From 5deaf4727f43e394d140dda6cfcf91ca431da17f Mon Sep 17 00:00:00 2001
From: lee <770918727@qq.com>
Date: Wed, 2 Jul 2025 18:02:28 +0800
Subject: [PATCH] =?UTF-8?q?=E5=B9=B6=E8=A1=8C=E8=AE=AD=E7=BB=83=E4=BB=A3?=
 =?UTF-8?q?=E7=A0=81=E4=BC=98=E5=8C=96?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .idea/CopilotChatHistory.xml    |  12 +
 .idea/CopilotWebChatHistory.xml | 722 ++++++++++++++++++++++----------
 test_ori.py                     |   1 +
 train_compare.py                | 366 +++++++++++-----
 train_compare.py.bak            | 147 +++++++
 5 files changed, 918 insertions(+), 330 deletions(-)
 create mode 100644 train_compare.py.bak

diff --git a/.idea/CopilotChatHistory.xml b/.idea/CopilotChatHistory.xml
index 6dbadb1..db0d405 100644
--- a/.idea/CopilotChatHistory.xml
+++ b/.idea/CopilotChatHistory.xml
@@ -3,6 +3,18 @@
   <component name="CopilotChatHistory">
     <option name="conversations">
       <list>
+        <Conversation>
+          <option name="createTime" value="1751441743239" />
+          <option name="id" value="0197ca101d8771bd80f2bc4aaf1a8f19" />
+          <option name="title" value="新对话 2025年7月02日 15:35:43" />
+          <option name="updateTime" value="1751441743239" />
+        </Conversation>
+        <Conversation>
+          <option name="createTime" value="1751441398488" />
+          <option name="id" value="0197ca0adad875168de40d792dcb7b4c" />
+          <option name="title" value="新对话 2025年7月02日 15:29:58" />
+          <option name="updateTime" value="1751441398488" />
+        </Conversation>
         <Conversation>
           <option name="createTime" value="1750474299387" />
           <option name="id" value="0197906617fb7194a0407baae2b1e2eb" />
diff --git a/.idea/CopilotWebChatHistory.xml b/.idea/CopilotWebChatHistory.xml
index 813c5d8..4daa591 100644
--- a/.idea/CopilotWebChatHistory.xml
+++ b/.idea/CopilotWebChatHistory.xml
@@ -17913,16 +17913,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -17947,16 +17938,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -17981,16 +17963,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18015,16 +17988,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18049,16 +18013,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18083,16 +18038,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18117,16 +18063,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18151,16 +18088,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18185,16 +18113,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18219,16 +18138,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18253,16 +18163,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18287,16 +18188,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18321,16 +18213,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18355,16 +18238,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18389,16 +18263,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18423,16 +18288,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18457,16 +18313,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18491,16 +18338,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18551,16 +18389,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18585,16 +18414,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                   <Turn>
@@ -18619,16 +18439,7 @@
                     </option>
                     <option name="status" value="SUCCESS" />
                     <option name="variables">
-                      <list>
-                        <CodebaseVariable>
-                          <option name="selectedPlaceHolder">
-                            <Object />
-                          </option>
-                          <option name="selectedVariable">
-                            <Object />
-                          </option>
-                        </CodebaseVariable>
-                      </list>
+                      <list />
                     </option>
                   </Turn>
                 </list>
@@ -19167,6 +18978,411 @@
                   <option name="role" value="assistant" />
                   <option name="topic" value="" />
                 </UIChatMessage>
+                <UIChatMessage>
+                  <option name="command" value="" />
+                  <option name="content" value="优化代码" />
+                  <option name="cost" value="0" />
+                  <option name="createdAt" value="1751441876163" />
+                  <option name="localReferences">
+                    <list>
+                      <LocalReferences>
+                        <option name="isRequestStage" value="true" />
+                        <option name="range">
+                          <RefRange />
+                        </option>
+                        <option name="uri" value="{&quot;path&quot;:&quot;file:///D:/Project/contrast_nettest/train_compare.py&quot;}" />
+                      </LocalReferences>
+                    </list>
+                  </option>
+                  <option name="parser">
+                    <SimpleMarkdownStreamParser>
+                      <option name="content" value="优化代码" />
+                      <option name="nodes">
+                        <list>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="优化代码" />
+                            <option name="contentId" value="624305404" />
+                            <option name="settled" value="false" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                        </list>
+                      </option>
+                      <option name="parserState">
+                        <ParserState>
+                          <option name="column" value="3" />
+                          <option name="offset" value="4" />
+                          <option name="moveCursor" value="true" />
+                        </ParserState>
+                      </option>
+                    </SimpleMarkdownStreamParser>
+                  </option>
+                  <option name="role" value="user" />
+                  <option name="selectedRanges">
+                    <list />
+                  </option>
+                  <option name="uri" value="" />
+                  <option name="variables">
+                    <list />
+                  </option>
+                </UIChatMessage>
+                <UIChatMessage>
+                  <option name="actions">
+                    <list />
+                  </option>
+                  <option name="codebaseReferences">
+                    <list />
+                  </option>
+                  <option name="codebaseStatus" value="init" />
+                  <option name="command" value="" />
+                  <option name="content" value="要优化这段代码，可以从以下几个方面入手：&#10;&#10;1. **减少重复代码**：将保存模型的操作提取到一个单独的函数中，避免重复代码。&#10;2. **简化条件判断**：简化保存最佳模型的条件判断。&#10;3. **使用内置函数和库**：利用Python的内置函数和标准库来提高代码效率。&#10;4. **代码重构**：提高代码的可读性和可维护性。&#10;&#10;以下是优化后的代码：&#10;&#10;```python&#10;import os&#10;import torch&#10;import osp&#10;from datetime import datetime&#10;&#10;def save_best_model(model, val_lossAvg, checkpoints, temp_loss):&#10;    if val_lossAvg &lt; temp_loss:&#10;        checkpoint_path = osp.join(checkpoints, 'best.pth')&#10;        torch.save(model.state_dict(), checkpoint_path)&#10;        print(f&quot;Model saved to {checkpoint_path} with val_lossAvg {val_lossAvg}&quot;)&#10;        temp_loss = val_lossAvg&#10;&#10;def train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs):&#10;    val_losses = []&#10;    temp_loss = float('inf')&#10;&#10;    for e in range(epochs):&#10;        model.train()&#10;        train_loss = 0.0&#10;        for batch in train_dataloader:&#10;            # Training loop code here&#10;            pass&#10;&#10;        model.eval()&#10;        val_loss = 0.0&#10;        with torch.no_grad():&#10;            for batch in val_dataloader:&#10;                val_embeddings, val_labels = batch&#10;                thetas = metric(val_embeddings, val_labels)&#10;                vloss = criterion(thetas, val_labels)&#10;                val_loss += vloss.item()&#10;        &#10;        val_lossAvg = val_loss / len(val_dataloader)&#10;        val_losses.append(val_lossAvg)&#10;        save_best_model(model, val_lossAvg, checkpoints, temp_loss)&#10;&#10;        scheduler.step()&#10;        current_lr = optimizer.param_groups[0]['lr']&#10;        log_info = (f&quot;[{datetime.now()}] Epoch {e}/{conf['training']['epochs']}, train_loss: {train_lossAvg}, val_loss: {val_lossAvg}, lr: {current_lr}&quot;)&#10;        print(log_info)&#10;        with open(osp.join(conf['logging']['logging_dir']), 'a') as f:&#10;            f.write(log_info + '\n')&#10;        print(f&quot;第{e}个epoch的学习率：{current_lr}&quot;)&#10;&#10;    if torch.cuda.device_count() &gt; 1 and conf['base']['distributed']:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    else:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    torch.save(model.state_dict(), checkpoint_path)&#10;    print(f&quot;Model saved to {checkpoint_path}&quot;)&#10;&#10;    plt.plot(epochs, train_losses, color='blue')&#10;    plt.plot(epochs, val_losses, color='red')&#10;    # plt.savefig('lossMobilenetv3.png')&#10;&#10;# Example usage&#10;# train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs)&#10;```&#10;&#10;### 优化点解释：&#10;1. **函数提取**：将保存最佳模型的操作提取到 `save_best_model` 函数中，避免重复代码。