update: seperate two classifiers

filter/RunningLogs.txt

@@ -24,3 +24,6 @@ Note
 2243: V3.11 # 256维嵌入
 2253: V3.11 # 1024维度嵌入（对比）
 2337: V3.12 # 级联分类
+2350: V3.13 # V3.12, 换用普通交叉熵损失
+0012: V3.11 # 换用普通交叉熵损失
+0039: V3.11 # 级联分类，但使用两个独立模型

filter/labeling_system.py

@@ -10,7 +10,7 @@ import tty
 import termios
 from sentence_transformers import SentenceTransformer
 from db_utils import fetch_entry_data, parse_entry_data
-from modelV3_9 import VideoClassifierV3_9
+from modelV3_10 import VideoClassifierV3_10
 
 class LabelingSystem:
     def __init__(self, mode='model_testing', database_path="./data/main.db", 
@@ -27,7 +27,7 @@ class LabelingSystem:
         self.model = None
         self.sentence_transformer = None
         if self.mode == 'model_testing':
-            self.model = VideoClassifierV3_9()
+            self.model = VideoClassifierV3_10()
             self.model.load_state_dict(torch.load(model_path))
             self.model.eval()
             self.sentence_transformer = SentenceTransformer("Thaweewat/jina-embedding-v3-m2v-1024")

filter/modelV3_10.py

@@ -3,13 +3,13 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 class VideoClassifierV3_10(nn.Module):
-    def __init__(self, embedding_dim=1024, hidden_dim=648, output_dim=3):
+    def __init__(self, embedding_dim=1024, hidden_dim=648, output_dim=3, temperature=1.7):
         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))
+        self.temperature = nn.Parameter(torch.tensor(temperature))
         
         # 带约束的通道权重（使用Sigmoid替代Softmax）
         self.channel_weights = nn.Parameter(torch.ones(self.num_channels))

filter/test.py

@@ -1,7 +1,7 @@
 from labeling_system import LabelingSystem
 
 DATABASE_PATH = "./data/main.db"
-MODEL_PATH = "./filter/checkpoints/best_model_V3.9.pt"
+MODEL_PATH = "./filter/checkpoints/best_model_V3.11.pt"
 OUTPUT_FILE = "./data/filter/real_test.jsonl"
 BATCH_SIZE = 50
 

filter/train.py

@@ -4,7 +4,7 @@ import numpy as np
 from torch.utils.data import DataLoader
 import torch.optim as optim
 from dataset import MultiChannelDataset
-from filter.modelV3_12 import VideoClassifierV3_12
+from filter.modelV3_10 import VideoClassifierV3_10
 from sklearn.metrics import f1_score, recall_score, precision_score, accuracy_score, classification_report
 import os
 import torch
@@ -51,30 +51,39 @@ class_weights = torch.tensor(
 )
 
 # 初始化模型和SentenceTransformer
-model = VideoClassifierV3_12()
+model1 = VideoClassifierV3_10(output_dim=2, hidden_dim=384)
-checkpoint_name = './filter/checkpoints/best_model_V3.12.pt'
+model2 = VideoClassifierV3_10(output_dim=2, hidden_dim=384)
+checkpoint1_name = './filter/checkpoints/best_model_V3.14-part1.pt'
+checkpoint2_name = './filter/checkpoints/best_model_V3.14-part2.pt'
 
 # 模型保存路径
 os.makedirs('./filter/checkpoints', exist_ok=True)
 
 # 优化器
-optimizer = optim.AdamW(model.parameters(), lr=4e-4)
+optimizer1 = optim.AdamW(model1.parameters(), lr=4e-4)
+optimizer2 = optim.AdamW(model2.parameters(), lr=4e-4)
 # Cross entropy loss
-criterion = nn.CrossEntropyLoss()
+criterion1 = nn.CrossEntropyLoss()
+criterion2 = nn.CrossEntropyLoss()
 
 def count_trainable_parameters(model):
     return sum(p.numel() for p in model.parameters() if p.requires_grad)
 
-def evaluate(model, dataloader):
+def evaluate(model1, model2, dataloader):
-    model.eval()
+    model1.eval()
+    model2.eval()
     all_preds = []
     all_labels = []
     
     with torch.no_grad():
         for batch in dataloader:
             batch_tensor = prepare_batch(batch['texts'])
-            logits = model(batch_tensor)
+            logits1 = model1(batch_tensor)
-            preds = torch.argmax(logits, dim=1)
+            logits2 = model2(batch_tensor)
+            preds1 = torch.argmax(logits1, dim=1)
+            preds2 = torch.argmax(logits2, dim=1)
+            # 如果preds1输出为0，那么预测结果为0，否则使用preds2的结果加上1
+            preds = torch.where(preds1 == 0, preds1, preds2 + 1)
             all_preds.extend(preds.cpu().numpy())
             all_labels.extend(batch['label'].cpu().numpy())
     
@@ -86,10 +95,10 @@ def evaluate(model, dataloader):
     
