update: v3.5

2025-01-25 03:29:47 +08:00 · 2025-01-25 03:29:47 +08:00 · 546a9f7bc8
commit 546a9f7bc8
parent 5a83120ad6
4 changed files with 247 additions and 7 deletions
--- a/filter/RunningLogs.txt
+++ b/filter/RunningLogs.txt
@ -14,3 +14,5 @@ Note
 0138: V4.3-test3
 0155: V5-test3 # V4 的效果也不是特别好
 0229: V3.3-test3 # 重新回到V3迭代
+0316: V3.4-test3 # 模型架构修改，自定义Loss与FC层修改
+0324: V3.5-test3 # 用回3.2的FC层试试
--- a/filter/modelV3_4.py
+++ b/filter/modelV3_4.py
@ -0,0 +1,111 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class VideoClassifierV3_4(nn.Module):
+    def __init__(self, embedding_dim=1024, hidden_dim=512, output_dim=3):
+        super().__init__()
+        self.num_channels = 4
+        self.channel_names = ['title', 'description', 'tags', 'author_info']
+        
+        # 可学习温度系数
+        self.temperature = nn.Parameter(torch.tensor(1.7))
+        
+        # 带约束的通道权重（使用Sigmoid替代Softmax）
+        self.channel_weights = nn.Parameter(torch.ones(self.num_channels))
+        
+        # 增强的非线性层
+        self.fc = nn.Sequential(
+            nn.Linear(embedding_dim * self.num_channels, hidden_dim*2),
+            nn.BatchNorm1d(hidden_dim*2),
+            nn.Dropout(0.3),
+            nn.GELU(),
+            nn.Linear(hidden_dim*2, hidden_dim),
+            nn.BatchNorm1d(hidden_dim),
+            nn.Dropout(0.2),
+            nn.GELU(),
+            nn.Linear(hidden_dim, output_dim)
+        )
+        
+        # 权重初始化
+        self._init_weights()
+    
+    def _init_weights(self):
+        for layer in self.fc:
+            if isinstance(layer, nn.Linear):
+                # 使用ReLU的初始化参数（GELU的近似）
+                nn.init.kaiming_normal_(layer.weight, nonlinearity='relu')  # 修改这里
+                
+                # 或者使用Xavier初始化（更适合通用场景）
+                # nn.init.xavier_normal_(layer.weight, gain=nn.init.calculate_gain('relu'))
+                
+                nn.init.zeros_(layer.bias)
+
+
+    def forward(self, input_texts, sentence_transformer):
+        # 合并文本进行批量编码
+        all_texts = [text for channel in self.channel_names for text in input_texts[channel]]
+        
+        # 冻结的文本编码
+        with torch.no_grad():
+            embeddings = torch.tensor(
+                sentence_transformer.encode(all_texts),
+                device=next(self.parameters()).device
+            )
+        
+        # 分割并加权通道特征
+        split_sizes = [len(input_texts[name]) for name in self.channel_names]
+        channel_features = torch.split(embeddings, split_sizes, dim=0)
+        channel_features = torch.stack(channel_features, dim=1)
+        
+        # 自适应通道权重（Sigmoid约束）
+        weights = torch.sigmoid(self.channel_weights)  # [0,1]范围
+        weighted_features = channel_features * weights.unsqueeze(0).unsqueeze(-1)
+        
+        # 特征拼接
+        combined = weighted_features.view(weighted_features.size(0), -1)
+        
+        return self.fc(combined)
+
+    def get_channel_weights(self):
+        """获取各通道权重（带温度调节）"""
+        return torch.softmax(self.channel_weights / self.temperature, dim=0).detach().cpu().numpy()
+    
+
+class AdaptiveRecallLoss(nn.Module):
+    def __init__(self, class_weights, alpha=0.8, gamma=2.0, fp_penalty=0.5):
+        """
+        Args:
+            class_weights (torch.Tensor): 类别权重
+            alpha (float): 召回率调节因子（0-1）
+            gamma (float): Focal Loss参数
+            fp_penalty (float): 类别0假阳性惩罚强度
+        """
+        super().__init__()
+        self.class_weights = class_weights
+        self.alpha = alpha
+        self.gamma = gamma
+        self.fp_penalty = fp_penalty
+
+    def forward(self, logits, targets):
+        # 基础交叉熵损失
+        ce_loss = F.cross_entropy(logits, targets, weight=self.class_weights, reduction='none')
+        
+        # Focal Loss组件
+        pt = torch.exp(-ce_loss)
+        focal_loss = ((1 - pt) ** self.gamma) * ce_loss
+        
+        # 召回率增强（对困难样本加权）
+        class_mask = F.one_hot(targets, num_classes=len(self.class_weights))
+        class_weights = (self.alpha + (1 - self.alpha) * pt.unsqueeze(-1)) * class_mask
+        recall_loss = (class_weights * focal_loss.unsqueeze(-1)).sum(dim=1)
+        
+        # 类别0假阳性惩罚
+        probs = F.softmax(logits, dim=1)
+        fp_mask = (targets != 0) & (torch.argmax(logits, dim=1) == 0)
+        fp_loss = self.fp_penalty * probs[:, 0][fp_mask].pow(2).sum()
+        
+        # 总损失
+        total_loss = recall_loss.mean() + fp_loss / len(targets)
+        
+        return total_loss
--- a/filter/modelV3_5.py
+++ b/filter/modelV3_5.py
@ -0,0 +1,109 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class VideoClassifierV3_5(nn.Module):