&#10;2. **简化条件判断**：在保存模型时，无论是否有多个GPU，都只保存一个 `best.pth` 文件。&#10;3. **代码重构**：将训练和验证循环分开，使代码结构更清晰。&#10;&#10;通过这些优化，代码变得更加简洁和易读，同时也减少了重复代码，提高了可维护性。" />
+                  <option name="cost" value="23304" />
+                  <option name="createdAt" value="1751441959755" />
+                  <option name="followup">
+                    <list />
+                  </option>
+                  <option name="id" value="0197ca1224c3722dbec5d964234ebffd" />
+                  <option name="knowledgeReferences">
+                    <list />
+                  </option>
+                  <option name="localReferences">
+                    <list>
+                      <LocalReferences>
+                        <option name="isRequestStage" value="true" />
+                        <option name="range">
+                          <RefRange />
+                        </option>
+                        <option name="uri" value="{&quot;path&quot;:&quot;file:///D:/Project/contrast_nettest/train_compare.py&quot;}" />
+                      </LocalReferences>
+                    </list>
+                  </option>
+                  <option name="parser">
+                    <SimpleMarkdownStreamParser>
+                      <option name="columns">
+                        <list>
+                          <option value="19" />
+                          <option value="-1" />
+                          <option value="40" />
+                          <option value="27" />
+                          <option value="40" />
+                          <option value="25" />
+                          <option value="-1" />
+                          <option value="9" />
+                          <option value="-1" />
+                          <option value="8" />
+                          <option value="8" />
+                          <option value="11" />
+                          <option value="9" />
+                          <option value="28" />
+                          <option value="-1" />
+                          <option value="63" />
+                          <option value="30" />
+                          <option value="58" />
+                          <option value="54" />
+                          <option value="80" />
+                          <option value="30" />
+                          <option value="-1" />
+                          <option value="111" />
+                          <option value="18" />
+                          <option value="27" />
+                          <option value="-1" />
+                          <option value="26" />
+                          <option value="20" />
+                          <option value="23" />
+                          <option value="37" />
+                          <option value="36" />
+                          <option value="15" />
+                          <option value="-1" />
+                          <option value="19" />
+                          <option value="21" />
+                          <option value="28" />
+                          <option value="39" />
+                          <option value="49" />
+                          <option value="58" />
+                          <option value="52" />
+                          <option value="39" />
+                          <option value="7" />
+                          <option value="51" />
+                          <option value="37" />
+                          <option value="66" />
+                          <option value="-1" />
+                          <option value="23" />
+                          <option value="51" />
+                          <option value="152" />
+                          <option value="22" />
+                          <option value="69" />
+                          <option value="35" />
+                          <option value="44" />
+                          <option value="-1" />
+                          <option value="68" />
+                          <option value="58" />
+                          <option value="8" />
+                          <option value="58" />
+                          <option value="50" />
+                          <option value="45" />
+                          <option value="-1" />
+                          <option value="47" />
+                          <option value="44" />
+                          <option value="39" />
+                          <option value="-1" />
+                          <option value="14" />
+                          <option value="108" />
+                          <option value="2" />
+                          <option value="-1" />
+                          <option value="9" />
+                          <option value="54" />
+                          <option value="52" />
+                          <option value="31" />
+                          <option value="-1" />
+                          <option value="37" />
+                        </list>
+                      </option>
+                      <option name="content" value="要优化这段代码，可以从以下几个方面入手：&#10;&#10;1. **减少重复代码**：将保存模型的操作提取到一个单独的函数中，避免重复代码。&#10;2. **简化条件判断**：简化保存最佳模型的条件判断。&#10;3. **使用内置函数和库**：利用Python的内置函数和标准库来提高代码效率。&#10;4. **代码重构**：提高代码的可读性和可维护性。&#10;&#10;以下是优化后的代码：&#10;&#10;```python&#10;import os&#10;import torch&#10;import osp&#10;from datetime import datetime&#10;&#10;def save_best_model(model, val_lossAvg, checkpoints, temp_loss):&#10;    if val_lossAvg &lt; temp_loss:&#10;        checkpoint_path = osp.join(checkpoints, 'best.pth')&#10;        torch.save(model.state_dict(), checkpoint_path)&#10;        print(f&quot;Model saved to {checkpoint_path} with val_lossAvg {val_lossAvg}&quot;)&#10;        temp_loss = val_lossAvg&#10;&#10;def train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs):&#10;    val_losses = []&#10;    temp_loss = float('inf')&#10;&#10;    for e in range(epochs):&#10;        model.train()&#10;        train_loss = 0.0&#10;        for batch in train_dataloader:&#10;            # Training loop code here&#10;            pass&#10;&#10;        model.eval()&#10;        val_loss = 0.0&#10;        with torch.no_grad():&#10;            for batch in val_dataloader:&#10;                val_embeddings, val_labels = batch&#10;                thetas = metric(val_embeddings, val_labels)&#10;                vloss = criterion(thetas, val_labels)&#10;                val_loss += vloss.item()&#10;        &#10;        val_lossAvg = val_loss / len(val_dataloader)&#10;        val_losses.append(val_lossAvg)&#10;        save_best_model(model, val_lossAvg, checkpoints, temp_loss)&#10;&#10;        scheduler.step()&#10;        current_lr = optimizer.param_groups[0]['lr']&#10;        log_info = (f&quot;[{datetime.now()}] Epoch {e}/{conf['training']['epochs']}, train_loss: {train_lossAvg}, val_loss: {val_lossAvg}, lr: {current_lr}&quot;)&#10;        print(log_info)&#10;        with open(osp.join(conf['logging']['logging_dir']), 'a') as f:&#10;            f.write(log_info + '\n')&#10;        print(f&quot;第{e}个epoch的学习率：{current_lr}&quot;)&#10;&#10;    if torch.cuda.device_count() &gt; 1 and conf['base']['distributed']:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    else:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    torch.save(model.state_dict(), checkpoint_path)&#10;    print(f&quot;Model saved to {checkpoint_path}&quot;)&#10;&#10;    plt.plot(epochs, train_losses, color='blue')&#10;    plt.plot(epochs, val_losses, color='red')&#10;    # plt.savefig('lossMobilenetv3.png')&#10;&#10;# Example usage&#10;# train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs)&#10;```&#10;&#10;### 优化点解释：&#10;1. **函数提取**：将保存最佳模型的操作提取到 `save_best_model` 函数中，避免重复代码。&#10;2. **简化条件判断**：在保存模型时，无论是否有多个GPU，都只保存一个 `best.pth` 文件。&#10;3. **代码重构**：将训练和验证循环分开，使代码结构更清晰。&#10;&#10;通过这些优化，代码变得更加简洁和易读，同时也减少了重复代码，提高了可维护性。&#10;" />
