数据分析

.idea/CopilotChatHistory.xml

@@ -3,6 +3,12 @@
   <component name="CopilotChatHistory">
     <option name="conversations">
       <list>
+        <Conversation>
+          <option name="createTime" value="1752195523240" />
+          <option name="id" value="0197f6fde2a87e68b893b3a36dfc838f" />
+          <option name="title" value="新对话 2025年7月11日 08:58:43" />
+          <option name="updateTime" value="1752195523240" />
+        </Conversation>
         <Conversation>
           <option name="createTime" value="1752114061266" />
           <option name="id" value="0197f222dfd27515a3dbfea638532ee5" />

configs/compare.yml

@@ -15,25 +15,25 @@ base:
 
 # 模型配置
 models:
-  backbone: 'resnet34'
+  backbone: 'resnet18'
-  channel_ratio: 1.0
+  channel_ratio: 0.75
 
 # 训练参数
 training:
   epochs: 600                         # 总训练轮次
   batch_size: 128                     # 批次大小
-  lr: 0.001                           # 初始学习率
+  lr: 0.007                           # 初始学习率
   optimizer: "sgd"                    # 优化器类型
   metric: 'arcface'                   # 损失函数类型（可选：arcface/cosface/sphereface/softmax）
   loss: "cross_entropy"               # 损失函数类型（可选：cross_entropy/cross_entropy_smooth/center_loss/center_loss_smooth/arcface/cosface/sphereface/softmax）
   lr_step: 10                         # 学习率调整间隔（epoch）
-  lr_decay: 0.98                      # 学习率衰减率
+  lr_decay: 0.95                      # 学习率衰减率
   weight_decay: 0.0005                  # 权重衰减
   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/resnet34_20250612_scale=1.0/" # 模型保存目录
+  checkpoints: "./checkpoints/resnet18_20250717_scale=0.75_nosub/" # 模型保存目录
-  restore: false
+  restore: true
-  restore_model: "resnet18_test/epoch_600.pth" # 模型恢复路径
+  restore_model: "./checkpoints/resnet18_20250716_scale=0.75_nosub/best.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
@@ -49,8 +49,8 @@ data:
   train_batch_size: 128               # 训练批次大小
   val_batch_size: 128                 # 验证批次大小
   num_workers: 32                      # 数据加载线程数
-  data_train_dir: "../data_center/contrast_learning/data_base/train"   # 训练数据集根目录
+  data_train_dir: "../data_center/contrast_data/v2/train"   # 训练数据集根目录
-  data_val_dir: "../data_center/contrast_learning/data_base/val"   # 验证数据集根目录
+  data_val_dir: "../data_center/contrast_data/v2/val"   # 验证数据集根目录
 
 transform:
   img_size: 224                     # 图像尺寸
@@ -62,7 +62,7 @@ transform:
 
 # 日志与监控
 logging:
-  logging_dir: "./logs/resnet50_log"                  # 日志保存目录
+  logging_dir: "./logs/resnet18_scale=0.75_nosub_log"                  # 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/similar_analysis.yml

@@ -0,0 +1,49 @@
+# configs/similar_analysis.yml
+# 专为模型训练对比设计的配置文件
+# 支持对比不同训练策略（如蒸馏vs独立训练）
+
+# 基础配置
+base:
+  experiment_name: "model_comparison"  # 实验名称（用于结果保存目录）
+  device: "cuda"                      # 训练设备（cuda/cpu）
+  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: 'resnet18'
+  channel_ratio: 0.75
+  model_path: "../checkpoints/resnet18_1009/best.pth"
+
+heatmap:
+  feature_layer: "layer4"
+  show_heatmap:  true
+# 数据配置
+data:
+  dataset: "imagenet"                 # 数据集名称（示例用，可替换为实际数据集）
+  train_batch_size: 128               # 训练批次大小
+  val_batch_size: 8                 # 验证批次大小
+  num_workers: 32                      # 数据加载线程数
+  data_dir: "/home/lc/data_center/image_analysis/error_compare_subimg"
+  image_joint_pth: "/home/lc/data_center/image_analysis/error_compare_result"
+
+transform:
+  img_size: 224                     # 图像尺寸
+  img_mean: 0.5                     # 图像均值
+  img_std: 0.5                      # 图像方差
+  RandomHorizontalFlip: 0.5         # 随机水平翻转概率
+  RandomRotation: 180               # 随机旋转角度
+  ColorJitter: 0.5                   # 随机颜色抖动强度
+
+# 日志与监控
+logging:
+  logging_dir: "./logs/resnet18_scale=0.75_nosub_log"                  # 日志保存目录
+  tensorboard: true                   # 是否启用TensorBoard
+  checkpoint_interval: 30             # 检查点保存间隔（epoch）
+
+# 分布式训练（可选）
+distributed:
+  enabled: false                      # 是否启用分布式训练
+  backend: "nccl"                     # 分布式后端（nccl/gloo）