+    def __init__(self, embedding_dim=1024, hidden_dim=512, output_dim=3):
+        super().__init__()
+        self.num_channels = 4
+        self.channel_names = ['title', 'description', 'tags', 'author_info']
+        
+        # 可学习温度系数
+        self.temperature = nn.Parameter(torch.tensor(1.7))
+        
+        # 带约束的通道权重（使用Sigmoid替代Softmax）
+        self.channel_weights = nn.Parameter(torch.ones(self.num_channels))
+        
+        # 增强的非线性层
+        self.fc = nn.Sequential(
+            nn.Linear(embedding_dim * self.num_channels, hidden_dim*2),
+            nn.BatchNorm1d(hidden_dim*2),
+            nn.Dropout(0.1),
+            nn.ReLU(),
+            nn.Linear(hidden_dim*2, hidden_dim),
+            nn.LayerNorm(hidden_dim),
+            nn.Linear(hidden_dim, output_dim)
+        )
+        
+        # 权重初始化
+        self._init_weights()
+    
+    def _init_weights(self):
+        for layer in self.fc:
+            if isinstance(layer, nn.Linear):
+                # 使用ReLU的初始化参数（GELU的近似）
+                nn.init.kaiming_normal_(layer.weight, nonlinearity='relu')  # 修改这里
+                
+                # 或者使用Xavier初始化（更适合通用场景）
+                # nn.init.xavier_normal_(layer.weight, gain=nn.init.calculate_gain('relu'))
+                
+                nn.init.zeros_(layer.bias)
+
+
+    def forward(self, input_texts, sentence_transformer):
+        # 合并文本进行批量编码
+        all_texts = [text for channel in self.channel_names for text in input_texts[channel]]
+        
+        # 冻结的文本编码
+        with torch.no_grad():
+            embeddings = torch.tensor(
+                sentence_transformer.encode(all_texts),
+                device=next(self.parameters()).device
+            )
+        
+        # 分割并加权通道特征
+        split_sizes = [len(input_texts[name]) for name in self.channel_names]
+        channel_features = torch.split(embeddings, split_sizes, dim=0)
+        channel_features = torch.stack(channel_features, dim=1)
+        
+        # 自适应通道权重（Sigmoid约束）
+        weights = torch.sigmoid(self.channel_weights)  # [0,1]范围
+        weighted_features = channel_features * weights.unsqueeze(0).unsqueeze(-1)
+        
+        # 特征拼接
+        combined = weighted_features.view(weighted_features.size(0), -1)
+        
+        return self.fc(combined)
+
+    def get_channel_weights(self):
+        """获取各通道权重（带温度调节）"""
+        return torch.softmax(self.channel_weights / self.temperature, dim=0).detach().cpu().numpy()
+    
+
+class AdaptiveRecallLoss(nn.Module):
+    def __init__(self, class_weights, alpha=0.8, gamma=2.0, fp_penalty=0.5):
+        """
+        Args:
+            class_weights (torch.Tensor): 类别权重
+            alpha (float): 召回率调节因子（0-1）
+            gamma (float): Focal Loss参数
+            fp_penalty (float): 类别0假阳性惩罚强度
+        """
+        super().__init__()
+        self.class_weights = class_weights
+        self.alpha = alpha
+        self.gamma = gamma
+        self.fp_penalty = fp_penalty
+
+    def forward(self, logits, targets):
+        # 基础交叉熵损失
+        ce_loss = F.cross_entropy(logits, targets, weight=self.class_weights, reduction='none')
+        
+        # Focal Loss组件
+        pt = torch.exp(-ce_loss)
+        focal_loss = ((1 - pt) ** self.gamma) * ce_loss
+        
+        # 召回率增强（对困难样本加权）
+        class_mask = F.one_hot(targets, num_classes=len(self.class_weights))
+        class_weights = (self.alpha + (1 - self.alpha) * pt.unsqueeze(-1)) * class_mask
+        recall_loss = (class_weights * focal_loss.unsqueeze(-1)).sum(dim=1)
+        
+        # 类别0假阳性惩罚
+        probs = F.softmax(logits, dim=1)
+        fp_mask = (targets != 0) & (torch.argmax(logits, dim=1) == 0)
+        fp_loss = self.fp_penalty * probs[:, 0][fp_mask].pow(2).sum()
+        
+        # 总损失
+        total_loss = recall_loss.mean() + fp_loss / len(targets)
+        
+        return total_loss
--- a/filter/train.py
+++ b/filter/train.py
@ -1,11 +1,11 @@
 import os
 os.environ["PYTORCH_ENABLE_MPS_FALLBACK"]="1"
+import numpy as np
 from torch.utils.data import DataLoader
 import torch.optim as optim
 from dataset import MultiChannelDataset
-from modelV3_3 import VideoClassifierV3_3
+from modelV3_5 import VideoClassifierV3_5, AdaptiveRecallLoss
 from sentence_transformers import SentenceTransformer
-import torch.nn as nn
 from sklearn.metrics import f1_score, recall_score, precision_score, accuracy_score, classification_report
 import os
 import torch
@ -37,17 +37,35 @@ train_loader = DataLoader(train_dataset, batch_size=24, shuffle=True)
 eval_loader = DataLoader(eval_dataset, batch_size=24, shuffle=False)
 test_loader = DataLoader(test_dataset, batch_size=24, shuffle=False)