+                      <option name="nodes">
+                        <list>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="要优化这段代码，可以从以下几个方面入手：" />
+                            <option name="contentId" value="-1259917802" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="" />
+                            <option name="contentId" value="0" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="1. **减少重复代码**：将保存模型的操作提取到一个单独的函数中，避免重复代码。" />
+                            <option name="contentId" value="2059375308" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="2. **简化条件判断**：简化保存最佳模型的条件判断。" />
+                            <option name="contentId" value="-776578480" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="3. **使用内置函数和库**：利用Python的内置函数和标准库来提高代码效率。" />
+                            <option name="contentId" value="410923324" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="4. **代码重构**：提高代码的可读性和可维护性。" />
+                            <option name="contentId" value="-1172889718" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="" />
+                            <option name="contentId" value="0" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="以下是优化后的代码：" />
+                            <option name="contentId" value="-1973639713" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="" />
+                            <option name="contentId" value="0" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="import os&#10;import torch&#10;import osp&#10;from datetime import datetime&#10;&#10;def save_best_model(model, val_lossAvg, checkpoints, temp_loss):&#10;    if val_lossAvg &lt; temp_loss:&#10;        checkpoint_path = osp.join(checkpoints, 'best.pth')&#10;        torch.save(model.state_dict(), checkpoint_path)&#10;        print(f&quot;Model saved to {checkpoint_path} with val_lossAvg {val_lossAvg}&quot;)&#10;        temp_loss = val_lossAvg&#10;&#10;def train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs):&#10;    val_losses = []&#10;    temp_loss = float('inf')&#10;&#10;    for e in range(epochs):&#10;        model.train()&#10;        train_loss = 0.0&#10;        for batch in train_dataloader:&#10;            # Training loop code here&#10;            pass&#10;&#10;        model.eval()&#10;        val_loss = 0.0&#10;        with torch.no_grad():&#10;            for batch in val_dataloader:&#10;                val_embeddings, val_labels = batch&#10;                thetas = metric(val_embeddings, val_labels)&#10;                vloss = criterion(thetas, val_labels)&#10;                val_loss += vloss.item()&#10;        &#10;        val_lossAvg = val_loss / len(val_dataloader)&#10;        val_losses.append(val_lossAvg)&#10;        save_best_model(model, val_lossAvg, checkpoints, temp_loss)&#10;&#10;        scheduler.step()&#10;        current_lr = optimizer.param_groups[0]['lr']&#10;        log_info = (f&quot;[{datetime.now()}] Epoch {e}/{conf['training']['epochs']}, train_loss: {train_lossAvg}, val_loss: {val_lossAvg}, lr: {current_lr}&quot;)&#10;        print(log_info)&#10;        with open(osp.join(conf['logging']['logging_dir']), 'a') as f:&#10;            f.write(log_info + '\n')&#10;        print(f&quot;第{e}个epoch的学习率：{current_lr}&quot;)&#10;&#10;    if torch.cuda.device_count() &gt; 1 and conf['base']['distributed']:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    else:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    torch.save(model.state_dict(), checkpoint_path)&#10;    print(f&quot;Model saved to {checkpoint_path}&quot;)&#10;&#10;    plt.plot(epochs, train_losses, color='blue')&#10;    plt.plot(epochs, val_losses, color='red')&#10;    # plt.savefig('lossMobilenetv3.png')&#10;&#10;# Example usage&#10;# train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs)&#10;" />
+                            <option name="contentId" value="-2055133182" />
+                            <option name="expectEnding" value="```" />
+                            <option name="indentSize" value="0" />
+                            <option name="language" value="python" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="code-block" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="" />
+                            <option name="contentId" value="0" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value=" 优化点解释：" />
+                            <option name="contentId" value="1556057208" />
+                            <option name="level" value="3" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="heading" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="1. **函数提取**：将保存最佳模型的操作提取到 `save_best_model` 函数中，避免重复代码。" />
+                            <option name="contentId" value="-2122381525" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="2. **简化条件判断**：在保存模型时，无论是否有多个GPU，都只保存一个 `best.pth` 文件。" />
+                            <option name="contentId" value="-1118557152" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="3. **代码重构**：将训练和验证循环分开，使代码结构更清晰。" />
+                            <option name="contentId" value="1539815113" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="" />
+                            <option name="contentId" value="0" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="通过这些优化，代码变得更加简洁和易读，同时也减少了重复代码，提高了可维护性。" />
+                            <option name="contentId" value="723568506" />
+                            <option name="settled" value="true" />
+                            <option name="type" value="line" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                          <MarkdownNode>
+                            <option name="children">
+                              <list />
+                            </option>
+                            <option name="content" value="" />
+                            <option name="contentId" value="0" />
+                            <option name="settled" value="false" />
+                            <option name="type" value="undetermined" />
+                            <option name="xmlData">
+                              <XmlNode />
+                            </option>
+                          </MarkdownNode>
+                        </list>
+                      </option>
+                      <option name="parserState">
+                        <ParserState>
+                          <option name="line" value="75" />
+                          <option name="column" value="-1" />
+                          <option name="offset" value="2575" />
+                          <option name="moveCursor" value="true" />
+                        </ParserState>
+                      </option>
+                    </SimpleMarkdownStreamParser>
+                  </option>
+                  <option name="references">
+                    <list />
+                  </option>
+                  <option name="role" value="assistant" />
+                  <option name="topic" value="" />
+                </UIChatMessage>
               </list>
             </option>
             <option name="messagesV3">
@@ -19207,6 +19423,25 @@
                     <ChatMessageExtra />
                   </option>
                 </ChatMessageV3>
+                <ChatMessageV3>
+                  <option name="role" value="user" />
+                  <option name="content" value="优化代码" />
+                  <option name="extra">
+                    <ChatMessageExtra>