configs/sub_data.yml

@@ -4,23 +4,29 @@
 
 # 数据配置
 data:
-  source_dir: "../../data_center/contrast_data/total"                 # 数据集名称（示例用，可替换为实际数据集）
+  source_dir: "../../data_center/contrast_data/total"     # 数据集名称（示例用，可替换为实际数据集）
   train_dir: "../../data_center/contrast_data/v1/train"   # 训练数据集根目录
   val_dir: "../../data_center/contrast_data/v1/val"   # 验证数据集根目录
   data_extra_dir: "../../data_center/contrast_data/v1/extra"
   max_files_ratio: 0.1
   min_files: 10
   split_ratio: 0.9
+  combine_scr_dir: "../../data_center/contrast_data/v1/val"  # 合并数据集源目录·
+  combine_dst_dir: "../../data_center/contrast_data/v2/val"  # 合并数据集目标目录
 
 
 extend:
   extend_same_dir: true
   extend_extra: true
-  extend_extra_dir: "../../data_center/contrast_data/v1/extra"
+  extend_extra_dir: "../../data_center/contrast_data/v1/extra"  # 扩展测试集数据
   extend_train: true
-  extend_train_dir: "../../data_center/contrast_data/v1/train"
+  extend_train_dir: "../../data_center/contrast_data/v1/train"  # 训练接数据扩展
 
 limit:
   count_limit: true
   limit_count: 200
-  limit_dir: "../../data_center/contrast_data/v1/train"
+  limit_dir: "../../data_center/contrast_data/v1/train"  # 限制单个样本数量
+
+control:
+  combine: true   # 是否进行子类数据集合并
+  split: false    # 子类数据集拆解与扩增

configs/test.yml

@@ -8,7 +8,7 @@ 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: false                  # 是否启用分布式训练
+  distributed: false                  # 是否启用分布式训练   启用热力图时不能用分布式训练
 
 # 模型配置
 models:

configs/write_feature.yml

@@ -22,7 +22,7 @@ data:
   test_batch_size: 128                 # 验证批次大小
   num_workers: 32                      # 数据加载线程数
   half:  true                         # 是否启用半精度数据
-  img_dirs_path: "/personalDocument/lic/base+stlib"
+  img_dirs_path: "/home/lc/data_center/baseStlib/pic/stlib_base"  # base标准库图片存储路径
 #  img_dirs_path: "/home/lc/contrast_nettest/data/feature_json"
   xlsx_pth: false  # 过滤商品， 默认None不进行过滤
 
@@ -42,7 +42,7 @@ logging:
 
 save:
   json_bin: "../search_library/yunhedian_05-09.json"       # 保存整个json文件
-  json_path: "../feature_json/base+stlib/"    # 保存单个json文件路径
+  json_path: "/home/lc/data_center/baseStlib/feature_json/stlib_base"    # 保存单个json文件路径
   error_barcodes: "error_barcodes.txt"
   barcodes_statistics: "../search_library/barcodes_statistics.txt"
   create_single_json: true                  # 是否保存单个json文件

data_preprocessing/combine_sub_class.py

@@ -0,0 +1,25 @@
+import os
+import shutil
+
+
+def combine_dirs(conf):
+    source_root = conf['data']['combine_scr_dir']
+    target_root = conf['data']['combine_dst_dir']
+    for roots, dir_names, files in os.walk(source_root):
+        for dir_name in dir_names:
+            source_dir = os.path.join(roots, dir_name)
+            target_dir = os.path.join(target_root, dir_name.split('_')[0])
+            if not os.path.exists(target_dir):
+                os.mkdir(target_dir)
+            for filename in os.listdir(source_dir):
+                print(filename)
+                source_file = os.sep.join([source_dir, filename])
+                target_file = os.sep.join([target_dir, filename])
+                shutil.copy(source_file, target_file)
+            # print(f"已复制目录 {source_dir} 到 {target_dir}")
+
+
+# if __name__ == '__main__':
+#     source_root = r'scatter_mini'
+#     target_root = r'C:\Users\123\Desktop\scatter-1'
+#     # combine_dirs(conf)