+train_labels = []
+for batch in train_loader:
+    train_labels.extend(batch['label'].tolist())
+    
+# 计算自适应类别权重
+class_counts = np.bincount(train_labels)
+median_freq = np.median(class_counts)
+class_weights = torch.tensor(
+    [median_freq / count for count in class_counts],
+    dtype=torch.float32,
+    device='cpu'
+)
+
 # 初始化模型和SentenceTransformer
 sentence_transformer = SentenceTransformer("Thaweewat/jina-embedding-v3-m2v-1024")
-model = VideoClassifierV3_3()
-checkpoint_name = './filter/checkpoints/best_model_V3.3.pt'
+model = VideoClassifierV3_5()
+checkpoint_name = './filter/checkpoints/best_model_V3.5.pt'

 # 模型保存路径
 os.makedirs('./filter/checkpoints', exist_ok=True)

 # 优化器
 optimizer = optim.AdamW(model.parameters(), lr=2e-3)
-criterion = nn.CrossEntropyLoss()
+criterion = AdaptiveRecallLoss(
+    class_weights=class_weights,
+    alpha=0.7,     # 召回率权重
+    gamma=1.5,     # 困难样本聚焦
+    fp_penalty=0.8 # 假阳性惩罚强度
+)

 def count_trainable_parameters(model):
    return sum(p.numel() for p in model.parameters() if p.requires_grad)
@ -81,7 +99,7 @@ print(f"Trainable parameters: {count_trainable_parameters(model)}")
 # 训练循环
 best_f1 = 0
 step = 0
-eval_interval = 50
+eval_interval = 20
 num_epochs = 8

 for epoch in range(num_epochs):