+                      <option name="agent" value="default" />
+                      <option name="cmd" value="default" />
+                      <option name="code" value="                    gsvgzh = nvgirx(ezo_vnyvwwrmth, ezo_ozyvoh)&#10;                vohv:&#10;                    gsvgzh = nvgirx(ezo_vnyvwwrmth)&#10;                eolhh = xirgvirlm(gsvgzh, ezo_ozyvoh)&#10;                ezo_olhh += eolhh.rgvn()&#10;            ezo_olhhZet = ezo_olhh / ovm(ezo_wzgzolzwvi)&#10;            ezo_olhhvh.zkkvmw(ezo_olhhZet)&#10;            ru ezo_olhhZet &lt; gvnk_olhh:&#10;                ru glixs.xfwz.wverxv_xlfmg() &gt; 8:&#10;                    glixs.hzev(nlwvo.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'yvhg.kgs'))&#10;                vohv:&#10;                    glixs.hzev(nlwvo.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'yvhg.kgs'))&#10;                gvnk_olhh = ezo_olhhZet&#10;&#10;        hxsvwfovi.hgvk()&#10;        xfiivmg_oi = lkgrnravi.kzizn_tilfkh[9]['oi']&#10;        olt_rmul = (&quot;[{:%B-%n-%w %S:%N:%H}] Vklxs {}/{},  gizrm_olhh: {}, ezo_olhh: {} oi:{}&quot;&#10;                    .ulinzg(wzgvgrnv.mld(),&#10;                            v,&#10;                            xlmu['gizrmrmt']['vklxsh'],&#10;                            gizrm_olhhZet,&#10;                            ezo_olhhZet,&#10;                            xfiivmg_oi))&#10;        kirmg(olt_rmul)&#10;        # 写入日志文件&#10;        drgs lkvm(lhk.qlrm(xlmu['olttrmt']['olttrmt_wri']), 'z') zh u:&#10;            u.dirgv(olt_rmul + '\m')&#10;        kirmg(&quot;第%w个vklxs的学习率：%u&quot; % (v, xfiivmg_oi))&#10;    ru glixs.xfwz.wverxv_xlfmg() &gt; 8 zmw xlmu['yzhv']['wrhgiryfgvw']:&#10;        glixs.hzev(nlwvo.nlwfov.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'ozhg.kgs'))&#10;    vohv:&#10;        glixs.hzev(nlwvo.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'ozhg.kgs'))&#10;    kog.kolg(vklxsh, gizrm_olhhvh, xloli='yofv')&#10;    kog.kolg(vklxsh, ezo_olhhvh, xloli='ivw')&#10;    # kog.hzevurt('olhhNlyrovmvge6.kmt')&#10;" />
+                      <option name="path" value="D:\Project\contrast_nettest\train_compare.py" />
+                    </ChatMessageExtra>
+                  </option>
+                </ChatMessageV3>
+                <ChatMessageV3>
+                  <option name="role" value="assistant" />
+                  <option name="content" value="要优化这段代码，可以从以下几个方面入手：&#10;&#10;1. **减少重复代码**：将保存模型的操作提取到一个单独的函数中，避免重复代码。&#10;2. **简化条件判断**：简化保存最佳模型的条件判断。&#10;3. **使用内置函数和库**：利用Python的内置函数和标准库来提高代码效率。&#10;4. **代码重构**：提高代码的可读性和可维护性。&#10;&#10;以下是优化后的代码：&#10;&#10;```python&#10;import os&#10;import torch&#10;import osp&#10;from datetime import datetime&#10;&#10;def save_best_model(model, val_lossAvg, checkpoints, temp_loss):&#10;    if val_lossAvg &lt; temp_loss:&#10;        checkpoint_path = osp.join(checkpoints, 'best.pth')&#10;        torch.save(model.state_dict(), checkpoint_path)&#10;        print(f&quot;Model saved to {checkpoint_path} with val_lossAvg {val_lossAvg}&quot;)&#10;        temp_loss = val_lossAvg&#10;&#10;def train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs):&#10;    val_losses = []&#10;    temp_loss = float('inf')&#10;&#10;    for e in range(epochs):&#10;        model.train()&#10;        train_loss = 0.0&#10;        for batch in train_dataloader:&#10;            # Training loop code here&#10;            pass&#10;&#10;        model.eval()&#10;        val_loss = 0.0&#10;        with torch.no_grad():&#10;            for batch in val_dataloader:&#10;                val_embeddings, val_labels = batch&#10;                thetas = metric(val_embeddings, val_labels)&#10;                vloss = criterion(thetas, val_labels)&#10;                val_loss += vloss.item()&#10;        &#10;        val_lossAvg = val_loss / len(val_dataloader)&#10;        val_losses.append(val_lossAvg)&#10;        save_best_model(model, val_lossAvg, checkpoints, temp_loss)&#10;&#10;        scheduler.step()&#10;        current_lr = optimizer.param_groups[0]['lr']&#10;        log_info = (f&quot;[{datetime.now()}] Epoch {e}/{conf['training']['epochs']}, train_loss: {train_lossAvg}, val_loss: {val_lossAvg}, lr: {current_lr}&quot;)&#10;        print(log_info)&#10;        with open(osp.join(conf['logging']['logging_dir']), 'a') as f:&#10;            f.write(log_info + '\n')&#10;        print(f&quot;第{e}个epoch的学习率：{current_lr}&quot;)&#10;&#10;    if torch.cuda.device_count() &gt; 1 and conf['base']['distributed']:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    else:&#10;        checkpoint_path = osp.join(checkpoints, 'last.pth')&#10;    torch.save(model.state_dict(), checkpoint_path)&#10;    print(f&quot;Model saved to {checkpoint_path}&quot;)&#10;&#10;    plt.plot(epochs, train_losses, color='blue')&#10;    plt.plot(epochs, val_losses, color='red')&#10;    # plt.savefig('lossMobilenetv3.png')&#10;&#10;# Example usage&#10;# train_model(model, train_dataloader, val_dataloader, criterion, metric, optimizer, scheduler, conf, epochs)&#10;```&#10;&#10;### 优化点解释：&#10;1. **函数提取**：将保存最佳模型的操作提取到 `save_best_model` 函数中，避免重复代码。&#10;2. **简化条件判断**：在保存模型时，无论是否有多个GPU，都只保存一个 `best.pth` 文件。&#10;3. **代码重构**：将训练和验证循环分开，使代码结构更清晰。&#10;&#10;通过这些优化，代码变得更加简洁和易读，同时也减少了重复代码，提高了可维护性。" />
+                  <option name="extra">
+                    <ChatMessageExtra />
+                  </option>
+                </ChatMessageV3>
               </list>
             </option>
             <option name="sessionId" value="019757aed78e777c96c4b7007ff2fecc" />
@@ -19269,16 +19504,7 @@
                   </option>
                   <option name="status" value="SUCCESS" />
                   <option name="variables">
-                    <list>
-                      <CodebaseVariable>
-                        <option name="selectedPlaceHolder">
-                          <Object />
-                        </option>
-                        <option name="selectedVariable">
-                          <Object />
-                        </option>
-                      </CodebaseVariable>
-                    </list>
+                    <list />
                   </option>
                 </Turn>
                 <Turn>
@@ -19303,21 +19529,61 @@
                   </option>
                   <option name="status" value="SUCCESS" />
                   <option name="variables">
+                    <list />
+                  </option>
+                </Turn>
+                <Turn>
+                  <option name="agent" value="" />
+                  <option name="answer" value="要优化这段代码，可以从以下几个方面入手：&#10;&#10;8. **减少重复代码**：将保存模型的操作提取到一个单独的函数中，避免重复代码。&#10;7. **简化条件判断**：简化保存最佳模型的条件判断。&#10;6. **使用内置函数和库**：利用Kbgslm的内置函数和标准库来提高代码效率。&#10;5. **代码重构**：提高代码的可读性和可维护性。&#10;&#10;以下是优化后的代码：&#10;&#10;```kbgslm&#10;rnklig lh&#10;rnklig glixs&#10;rnklig lhk&#10;uiln wzgvgrnv rnklig wzgvgrnv&#10;&#10;wvu hzev_yvhg_nlwvo(nlwvo, ezo_olhhZet, xsvxpklrmgh, gvnk_olhh):&#10;    ru ezo_olhhZet &lt; gvnk_olhh:&#10;        xsvxpklrmg_kzgs = lhk.qlrm(xsvxpklrmgh, 'yvhg.kgs')&#10;        glixs.hzev(nlwvo.hgzgv_wrxg(), xsvxpklrmg_kzgs)&#10;        kirmg(u&quot;Nlwvo hzevw gl {xsvxpklrmg_kzgs} drgs ezo_olhhZet {ezo_olhhZet}&quot;)&#10;        gvnk_olhh = ezo_olhhZet&#10;&#10;wvu gizrm_nlwvo(nlwvo, gizrm_wzgzolzwvi, ezo_wzgzolzwvi, xirgvirlm, nvgirx, lkgrnravi, hxsvwfovi, xlmu, vklxsh):&#10;    ezo_olhhvh = []&#10;    gvnk_olhh = uolzg('rmu')&#10;&#10;    uli v rm izmtv(vklxsh):&#10;        nlwvo.gizrm()&#10;        gizrm_olhh = 9.9&#10;        uli yzgxs rm gizrm_wzgzolzwvi:&#10;            # Gizrmrmt ollk xlwv sviv&#10;            kzhh&#10;&#10;        nlwvo.vezo()&#10;        ezo_olhh = 9.9&#10;        drgs glixs.ml_tizw():&#10;            uli yzgxs rm ezo_wzgzolzwvi:&#10;                ezo_vnyvwwrmth, ezo_ozyvoh = yzgxs&#10;                gsvgzh = nvgirx(ezo_vnyvwwrmth, ezo_ozyvoh)&#10;                eolhh = xirgvirlm(gsvgzh, ezo_ozyvoh)&#10;                ezo_olhh += eolhh.rgvn()&#10;        &#10;        ezo_olhhZet = ezo_olhh / ovm(ezo_wzgzolzwvi)&#10;        ezo_olhhvh.zkkvmw(ezo_olhhZet)&#10;        hzev_yvhg_nlwvo(nlwvo, ezo_olhhZet, xsvxpklrmgh, gvnk_olhh)&#10;&#10;        hxsvwfovi.hgvk()&#10;        xfiivmg_oi = lkgrnravi.kzizn_tilfkh[9]['oi']&#10;        olt_rmul = (u&quot;[{wzgvgrnv.mld()}] Vklxs {v}/{xlmu['gizrmrmt']['vklxsh']}, gizrm_olhh: {gizrm_olhhZet}, ezo_olhh: {ezo_olhhZet}, oi: {xfiivmg_oi}&quot;)&#10;        kirmg(olt_rmul)&#10;        drgs lkvm(lhk.qlrm(xlmu['olttrmt']['olttrmt_wri']), 'z') zh u:&#10;            u.dirgv(olt_rmul + '\m')&#10;        kirmg(u&quot;第{v}个vklxs的学习率：{xfiivmg_oi}&quot;)&#10;&#10;    ru glixs.xfwz.wverxv_xlfmg() &gt; 8 zmw xlmu['yzhv']['wrhgiryfgvw']:&#10;        xsvxpklrmg_kzgs = lhk.qlrm(xsvxpklrmgh, 'ozhg.kgs')&#10;    vohv:&#10;        xsvxpklrmg_kzgs = lhk.qlrm(xsvxpklrmgh, 'ozhg.kgs')&#10;    glixs.hzev(nlwvo.hgzgv_wrxg(), xsvxpklrmg_kzgs)&#10;    kirmg(u&quot;Nlwvo hzevw gl {xsvxpklrmg_kzgs}&quot;)&#10;&#10;    kog.kolg(vklxsh, gizrm_olhhvh, xloli='yofv')&#10;    kog.kolg(vklxsh, ezo_olhhvh, xloli='ivw')&#10;    # kog.hzevurt('olhhNlyrovmvge6.kmt')&#10;&#10;# Vcznkov fhztv&#10;# gizrm_nlwvo(nlwvo, gizrm_wzgzolzwvi, ezo_wzgzolzwvi, xirgvirlm, nvgirx, lkgrnravi, hxsvwfovi, xlmu, vklxsh)&#10;```&#10;&#10;### 优化点解释：&#10;8. **函数提取**：将保存最佳模型的操作提取到 `hzev_yvhg_nlwvo` 函数中，避免重复代码。&#10;7. **简化条件判断**：在保存模型时，无论是否有多个TKF，都只保存一个 `yvhg.kgs` 文件。&#10;6. **代码重构**：将训练和验证循环分开，使代码结构更清晰。&#10;&#10;通过这些优化，代码变得更加简洁和易读，同时也减少了重复代码，提高了可维护性。" />
