 周期长度倍率
+  cosine_eta_min: 0.00001               # 最小学习率
 
 # 验证参数
 validation:

configs/test.yml

@@ -8,13 +8,13 @@ base:
   log_level: "info"                   # 日志级别（debug/info/warning/error）
   embedding_size: 256                 # 特征维度
   pin_memory: true                    # 是否启用pin_memory
-  distributed: true                  # 是否启用分布式训练
+  distributed: false                  # 是否启用分布式训练
 
 # 模型配置
 models:
   backbone: 'resnet18'
-  channel_ratio: 1.0
-  model_path: "./checkpoints/resnet18_scatter_6.2/best.pth"
+  channel_ratio: 0.75
+  model_path: "./checkpoints/resnet18_0515/best.pth"
   half: false                        # 是否启用半精度测试（fp16）
 
 # 数据配置
@@ -22,9 +22,9 @@ data:
   group_test: False                 # 数据集名称（示例用，可替换为实际数据集）
   test_batch_size: 128               # 训练批次大小
   num_workers: 32                      # 数据加载线程数
-  test_dir: "../data_center/scatter/"   # 验证数据集根目录
+  test_dir: "../data_center/contrast_learning/contrast_test_data"   # 验证数据集根目录
   test_group_json: "../data_center/contrast_learning/model_test_data/test/inner_group_pairs.json"
-  test_list: "../data_center/scatter/val_pair.txt"
+  test_list: "../data_center/contrast_learning/contrast_test_data/test_pair.txt"
 
 transform:
   img_size: 224                     # 图像尺寸

configs/transform.yml

@@ -0,0 +1,27 @@
+# configs/transform.yml
+# pth转换onnx配置文件
+
+# 基础配置
+base:
+  experiment_name: "model_comparison"  # 实验名称（用于结果保存目录）
+  seed: 42                            # 随机种子（保证可复现性）
+  device: "cuda"                      # 训练设备（cuda/cpu）
+  log_level: "info"                   # 日志级别（debug/info/warning/error）
+  embedding_size: 256                 # 特征维度
+  pin_memory: true                    # 是否启用pin_memory
+  distributed: true                  # 是否启用分布式训练
+
+
+# 模型配置
+models:
+  backbone: 'resnet50'
+  channel_ratio: 1.0
+  model_path: "../checkpoints/resnet50_0519/best.pth"
+  onnx_model: "../checkpoints/resnet50_0519/best.onnx"
+  rknn_model: "../checkpoints/resnet50_0519/best.rknn"
+
+# 日志与监控
+logging:
+  logging_dir: "./logs"                  # 日志保存目录
+  tensorboard: true                   # 是否启用TensorBoard
+  checkpoint_interval: 30             # 检查点保存间隔（epoch）

configs/utils.py

@@ -1,4 +1,4 @@
-from model import (resnet18, mobilevit_s, MobileNetV3_Small, MobileNetV3_Large, mobilenet_v1,
+from model import (resnet18, resnet34, resnet50, mobilevit_s, MobileNetV3_Small, MobileNetV3_Large, mobilenet_v1,
                    PPLCNET_x1_0, PPLCNET_x0_5, PPLCNET_x2_5)
 from timm.models import vit_base_patch16_224 as vit_base_16
 from model.metric import ArcFace, CosFace
@@ -14,6 +14,8 @@ class trainer_tools:
     def get_backbone(self):
         backbone_mapping = {
             'resnet18': lambda: resnet18(scale=self.conf['models']['channel_ratio']),
+            'resnet34': lambda: resnet34(scale=self.conf['models']['channel_ratio']),
+            'resnet50': lambda: resnet50(scale=self.conf['models']['channel_ratio']),
             'mobilevit_s': lambda: mobilevit_s(),
             'mobilenetv3_small': lambda: MobileNetV3_Small(),
             'PPLCNET_x1_0': lambda: PPLCNET_x1_0(),
@@ -54,3 +56,24 @@ class trainer_tools:
             )
         }
         return optimizer_mapping
+
+    def get_scheduler(self, optimizer):
+        scheduler_mapping = {
+            'step': lambda: optim.lr_scheduler.StepLR(
+                optimizer,
+                step_size=self.conf['training']['lr_step'],
+                gamma=self.conf['training']['lr_decay']
+            ),
+            'cosine': lambda: optim.lr_scheduler.CosineAnnealingLR(
+                optimizer,
+                T_max=self.conf['training']['epochs'],
+                eta_min=self.conf['training']['cosine_eta_min']
+            ),
+            'cosine_warm': lambda: optim.lr_scheduler.CosineAnnealingWarmRestarts(
+                optimizer,
+                T_0=self.conf['training'].get('cosine_t_0', 10),
+                T_mult=self.conf['training'].get('cosine_t_mult', 1),
+                eta_min=self.conf['training'].get('cosine_eta_min', 0)
+            )
+        }
+        return scheduler_mapping