data_preprocessing/sub_data_preprocessing.py

@@ -1,14 +1,18 @@
 from create_extra import split_subdirs
 from data_split import split_directory
 from extend import ImageExtendProcessor
+from combine_sub_class import combine_dirs
 import yaml
 
 
 def data_preprocessing(conf):
-    # split_subdirs(conf)
+    if conf['control']['split']:
-    # image_ex = ImageExtendProcessor(conf)
+        split_subdirs(conf)
-    # image_ex.control_number()
+        image_ex = ImageExtendProcessor(conf)
-    split_directory(conf)
+        image_ex.control_number()
+        split_directory(conf)
+    if conf['control']['combine']:
+        combine_dirs(conf)
 
 
 if __name__ == '__main__':

test_ori.py

@@ -17,7 +17,7 @@ from configs import trainer_tools
 import yaml
 from datetime import datetime
 
-with open('configs/test.yml', 'r') as f:
+with open('../configs/test.yml', 'r') as f:
     conf = yaml.load(f, Loader=yaml.FullLoader)
 
 # Constants from config

tools/image_joint.py

@@ -3,7 +3,7 @@ from tools.getHeatMap import cal_cam
 import os
 
 
-def merge_imgs(img1_path, img2_path, conf, similar=None, label=None, cam=None):
+def merge_imgs(img1_path, img2_path, conf, similar=None, label=None, cam=None, save_path=None):
     save = True
     position = (50, 50)  # 文字的左上角坐标
     color = (255, 0, 0)  # 红色文字，格式为 RGB
@@ -11,22 +11,25 @@ def merge_imgs(img1_path, img2_path, conf, similar=None, label=None, cam=None):
     #     os.makedirs(os.sep.join([save_path, str(label)]))
     # save_path = os.sep.join([save_path, str(label)])
     # img_name = os.path.basename(img1_path).split('.')[0] + '_' + os.path.basename(img2_path).split('.')[0] + '.png'
+    if save_path is None:
+        save_path = conf['data']['image_joint_pth']
     if not conf['heatmap']['show_heatmap']:
         img1 = Image.open(img1_path)
         img2 = Image.open(img2_path)
         img1 = img1.resize((224, 224))
         img2 = img2.resize((224, 224))
-        save_path = conf['data']['image_joint_pth']
+        # save_path = conf['data']['image_joint_pth']
     else:
         assert cam is not None, 'cam is None'
         img1 = cam.get_hot_map(img1_path)
         img2 = cam.get_hot_map(img2_path)
-        save_path = conf['heatmap']['image_joint_pth']
+        # save_path = conf['heatmap']['image_joint_pth']
     # print('img1_path', img1)
     # print('img2_path', img2)
     if not os.path.exists(os.sep.join([save_path, str(label)])):
         os.makedirs(os.sep.join([save_path, str(label)]))
-    save_path = os.sep.join([save_path, str(label)])
+    if save_path is None:
+        save_path = os.sep.join([save_path, str(label)])
     img_name = os.path.basename(img1_path).split('.')[0] + '_' + os.path.basename(img2_path).split('.')[0] + '.png'
     assert img1.height == img2.height
 