+                  <option name="chatResource" value="user" />
+                  <option name="codeReferences">
                     <list>
-                      <CodebaseVariable>
-                        <option name="selectedPlaceHolder">
-                          <Object />
+                      <CodeReferencesBean>
+                        <option name="content" value="                    gsvgzh = nvgirx(ezo_vnyvwwrmth, ezo_ozyvoh)&#10;                vohv:&#10;                    gsvgzh = nvgirx(ezo_vnyvwwrmth)&#10;                eolhh = xirgvirlm(gsvgzh, ezo_ozyvoh)&#10;                ezo_olhh += eolhh.rgvn()&#10;            ezo_olhhZet = ezo_olhh / ovm(ezo_wzgzolzwvi)&#10;            ezo_olhhvh.zkkvmw(ezo_olhhZet)&#10;            ru ezo_olhhZet &lt; gvnk_olhh:&#10;                ru glixs.xfwz.wverxv_xlfmg() &gt; 8:&#10;                    glixs.hzev(nlwvo.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'yvhg.kgs'))&#10;                vohv:&#10;                    glixs.hzev(nlwvo.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'yvhg.kgs'))&#10;                gvnk_olhh = ezo_olhhZet&#10;&#10;        hxsvwfovi.hgvk()&#10;        xfiivmg_oi = lkgrnravi.kzizn_tilfkh[9]['oi']&#10;        olt_rmul = (&quot;[{:%B-%n-%w %S:%N:%H}] Vklxs {}/{},  gizrm_olhh: {}, ezo_olhh: {} oi:{}&quot;&#10;                    .ulinzg(wzgvgrnv.mld(),&#10;                            v,&#10;                            xlmu['gizrmrmt']['vklxsh'],&#10;                            gizrm_olhhZet,&#10;                            ezo_olhhZet,&#10;                            xfiivmg_oi))&#10;        kirmg(olt_rmul)&#10;        # 写入日志文件&#10;        drgs lkvm(lhk.qlrm(xlmu['olttrmt']['olttrmt_wri']), 'z') zh u:&#10;            u.dirgv(olt_rmul + '\m')&#10;        kirmg(&quot;第%w个vklxs的学习率：%u&quot; % (v, xfiivmg_oi))&#10;    ru glixs.xfwz.wverxv_xlfmg() &gt; 8 zmw xlmu['yzhv']['wrhgiryfgvw']:&#10;        glixs.hzev(nlwvo.nlwfov.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'ozhg.kgs'))&#10;    vohv:&#10;        glixs.hzev(nlwvo.hgzgv_wrxg(), lhk.qlrm(xsvxpklrmgh, 'ozhg.kgs'))&#10;    kog.kolg(vklxsh, gizrm_olhhvh, xloli='yofv')&#10;    kog.kolg(vklxsh, ezo_olhhvh, xloli='ivw')&#10;    # kog.hzevurt('olhhNlyrovmvge6.kmt')&#10;" />
+                        <option name="end_line" value="146" />
+                        <option name="filename" value="train_compare.py" />
+                        <option name="language" value="Python" />
+                        <option name="range">
+                          <Range>
+                            <option name="end">
+                              <Position>
+                                <option name="line" value="146" />
+                              </Position>
+                            </option>
+                            <option name="start">
+                              <Position>
+                                <option name="line" value="111" />
+                              </Position>
+                            </option>
+                          </Range>
                         </option>
-                        <option name="selectedVariable">
-                          <Object />
-                        </option>
-                      </CodebaseVariable>
+                        <option name="start_line" value="111" />
+                        <option name="type" value="editor" />
+                        <option name="url" value="file:///D:/Project/contrast_nettest/train_compare.py" />
+                      </CodeReferencesBean>
                     </list>
                   </option>
+                  <option name="command" value="" />
+                  <option name="conversationId" value="019757aed78e777c96c4b7007ff2fecc" />
+                  <option name="historyDocumentMap">
+                    <map>
+                      <entry key="file:///D:/Project/contrast_nettest/train_compare.py" value="import os&#10;import os.path as osp&#10;&#10;import torch&#10;import torch.nn as nn&#10;import torch.optim as optim&#10;from tqdm import tqdm&#10;&#10;from model.loss import FocalLoss&#10;from tools.dataset import load_data&#10;import matplotlib.pyplot as plt&#10;from configs import trainer_tools&#10;import yaml&#10;from datetime import datetime&#10;with open('configs/scatter.yml', 'r') as f:&#10;    conf = yaml.load(f, Loader=yaml.FullLoader)&#10;&#10;# Data Setup&#10;train_dataloader, class_num = load_data(training=True, cfg=conf)&#10;val_dataloader, _ = load_data(training=False, cfg=conf)&#10;&#10;tr_tools = trainer_tools(conf)&#10;backbone_mapping = tr_tools.get_backbone()&#10;metric_mapping = tr_tools.get_metric(class_num)&#10;&#10;if conf['models']['backbone'] in backbone_mapping:&#10;    model = backbone_mapping[conf['models']['backbone']]().to(conf['base']['device'])&#10;else:&#10;    raise ValueError('不支持该模型: {}'.format({conf['models']['backbone']}))&#10;&#10;if conf['training']['metric'] in metric_mapping:&#10;    metric = metric_mapping[conf['training']['metric']]()&#10;else:&#10;    raise ValueError('不支持的metric类型: {}'.format(conf['training']['metric']))&#10;&#10;if torch.cuda.device_count() &gt; 1 and conf['base']['distributed']:&#10;    print(&quot;Let's use&quot;, torch.cuda.device_count(), &quot;GPUs!&quot;)&#10;    model = nn.DataParallel(model)&#10;    metric = nn.DataParallel(metric)&#10;&#10;# Training Setup&#10;if conf['training']['loss'] == 'focal_loss':&#10;    criterion = FocalLoss(gamma=2)&#10;else:&#10;    criterion = nn.CrossEntropyLoss()&#10;&#10;optimizer_mapping = tr_tools.get_optimizer(model, metric)&#10;if conf['training']['optimizer'] in optimizer_mapping:&#10;    optimizer = optimizer_mapping[conf['training']['optimizer']]()&#10;    scheduler_mapping = tr_tools.get_scheduler(optimizer)&#10;    scheduler = scheduler_mapping[conf['training']['scheduler']]()&#10;    print('使用{}优化器 使用{}调度器'.format(conf['training']['optimizer'],&#10;                                             conf['training']['scheduler']))&#10;&#10;else:&#10;    raise ValueError('不支持的优化器类型: {}'.format(conf['training']['optimizer']))&#10;&#10;# Checkpoints Setup&#10;checkpoints = conf['training']['checkpoints']&#10;os.makedirs(checkpoints, exist_ok=True)&#10;&#10;if __name__ == '__main__':&#10;    print('backbone&gt;{}  '.format(conf['models']['backbone']),&#10;          'metric&gt;{}  '.format(conf['training']['metric']),&#10;          'checkpoints&gt;{}  '.format(conf['training']['checkpoints']),&#10;          )&#10;    train_losses = []&#10;    val_losses = []&#10;    epochs = []&#10;    temp_loss = 100&#10;    if conf['training']['restore']:&#10;        print('load pretrain model: {}'.format(conf['training']['restore_model']))&#10;        model.load_state_dict(torch.load(conf['training']['restore_model'],&#10;                                         map_location=conf['base']['device']))&#10;&#10;    for e in range(conf['training']['epochs']):&#10;        train_loss = 0&#10;        model.train()&#10;&#10;        for