chore: 调整损失值计算

ginka/train_pretrain_split.py

@@ -37,7 +37,7 @@ from .utils import masked_focal
 BATCH_SIZE   = 64
 NUM_CLASSES  = 7
 MAP_SIZE     = 13 * 13
-FOCAL_GAMMA  = 2.0
+FOCAL_GAMMA  = 1.0
 
 # 通道 1：空间骨架（floor+wall）
 CH1_KEEP     = {0, 1}          # 编码器输入保留的 tile
@@ -47,13 +47,13 @@ CH1_NHEAD    = 8
 
 # 通道 2：关卡门控
 CH2_KEEP     = {0, 1, 2, 4, 5}
-CH2_LOSS     = {0, 1, 2, 4, 5}
+CH2_LOSS     = {2, 4, 5}
 CH2_D_MODEL  = 64
 CH2_NHEAD    = 8
 
 # 通道 3：收集资源
 CH3_KEEP     = None            # 完整地图，无需切片
-CH3_LOSS     = {0, 1, 2, 3, 4, 5}
+CH3_LOSS     = {3}
 CH3_D_MODEL  = 64
 CH3_NHEAD    = 8
 

ginka/utils.py

@@ -75,9 +75,9 @@ def masked_focal(
     # count[0]，导致 weight[0] 趋近于 0、非专属位置损失被消除的问题
     class_weight = None
     if balance:
-        flat   = target.view(-1)                                # [B*S] 原始标签
+        flat   = corrected.view(-1)                                # [B*S] 原始标签
         counts = torch.bincount(flat, minlength=C).float()     # [C]
-        class_weight = flat.numel() / (counts.clamp(min=1.0) * C)
+        class_weight = torch.sqrt(flat.numel() / (counts.clamp(min=1.0) * C))
         class_weight[counts == 0] = 0.0                        # 未出现类别不参与
 
     ce = F.cross_entropy(

train_full.sh

@@ -24,17 +24,17 @@ TRAIN_DATA="ginka-dataset.json"
 EVAL_DATA="ginka-eval.json"
 
 # 阶段 0：三通道分拆预训练
-P0_EPOCHS=10
-P0_CHECKPOINT=5
+P0_EPOCHS=20
+P0_CHECKPOINT=10
 P0_FINAL="result/pretrain_split/split_final.pth"
 
 # 阶段 1：冻结编码器热身
-P1_EPOCHS=10
-P1_CHECKPOINT=5
+P1_EPOCHS=30
+P1_CHECKPOINT=10
 P1_FINAL="result/joint/warmup_final.pth"
 
 # 阶段 2：完整联合训练
-P2_EPOCHS=400
+P2_EPOCHS=470
 P2_CHECKPOINT=20
 
 # 从哪个阶段开始（0 = 从头）；命令行 --skip N 可覆盖此值