update: idfk

filter/RunningLogs.txt

@@ -35,4 +35,3 @@ Note
 0242: V6.1  # 自定义loss
 0259: V6.1  # 调整学习率 (自定义loss)
 0314: V6.1  # 调整学习率（交叉熵损失）
-0349: V6.3  # 增加层数至2

filter/checkpoint_conversion.py

@@ -1,26 +0,0 @@
-import torch
-
-from modelV3_10 import VideoClassifierV3_10
-from modelV3_9 import VideoClassifierV3_9
-
-
-def convert_checkpoint(original_model, new_model):
-    """转换原始checkpoint到新结构"""
-    state_dict = original_model.state_dict()
-    
-    # 直接复制所有参数（因为结构保持兼容）
-    new_model.load_state_dict(state_dict)
-    return new_model
-
-# 使用示例
-original_model = VideoClassifierV3_9()
-new_model = VideoClassifierV3_10()
-
-# 加载原始checkpoint
-original_model.load_state_dict(torch.load('./filter/checkpoints/best_model_V3.9.pt'))
-
-# 转换参数
-converted_model = convert_checkpoint(original_model, new_model)
-
-# 保存转换后的模型
-torch.save(converted_model.state_dict(), './filter/checkpoints/best_model_V3.10.pt')

filter/train.py

@@ -76,9 +76,9 @@ eval_interval = 20
 num_epochs = 20
 total_steps = samples_count * num_epochs / batch_size
 warmup_rate = 0.1
-optimizer = optim.AdamW(model.parameters(), lr=5e-5, weight_decay=1e-5)
+optimizer = optim.AdamW(model.parameters(), lr=1e-5, weight_decay=1e-3)
 cosine_annealing_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=total_steps - int(total_steps * warmup_rate))
-warmup_scheduler = optim.lr_scheduler.LinearLR(optimizer, start_factor=0.1, end_factor=1.0, total_iters=int(total_steps * warmup_rate))
+warmup_scheduler = optim.lr_scheduler.LinearLR(optimizer, start_factor=0.4, end_factor=1.0, total_iters=int(total_steps * warmup_rate))
 scheduler = optim.lr_scheduler.SequentialLR(optimizer, schedulers=[warmup_scheduler, cosine_annealing_scheduler], milestones=[int(total_steps * warmup_rate)])
 criterion = nn.CrossEntropyLoss(weight=class_weights).to(device)