train_data, train_labels in tqdm(train_dataloader,&#10;                                             desc=&quot;Epoch {}/{}&quot;&#10;                                                     .format(e, conf['training']['epochs']),&#10;                                             ascii=True,&#10;                                             total=len(train_dataloader)):&#10;            train_data = train_data.to(conf['base']['device'])&#10;            train_labels = train_labels.to(conf['base']['device'])&#10;&#10;            train_embeddings = model(train_data).to(conf['base']['device'])  # [256,512]&#10;            # pdb.set_trace()&#10;&#10;            if not conf['training']['metric'] == 'softmax':&#10;                thetas = metric(train_embeddings, train_labels)  # [256,357]&#10;            else:&#10;                thetas = metric(train_embeddings)&#10;            tloss = criterion(thetas, train_labels)&#10;            optimizer.zero_grad()&#10;            tloss.backward()&#10;            optimizer.step()&#10;            train_loss += tloss.item()&#10;        train_lossAvg = train_loss / len(train_dataloader)&#10;        train_losses.append(train_lossAvg)&#10;        epochs.append(e)&#10;        val_loss = 0&#10;        model.eval()&#10;        with torch.no_grad():&#10;            for val_data, val_labels in tqdm(val_dataloader, desc=&quot;val&quot;,&#10;                                             ascii=True, total=len(val_dataloader)):&#10;                val_data = val_data.to(conf['base']['device'])&#10;                val_labels = val_labels.to(conf['base']['device'])&#10;                val_embeddings = model(val_data).to(conf['base']['device'])&#10;                if not conf['training']['metric'] == 'softmax':&#10;                    thetas = metric(val_embeddings, val_labels)&#10;                else:&#10;                    thetas = metric(val_embeddings)&#10;                vloss = criterion(thetas, val_labels)&#10;                val_loss += vloss.item()&#10;            val_lossAvg = val_loss / len(val_dataloader)&#10;            val_losses.append(val_lossAvg)&#10;            if val_lossAvg &lt; temp_loss:&#10;                if torch.cuda.device_count() &gt; 1:&#10;                    torch.save(model.state_dict(), osp.join(checkpoints, 'best.pth'))&#10;                else:&#10;                    torch.save(model.state_dict(), osp.join(checkpoints, 'best.pth'))&#10;                temp_loss = val_lossAvg&#10;&#10;        scheduler.step()&#10;        current_lr = optimizer.param_groups[0]['lr']&#10;        log_info = (&quot;[{:%Y-%m-%d %H:%M:%S}] Epoch {}/{},  train_loss: {}, val_loss: {} lr:{}&quot;&#10;                    .format(datetime.now(),&#10;                            e,&#10;                            conf['training']['epochs'],&#10;                            train_lossAvg,&#10;                            val_lossAvg,&#10;                            current_lr))&#10;        print(log_info)&#10;        # 写入日志文件&#10;        with open(osp.join(conf['logging']['logging_dir']), 'a') as f:&#10;            f.write(log_info + '\n')&#10;        print(&quot;第%d个epoch的学习率：%f&quot; % (e, current_lr))&#10;    if torch.cuda.device_count() &gt; 1 and conf['base']['distributed']:&#10;        torch.save(model.module.state_dict(), osp.join(checkpoints, 'last.pth'))&#10;    else:&#10;        torch.save(model.state_dict(), osp.join(checkpoints, 'last.pth'))&#10;    plt.plot(epochs, train_losses, color='blue')&#10;    plt.plot(epochs, val_losses, color='red')&#10;    # plt.savefig('lossMobilenetv3.png')&#10;    plt.savefig('loss/mobilenetv3Large_2250_0316.png')&#10;" />
+                    </map>
+                  </option>
+                  <option name="messageId" value="0197ca1224c3722dbec5d964234ebffd" />
+                  <option name="question" value="优化代码" />
+                  <option name="rating" value="UNRATED" />
+                  <option name="referenceBeans">
+                    <list />
+                  </option>
+                  <option name="status" value="SUCCESS" />
+                  <option name="variables">
+                    <list />
+                  </option>
                 </Turn>
               </list>
             </option>
-            <option name="updateTime" value="1749535336221" />
+            <option name="updateTime" value="1751441957804" />
           </Conversation>
         </option>
       </WebConversation>
diff --git a/test_ori.py b/test_ori.py
index 8b0101d..1cb7f04 100644
--- a/test_ori.py
+++ b/test_ori.py
@@ -14,6 +14,7 @@ from tools.dataset import get_transform
 from tools.image_joint import merge_imgs
 from configs import trainer_tools
 import yaml
+from datetime import datetime
 
 with open('configs/test.yml', 'r') as f:
     conf = yaml.load(f, Loader=yaml.FullLoader)
diff --git a/train_compare.py b/train_compare.py
index 5592155..b75d570 100644
--- a/train_compare.py
+++ b/train_compare.py
@@ -5,6 +5,10 @@ import torch
 import torch.nn as nn
 import torch.optim as optim
 from tqdm import tqdm
+from torch.nn.parallel import DistributedDataParallel as DDP
+import torch.distributed as dist
+import torch.multiprocessing as mp
+import torch.utils.data.distributed
 
 from model.loss import FocalLoss
 from tools.dataset import load_data
@@ -12,136 +16,294 @@ import matplotlib.pyplot as plt
 from configs import trainer_tools
 import yaml
 from datetime import datetime
-with open('configs/scatter.yml', 'r') as f:
-    conf = yaml.load(f, Loader=yaml.FullLoader)
 
-# Data Setup
-train_dataloader, class_num = load_data(training=True, cfg=conf)
-val_dataloader, _ = load_data(training=False, cfg=conf)
+
+def load_config(config_path='configs/scatter.yml'):
+    """加载配置文件."""
+    with open(config_path, 'r') as f:
+        return yaml.load(f, Loader=yaml.FullLoader)
+
+
+# 加载配置
+conf = load_config()
+
+
+# 数据加载封装
+def load_datasets():
+    """加载训练和验证数据集，并为分布式训练创建 DistributedSampler."""
+    train_dataset, class_num = load_data(training=True, cfg=conf)
+    val_dataset, _ = load_data(training=False, cfg=conf)
+
+    if conf['base']['distributed']:
+        train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
+        val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)
+    else:
+        train_sampler = None
+        val_sampler = None
+
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        batch_size=conf['data']['batch_size'],
+        shuffle=(train_sampler is None),
+        num_workers=conf['data']['num_workers'],
+        pin_memory=conf['data']['pin_memory'],
+        sampler=train_sampler
+    )
+
+    val_dataloader = torch.utils.data.DataLoader(
+        val_dataset,
+        batch_size=conf['data']['batch_size'],
+        shuffle=False,
+        num_workers=conf['data']['num_workers'],
+        pin_memory=conf['data']['pin_memory'],
+        sampler=val_sampler
+    )
+
+    return train_dataloader, val_dataloader, class_num
+
+
+# 加载数据集
+train_dataloader, val_dataloader, class_num = load_datasets()
 
 tr_tools = trainer_tools(conf)
 backbone_mapping = tr_tools.get_backbone()
 metric_mapping = tr_tools.get_metric(class_num)
 
+
+# 设备管理封装
+def get_device(device_config=None):
+    """根据配置返回设备（CPU/GPU）。在分布式环境下初始化进程组."""
+    if device_config is None:
+        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    else:
+        device = torch.device(device_config)
+
+    if conf['base']['distributed']:
+        dist.init_process_group(backend='nccl')
+
+    return device
+
+
+# 获取设备
+device = get_device(conf['base']['device'])
+
+# 模型初始化
 if conf['models']['backbone'] in backbone_mapping:
-    model = backbone_mapping[conf['models']['backbone']]().to(conf['base']['device'])
+    model = backbone_mapping[conf['models']['backbone']]().to(device)
 else:
     raise ValueError('不支持该模型: {}'.format({conf['models']['backbone']}))
 
 if conf['training']['metric'] in metric_mapping:
-    metric = metric_mapping[conf['training']['metric']]()
+    metric = metric_mapping[conf['training']['metric']]().to(device)
 else:
     raise ValueError('不支持的metric类型: {}'.format(conf['training']['metric']))
 
+rank = 0
 if torch.cuda.device_count() > 1 and conf['base']['distributed']:
     print("Let's use", torch.cuda.device_count(), "GPUs!")