     # 获取每个类别的详细指标
     class_report = classification_report(all_labels, all_preds, output_dict=True)
-    
-    return f1, recall, precision, accuracy, class_report
 
-print(f"Trainable parameters: {count_trainable_parameters(model)}")
+    return f1, recall, precision, accuracy, class_report
+    
+print(f"Trainable parameters: {count_trainable_parameters(model1) + count_trainable_parameters(model2)}")
 
 # 训练循环
 best_f1 = 0
@@ -98,29 +107,44 @@ eval_interval = 20
 num_epochs = 8
 
 for epoch in range(num_epochs):
-    model.train()
+    model1.train()
-    epoch_loss = 0
+    model2.train()
+    epoch_loss_1 = 0
+    epoch_loss_2 = 0
     
     # 训练阶段
     for batch_idx, batch in enumerate(train_loader):
-        optimizer.zero_grad()
+        optimizer1.zero_grad()
+        optimizer2.zero_grad()
         
         batch_tensor = prepare_batch(batch['texts'])
+        batch_tensor_1 = batch_tensor
+        mask = batch['label'] != 0 
+        batch_tensor_2 = batch_tensor_1[mask]
 
-        logits = model(batch_tensor)
+        logits1 = model1(batch_tensor_1)
+        logits2 = model2(batch_tensor_2)
         
-        loss = criterion(logits, batch['label'])
+        label1 = torch.where(batch['label'] == 0, 0, 1)
-        loss.backward()
+        label2 = torch.where(batch['label'][mask] == 1, 0, 1)
-        optimizer.step()
+        loss1 = criterion1(logits1, label1)
-        epoch_loss += loss.item()
+        loss1.backward()
+        loss2 = criterion2(logits2, label2)
+        loss2.backward()
+        optimizer1.step()
+        optimizer2.step()
+        epoch_loss_1 += loss1.item()
+        epoch_loss_2 += loss2.item()
         
         # 记录训练损失
-        writer.add_scalar('Train/Loss', loss.item(), step)
+        writer.add_scalar('Train/Loss-1', loss1.item(), step)
+        writer.add_scalar('Train/Loss-2', loss2.item(), step)
         step += 1
         
         # 每隔 eval_interval 步执行验证
         if step % eval_interval == 0:
-            eval_f1, eval_recall, eval_precision, eval_accuracy, eval_class_report = evaluate(model, eval_loader)
+            eval_f1, eval_recall, eval_precision, eval_accuracy, eval_class_report = evaluate(model1, model2, eval_loader)
+
             writer.add_scalar('Eval/F1', eval_f1, step)
             writer.add_scalar('Eval/Recall', eval_recall, step)
             writer.add_scalar('Eval/Precision', eval_precision, step)
@@ -136,17 +160,19 @@ for epoch in range(num_epochs):
             # 保存最佳模型
             if eval_f1 > best_f1:
                 best_f1 = eval_f1
-                torch.save(model.state_dict(), checkpoint_name)
+                torch.save(model1.state_dict(), checkpoint1_name)
+                torch.save(model2.state_dict(), checkpoint2_name)
                 print("  Saved best model")
-                print("Channel weights: ", model.get_channel_weights())
+                print("Channel weights: ", model1.get_channel_weights())
+                print("Channel weights: ", model2.get_channel_weights())
     
     # 记录每个 epoch 的平均训练损失
-    avg_epoch_loss = epoch_loss / len(train_loader)
+    avg_epoch_loss = (epoch_loss_1 + epoch_loss_2) / 2 / len(train_loader)
     writer.add_scalar('Train/Epoch_Loss', avg_epoch_loss, epoch)
     
     # 每个 epoch 结束后执行一次完整验证
-    train_f1, train_recall, train_precision, train_accuracy, train_class_report = evaluate(model, train_loader)
+    train_f1, train_recall, train_precision, train_accuracy, train_class_report = evaluate(model1, model2, train_loader)
-    eval_f1, eval_recall, eval_precision, eval_accuracy, eval_class_report = evaluate(model, eval_loader)
+    eval_f1, eval_recall, eval_precision, eval_accuracy, eval_class_report = evaluate(model1, model2, eval_loader)
     
     writer.add_scalar('Train/Epoch_F1', train_f1, epoch)
     writer.add_scalar('Train/Epoch_Recall', train_recall, epoch)
@@ -173,8 +199,9 @@ for epoch in range(num_epochs):
 
 # 测试阶段
 print("\nTesting...")
-model.load_state_dict(torch.load(checkpoint_name))
+model1.load_state_dict(torch.load(checkpoint1_name))
-test_f1, test_recall, test_precision, test_accuracy, test_class_report = evaluate(model, test_loader)
+model2.load_state_dict(torch.load(checkpoint2_name))
+test_f1, test_recall, test_precision, test_accuracy, test_class_report = evaluate(model1, model2, test_loader)
 writer.add_scalar('Test/F1', test_f1, step)
 writer.add_scalar('Test/Recall', test_recall, step)
 writer.add_scalar('Test/Precision', test_precision, step)