tools/similar_analysis.py

@@ -0,0 +1,106 @@
+from configs.utils import trainer_tools
+from test_ori import group_image, featurize, cosin_metric
+from tools.dataset import get_transform
+from tools.getHeatMap import cal_cam
+from tools.image_joint import merge_imgs
+import torch.nn as nn
+import torch
+from collections import ChainMap
+import yaml
+import os
+
+
+class analysis:
+    def __init__(self):
+        with open('../configs/similar_analysis.yml', 'r') as f:
+            self.conf = yaml.load(f, Loader=yaml.FullLoader)
+        self.model = self.initialize_model(self.conf)
+        _, self.test_transform = get_transform(self.conf)
+        self.cam = cal_cam(self.model, self.conf)
+
+    def initialize_model(self, conf):
+        """初始化模型和度量方法"""
+        tr_tools = trainer_tools(conf)
+        backbone_mapping = tr_tools.get_backbone()
+
+        if conf['models']['backbone'] in backbone_mapping:
+            model = backbone_mapping[conf['models']['backbone']]()
+        else:
+            raise ValueError('不支持该模型: {}'.format({conf['models']['backbone']}))
+        try:
+            model.load_state_dict(torch.load(conf['models']['model_path'],
+                                             map_location=conf['base']['device']))
+        except:
+            state_dict = torch.load(conf['models']['model_path'],
+                                    map_location=conf['base']['device'])
+            new_state_dict = {}
+            for k, v in state_dict.items():
+                new_key = k.replace("module.", "")
+                new_state_dict[new_key] = v
+            model.load_state_dict(new_state_dict, strict=False)
+        return model.eval()
+
+    def get_feature(self, img_pth):
+        group = group_image([img_pth], self.conf['data']['val_batch_size'])
+        feature = featurize(group[0], self.test_transform, self.model, self.conf['base']['device'])
+        return feature
+
+    def get_similarity(self, feature_dict1, feature_dict2):
+        similarity = cosin_metric(feature_dict1, feature_dict2)
+        print(f"Similarity: {similarity}")
+        return similarity
+
+    def get_feature_map(self, all_imgs):
+        feature_dicts = {}
+        for img_pth in all_imgs:
+            feature_dict = self.get_feature(img_pth)
+            feature_dicts = dict(ChainMap(feature_dict, feature_dicts))
+        return feature_dicts
+
+    def get_image_map(self):
+        all_compare_img = []
+        for root, dirs, files in os.walk(self.conf['data']['data_dir']):
+            if len(dirs) == 2:
+                dir_pth_1 = os.sep.join([root, dirs[0]])
+                dir_pth_2 = os.sep.join([root, dirs[1]])
+                for img_name_1 in os.listdir(dir_pth_1):
+                    for img_name_2 in os.listdir(dir_pth_2):
+                        all_compare_img.append((os.sep.join([dir_pth_1, img_name_1]),
+                                                os.sep.join([dir_pth_2, img_name_2])))
+        return all_compare_img
+
+    def create_total_feature(self):
+        all_imgs = []
+        for root, dirs, files in os.walk(self.conf['data']['data_dir']):
+            if len(dirs) == 2:
+                for dir_name in dirs:
+                    dir_pth = os.sep.join([root, dir_name])
+                    for img_name in os.listdir(dir_pth):
+                        all_imgs.append(os.sep.join([dir_pth, img_name]))
+        return all_imgs
+
+    def get_contrast_result(self, feature_dicts, all_compare_img):
+        for img_pth1, img_pth2 in all_compare_img:
+            feature_dict1 = feature_dicts[img_pth1]
+            feature_dict2 = feature_dicts[img_pth2]
+            similarity = self.get_similarity(feature_dict1.cpu().numpy(),
+                                             feature_dict2.cpu().numpy())
+            dir_name = img_pth1.split(os.sep)[-3]
+            save_path = os.sep.join([self.conf['data']['image_joint_pth'], dir_name])
+            if similarity > 0.7:
+                merge_imgs(img_pth1,
+                           img_pth2,
+                           self.conf,
+                           similarity,
+                           label=None,
+                           cam=self.cam,
+                           save_path=save_path)
+            print(similarity)
+
+
+if __name__ == '__main__':
+    ana = analysis()
+    all_imgs = ana.create_total_feature()
+    feature_dicts = ana.get_feature_map(all_imgs)
+    all_compare_img = ana.get_image_map()
+    ana.get_contrast_result(feature_dicts, all_compare_img)

train_compare.py

@@ -17,7 +17,7 @@ import yaml
 from datetime import datetime
 
 
-def load_configuration(config_path='configs/scatter.yml'):
+def load_configuration(config_path='configs/compare.yml'):
     """加载配置文件"""
     with open(config_path, 'r') as f:
         return yaml.load(f, Loader=yaml.FullLoader)
@@ -74,13 +74,13 @@ def train_one_epoch(model, metric, criterion, optimizer, dataloader, device, sca
         data = data.to(device)
         labels = labels.to(device)
 
-        with torch.cuda.amp.autocast():
+        # with torch.cuda.amp.autocast():
-            embeddings = model(data)
+        embeddings = model(data)
-            if not conf['training']['metric'] == 'softmax':
+        if not conf['training']['metric'] == 'softmax':
-                thetas = metric(embeddings, labels)
+            thetas = metric(embeddings, labels)
-            else:
+        else:
-                thetas = metric(embeddings)
+            thetas = metric(embeddings)
-            loss = criterion(thetas, labels)
+        loss = criterion(thetas, labels)
 
         optimizer.zero_grad()
         scaler.scale(loss).backward()