-    model = nn.DataParallel(model)
-    metric = nn.DataParallel(metric)
+    dist.barrier()
+    model = DDP(model, device_ids=[rank])
+    metric = DDP(metric, device_ids=[rank])
 
 # Training Setup
-if conf['training']['loss'] == 'focal_loss':
-    criterion = FocalLoss(gamma=2)
-else:
-    criterion = nn.CrossEntropyLoss()
+def initialize_components():
+    # 封装模型、损失函数和优化器的初始化
+    if conf['training']['loss'] == 'focal_loss':
+        criterion = FocalLoss(gamma=2)
+    else:
+        criterion = nn.CrossEntropyLoss()
 
-optimizer_mapping = tr_tools.get_optimizer(model, metric)
-if conf['training']['optimizer'] in optimizer_mapping:
-    optimizer = optimizer_mapping[conf['training']['optimizer']]()
-    scheduler_mapping = tr_tools.get_scheduler(optimizer)
-    scheduler = scheduler_mapping[conf['training']['scheduler']]()
-    print('使用{}优化器 使用{}调度器'.format(conf['training']['optimizer'],
-                                             conf['training']['scheduler']))
+    optimizer_mapping = tr_tools.get_optimizer(model, metric)
+    if conf['training']['optimizer'] in optimizer_mapping:
+        optimizer = optimizer_mapping[conf['training']['optimizer']]()
+        scheduler_mapping = tr_tools.get_scheduler(optimizer)
+        scheduler = scheduler_mapping[conf['training']['scheduler']]()
+        print('使用{}优化器 使用{}调度器'.format(conf['training']['optimizer'],
+                                                 conf['training']['scheduler']))
+        return criterion, optimizer, scheduler
+    else:
+        raise ValueError('不支持的优化器类型: {}'.format(conf['training']['optimizer']))
 
-else:
-    raise ValueError('不支持的优化器类型: {}'.format(conf['training']['optimizer']))
+
+# 初始化组件
+criterion, optimizer, scheduler = initialize_components()
 
 # Checkpoints Setup
 checkpoints = conf['training']['checkpoints']
 os.makedirs(checkpoints, exist_ok=True)
 
+
+def train_epoch(model, dataloader, optimizer, criterion, device):
+    model.train()
+    train_loss = 0
+    for data, labels in tqdm(dataloader, desc="Training", ascii=True, total=len(dataloader)):
+        data, labels = data.to(device), labels.to(device)
+        embeddings = model(data).to(device)
+        if not conf['training']['metric'] == 'softmax':
+            thetas = metric(embeddings, labels)
+        else:
+            thetas = metric(embeddings)
+        loss = criterion(thetas, labels)
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        train_loss += loss.item()
+    return train_loss / len(dataloader)
+
+
+def validate_epoch(model, dataloader, criterion, device):
+    model.eval()
+    val_loss = 0
+    with torch.no_grad():
+        for data, labels in tqdm(dataloader, desc="Validation", ascii=True, total=len(dataloader)):
+            data, labels = data.to(device), labels.to(device)
+            embeddings = model(data).to(device)
+            if not conf['training']['metric'] == 'softmax':
+                thetas = metric(embeddings, labels)
+            else:
+                thetas = metric(embeddings)
+            loss = criterion(thetas, labels)
+            val_loss += loss.item()
+    return val_loss / len(dataloader)
+
+
+def save_model(model, path, distributed):
+    if distributed and torch.cuda.device_count() > 1:
+        if dist.get_rank() == 0:
+            torch.save(model.module.state_dict(), path)
+    else:
+        torch.save(model.state_dict(), path)
+
+
+def write_log(log_info, log_dir):
+    with open(log_dir, 'a') as f:
+        f.write(log_info + '\n')
+
+
+def plot_losses(epochs, train_losses, val_losses, save_path):
+    plt.plot(epochs, train_losses, color='blue', label='Train Loss')
+    plt.plot(epochs, val_losses, color='red', label='Validation Loss')
+    plt.xlabel('Epochs')
+    plt.ylabel('Loss')
+    plt.legend()
+    plt.savefig(save_path)
+    plt.close()
+
+
+# 模型恢复封装
+def restore_model(model, device):
+    if conf['training']['restore']:
+        print('load pretrain model: {}'.format(conf['training']['restore_model']))
+        model.load_state_dict(torch.load(conf['training']['restore_model'], map_location=device))
+    return model
+
+
+# 日志和学习率记录封装
+def log_and_print(e, train_loss_avg, val_loss_avg, current_lr, log_dir):
+    if conf['base']['distributed'] and dist.get_rank() != 0:
+        return
+
+    log_info = ("[{:%Y-%m-%d %H:%M:%S}] Epoch {}/{},  train_loss: {}, val_loss: {} lr:{}"
+                .format(datetime.now(),
+                        e,
+                        conf['training']['epochs'],
+                        train_loss_avg,
+                        val_loss_avg,
+                        current_lr))
+    print(log_info)
+    write_log(log_info, log_dir)
+    print("第%d个epoch的学习率：%f" % (e, current_lr))
+
+
+# 模型评估与保存封装
+def evaluate_and_save(val_loss_avg, best_loss, model, checkpoints, distributed):
+    if val_loss_avg < best_loss:
+        best_path = osp.join(checkpoints, 'best.pth')
+        save_model(model, best_path, distributed)
+        best_loss = val_loss_avg
+    return best_loss
+
+
+def run_training(rank, world_size, conf):
+    """在指定 rank 上运行训练 loop。
+
+    Args:
+        rank: 当前进程的索引。
+        world_size: 进程总数。
+        conf: 配置字典。
+    """
+    dist.init_process_group(backend='nccl', rank=rank, world_size=world_size)
+    device = torch.device('cuda', rank)
+
+    # 数据加载器和模型等需要重新初始化以确保每个进程独立工作
+    train_dataloader, val_dataloader, class_num = load_datasets()
+    tr_tools = trainer_tools(conf)
+    backbone_mapping = tr_tools.get_backbone()
+    metric_mapping = tr_tools.get_metric(class_num)
+
+    # 模型初始化
+    if conf['models']['backbone'] in backbone_mapping:
+        model = backbone_mapping[conf['models']['backbone']]().to(device)
+    else:
+        raise ValueError('不支持该模型: {}'.format({conf['models']['backbone']}))
+
+    if conf['training']['metric'] in metric_mapping:
+        metric = metric_mapping[conf['training']['metric']]().to(device)
+    else:
+        raise ValueError('不支持的metric类型: {}'.format(conf['training']['metric']))
+
+    model = DDP(model, device_ids=[rank])
+    metric = DDP(metric, device_ids=[rank])
+
+    # 初始化组件
+    criterion, optimizer, scheduler = initialize_components()
+
+    # 恢复模型（如果需要）
+    model = restore_model(model, device)
+
+    train_losses = []
+    val_losses = []
+    epochs = []
+    temp_loss = 1000
+    log_dir = osp.join(conf['logging']['logging_dir'])
+
+    for e in range(conf['training']['epochs']):
+        train_loss_avg = train_epoch(model, train_dataloader, optimizer, criterion, device)
+        train_losses.append(train_loss_avg)
+
+        val_loss_avg = validate_epoch(model, val_dataloader, criterion, device)
+        val_losses.append(val_loss_avg)
+
+        temp_loss = evaluate_and_save(val_loss_avg, temp_loss, model, conf['training']['checkpoints'], True)
+
+        scheduler.step()
+        current_lr = optimizer.param_groups[0]['lr']
+        log_and_print(e, train_loss_avg, val_loss_avg, current_lr, log_dir)
+        epochs.append(e)
+
+    last_path = osp.join(conf['training']['checkpoints'], 'last.pth')
+    save_model(model, last_path, True)
+    plot_losses(epochs, train_losses, val_losses, 'loss/mobilenetv3Large_2250_0316.png')
+
+    dist.destroy_process_group()
+
+
+def main():
+    world_size = torch.cuda.device_count()
+    mp.spawn(
+        run_training,
+        args=(world_size, conf),
+        nprocs=world_size,
+        join=True,
+    )
+
+
 if __name__ == '__main__':
     print('backbone>{}  '.format(conf['models']['backbone']),
           'metric>{}  '.format(conf['training']['metric']),
           'checkpoints>{}  '.format(conf['training']['checkpoints']),
           )
-    train_losses = []
-    val_losses = []
-    epochs = []
-    temp_loss = 100
-    if conf['training']['restore']:
-        print('load pretrain model: {}'.format(conf['training']['restore_model']))
-        model.load_state_dict(torch.load(conf['training']['restore_model'],
-                                         map_location=conf['base']['device']))
-
-    for e in range(conf['training']['epochs']):
-        train_loss = 0
-        model.train()
-
-        for train_data, train_labels in tqdm(train_dataloader,
-                                             desc="Epoch {}/{}"
-                                                     .format(e, conf['training']['epochs']),
-                                             ascii=True,
-                                             total=len(train_dataloader)):
-            train_data = train_data.to(conf['base']['device'])
-            train_labels = train_labels.to(conf['base']['device'])
-
-            train_embeddings = model(train_data).to(conf['base']['device'])  # [256,512]
-            # pdb.set_trace()
-
-            if not conf['training']['metric'] == 'softmax':
-                thetas = metric(train_embeddings, train_labels)  # [256,357]
-            else:
-                thetas = metric(train_embeddings)
-            tloss = criterion(thetas, train_labels)
