diff --git a/filter/train.py b/filter/train.py
index 99566d9..b495aaa 100644
--- a/filter/train.py
+++ b/filter/train.py
@@ -61,7 +61,7 @@ class_weights = torch.tensor(
     device=device
 )
 
-model = VideoClassifierV6_1().to(device)
+model = VideoClassifierV6_1(num_heads=8).to(device)
 checkpoint_name = './filter/checkpoints/best_model_V6.1.pt'
 
 # 初始化tokenizer和embedding模型
@@ -75,10 +75,10 @@ os.makedirs('./filter/checkpoints', exist_ok=True)
 eval_interval = 20
 num_epochs = 20
 total_steps = samples_count * num_epochs / batch_size
-warmup_rate = 0.1
-optimizer = optim.AdamW(model.parameters(), lr=3e-5, weight_decay=1e-3)
+warmup_rate = 0.15
+optimizer = optim.AdamW(model.parameters(), lr=3e-5, weight_decay=1e-4)
 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=0.8, 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)