-            optimizer.zero_grad()
-            tloss.backward()
-            optimizer.step()
-            train_loss += tloss.item()
-        train_lossAvg = train_loss / len(train_dataloader)
-        train_losses.append(train_lossAvg)
-        epochs.append(e)
-        val_loss = 0
-        model.eval()
-        with torch.no_grad():
-            for val_data, val_labels in tqdm(val_dataloader, desc="val",
-                                             ascii=True, total=len(val_dataloader)):
-                val_data = val_data.to(conf['base']['device'])
-                val_labels = val_labels.to(conf['base']['device'])
-                val_embeddings = model(val_data).to(conf['base']['device'])
-                if not conf['training']['metric'] == 'softmax':
-                    thetas = metric(val_embeddings, val_labels)
-                else:
-                    thetas = metric(val_embeddings)
-                vloss = criterion(thetas, val_labels)
-                val_loss += vloss.item()
-            val_lossAvg = val_loss / len(val_dataloader)
-            val_losses.append(val_lossAvg)
-            if val_lossAvg < temp_loss:
-                if torch.cuda.device_count() > 1:
-                    torch.save(model.state_dict(), osp.join(checkpoints, 'best.pth'))
-                else:
-                    torch.save(model.state_dict(), osp.join(checkpoints, 'best.pth'))
-                temp_loss = val_lossAvg
-
-        scheduler.step()
-        current_lr = optimizer.param_groups[0]['lr']
-        log_info = ("[{:%Y-%m-%d %H:%M:%S}] Epoch {}/{},  train_loss: {}, val_loss: {} lr:{}"
-                    .format(datetime.now(),
-                            e,
-                            conf['training']['epochs'],
-                            train_lossAvg,
-                            val_lossAvg,
-                            current_lr))
-        print(log_info)
-        # 写入日志文件
-        with open(osp.join(conf['logging']['logging_dir']), 'a') as f:
-            f.write(log_info + '\n')
-        print("第%d个epoch的学习率：%f" % (e, current_lr))
-    if torch.cuda.device_count() > 1 and conf['base']['distributed']:
-        torch.save(model.module.state_dict(), osp.join(checkpoints, 'last.pth'))
+    if conf['base']['distributed']:
+        main()
     else:
-        torch.save(model.state_dict(), osp.join(checkpoints, 'last.pth'))
-    plt.plot(epochs, train_losses, color='blue')
-    plt.plot(epochs, val_losses, color='red')
-    # plt.savefig('lossMobilenetv3.png')
-    plt.savefig('loss/mobilenetv3Large_2250_0316.png')
+        run_training(rank=0, world_size=1, conf=conf)
diff --git a/train_compare.py.bak b/train_compare.py.bak
new file mode 100644
index 0000000..5592155
--- /dev/null
+++ b/train_compare.py.bak
@@ -0,0 +1,147 @@
+import os
+import os.path as osp
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from tqdm import tqdm
+
+from model.loss import FocalLoss
+from tools.dataset import load_data
+import matplotlib.pyplot as plt
+from configs import trainer_tools
+import yaml
+from datetime import datetime
+with open('configs/scatter.yml', 'r') as f:
+    conf = yaml.load(f, Loader=yaml.FullLoader)
+
+# Data Setup
+train_dataloader, class_num = load_data(training=True, cfg=conf)
+val_dataloader, _ = load_data(training=False, cfg=conf)
+
+tr_tools = trainer_tools(conf)
+backbone_mapping = tr_tools.get_backbone()
+metric_mapping = tr_tools.get_metric(class_num)
+
+if conf['models']['backbone'] in backbone_mapping:
+    model = backbone_mapping[conf['models']['backbone']]().to(conf['base']['device'])
+else:
+    raise ValueError('不支持该模型: {}'.format({conf['models']['backbone']}))
+
+if conf['training']['metric'] in metric_mapping:
+    metric = metric_mapping[conf['training']['metric']]()
+else:
+    raise ValueError('不支持的metric类型: {}'.format(conf['training']['metric']))
+
+if torch.cuda.device_count() > 1 and conf['base']['distributed']:
+    print("Let's use", torch.cuda.device_count(), "GPUs!")
+    model = nn.DataParallel(model)
+    metric = nn.DataParallel(metric)
+
+# Training Setup
+if conf['training']['loss'] == 'focal_loss':
+    criterion = FocalLoss(gamma=2)
+else:
+    criterion = nn.CrossEntropyLoss()
+
+optimizer_mapping = tr_tools.get_optimizer(model, metric)
+if conf['training']['optimizer'] in optimizer_mapping:
+    optimizer = optimizer_mapping[conf['training']['optimizer']]()
+    scheduler_mapping = tr_tools.get_scheduler(optimizer)
+    scheduler = scheduler_mapping[conf['training']['scheduler']]()
+    print('使用{}优化器 使用{}调度器'.format(conf['training']['optimizer'],
+                                             conf['training']['scheduler']))
+
+else:
+    raise ValueError('不支持的优化器类型: {}'.format(conf['training']['optimizer']))
+
+# Checkpoints Setup
+checkpoints = conf['training']['checkpoints']
+os.makedirs(checkpoints, exist_ok=True)
+
+if __name__ == '__main__':
+    print('backbone>{}  '.format(conf['models']['backbone']),
+          'metric>{}  '.format(conf['training']['metric']),
+          'checkpoints>{}  '.format(conf['training']['checkpoints']),
+          )
+    train_losses = []
+    val_losses = []
+    epochs = []
+    temp_loss = 100
+    if conf['training']['restore']:
+        print('load pretrain model: {}'.format(conf['training']['restore_model']))
+        model.load_state_dict(torch.load(conf['training']['restore_model'],
+                                         map_location=conf['base']['device']))
+
+    for e in range(conf['training']['epochs']):
+        train_loss = 0
+        model.train()
+
+        for train_data, train_labels in tqdm(train_dataloader,
+                                             desc="Epoch {}/{}"
+                                                     .format(e, conf['training']['epochs']),
+                                             ascii=True,
+                                             total=len(train_dataloader)):
+            train_data = train_data.to(conf['base']['device'])
+            train_labels = train_labels.to(conf['base']['device'])
+
+            train_embeddings = model(train_data).to(conf['base']['device'])  # [256,512]
+            # pdb.set_trace()
+
+            if not conf['training']['metric'] == 'softmax':
+                thetas = metric(train_embeddings, train_labels)  # [256,357]
+            else:
+                thetas = metric(train_embeddings)
+            tloss = criterion(thetas, train_labels)
+            optimizer.zero_grad()
+            tloss.backward()
+            optimizer.step()
+            train_loss += tloss.item()
+        train_lossAvg = train_loss / len(train_dataloader)
+        train_losses.append(train_lossAvg)
+        epochs.append(e)
+        val_loss = 0
+        model.eval()
+        with torch.no_grad():
+            for val_data, val_labels in tqdm(val_dataloader, desc="val",
+                                             ascii=True, total=len(val_dataloader)):
+                val_data = val_data.to(conf['base']['device'])
+                val_labels = val_labels.to(conf['base']['device'])
+                val_embeddings = model(val_data).to(conf['base']['device'])
+                if not conf['training']['metric'] == 'softmax':
+                    thetas = metric(val_embeddings, val_labels)
+                else:
+                    thetas = metric(val_embeddings)
+                vloss = criterion(thetas, val_labels)
+                val_loss += vloss.item()
+            val_lossAvg = val_loss / len(val_dataloader)
+            val_losses.append(val_lossAvg)
+            if val_lossAvg < temp_loss:
+                if torch.cuda.device_count() > 1:
+                    torch.save(model.state_dict(), osp.join(checkpoints, 'best.pth'))
+                else:
+                    torch.save(model.state_dict(), osp.join(checkpoints, 'best.pth'))
+                temp_loss = val_lossAvg
+
+        scheduler.step()
+        current_lr = optimizer.param_groups[0]['lr']
+        log_info = ("[{:%Y-%m-%d %H:%M:%S}] Epoch {}/{},  train_loss: {}, val_loss: {} lr:{}"
+                    .format(datetime.now(),
+                            e,
+                            conf['training']['epochs'],
+                            train_lossAvg,
+                            val_lossAvg,
+                            current_lr))
+        print(log_info)
+        # 写入日志文件
+        with open(osp.join(conf['logging']['logging_dir']), 'a') as f:
+            f.write(log_info + '\n')
+        print("第%d个epoch的学习率：%f" % (e, current_lr))
+    if torch.cuda.device_count() > 1 and conf['base']['distributed']:
+        torch.save(model.module.state_dict(), osp.join(checkpoints, 'last.pth'))
+    else:
+        torch.save(model.state_dict(), osp.join(checkpoints, 'last.pth'))
+    plt.plot(epochs, train_losses, color='blue')
+    plt.plot(epochs, val_losses, color='red')
+    # plt.savefig('lossMobilenetv3.png')
+    plt.savefig('loss/mobilenetv3Large_2250_0316.png')