feat: 提高 VQ 解码头容量

Co-authored-by: Copilot <copilot@github.com>

ginka/train_pretrain.py

@@ -50,8 +50,10 @@ VQ_DIM_FF = 512
 VQ_BETA   = 0.5
 VQ_GAMMA  = 0.0
 
-# 解码头超参
-DH_NHEAD  = 8    # Cross-Attention 头数（VQ_D_Z=128 可被 8 整除）
+# 解码头超参（与编码器对称：同等层数和 FFN 宽度）
+DH_NHEAD   = 8    # Cross-Attention 头数（VQ_D_Z=128 可被 8 整除）
+DH_DIM_FF  = 512  # FFN 隐层维度（与编码器 VQ_DIM_FF 一致）
+DH_LAYERS  = 4    # 解码层数（与编码器 VQ_LAYERS 一致）
 
 # ---------------------------------------------------------------------------
 # 设备
@@ -185,6 +187,8 @@ def train():
         d_z=VQ_D_Z,
         map_size=MAP_SIZE,
         nhead=DH_NHEAD,
+        dim_ff=DH_DIM_FF,
+        num_layers=DH_LAYERS,
     ).to(device)
 
     vq_params = sum(p.numel() for p in model_vq.parameters())

ginka/vqvae/model.py

@@ -4,18 +4,42 @@ from .quantize import VectorQuantizer
 from typing import Tuple
 
 
+class _DecodeLayer(nn.Module):
+    """单个解码层：Pre-LN Cross-Attention + Pre-LN FFN。"""
+
+    def __init__(self, d_z: int, nhead: int, dim_ff: int):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(d_z)
+        self.cross_attn = nn.MultiheadAttention(
+            embed_dim=d_z,
+            num_heads=nhead,
+            batch_first=True,
+            dropout=0.0,
+        )
+        self.norm2 = nn.LayerNorm(d_z)
+        self.ffn = nn.Sequential(
+            nn.Linear(d_z, dim_ff),
+            nn.GELU(),
+            nn.Linear(dim_ff, d_z),
+        )
+
+    def forward(self, x: torch.Tensor, z_q: torch.Tensor) -> torch.Tensor:
+        x = x + self.cross_attn(self.norm1(x), z_q, z_q)[0]   # Pre-LN cross-attn
+        x = x + self.ffn(self.norm2(x))                        # Pre-LN FFN
+        return x
+
+
 class VQDecodeHead(nn.Module):
     """
-    VQ-VAE 预训练用轻量解码头（Cross-Attention 架构）。
+    VQ-VAE 预训练用解码头（堆叠 Cross-Attention + FFN，Pre-LN 风格）。
 
-    将 z_q [B, L, d_z] 通过 Cross-Attention 还原为地图 logits [B, H*W, num_classes]。
+    将 z_q [B, L, d_z] 通过多层 Cross-Attention 解码为地图 logits [B, H*W, num_classes]。
     预训练结束后此模块被丢弃，不影响联合训练路径。
 
-    架构：
-        可学习位置查询 [B, H*W, d_z]
-        → Cross-Attention (query=位置查询, key/value=z_q)
-        → LayerNorm
-        → 线性分类头 → logits [B, H*W, num_classes]
+    架构（每层）：
+        Pre-LN Cross-Attention（query=可学习位置查询, key/value=z_q）
+        Pre-LN FFN
+    × num_layers 层 → LayerNorm → 线性分类头
     """
 
     def __init__(
@@ -23,7 +47,9 @@ class VQDecodeHead(nn.Module):
         num_classes: int,
         d_z: int,
         map_size: int,
-        nhead: int = 4,
+        nhead: int = 8,
+        dim_ff: int = 512,
+        num_layers: int = 4,
     ):
         """
         Args:
@@ -31,22 +57,20 @@ class VQDecodeHead(nn.Module):
             d_z:         z 向量维度（须与 GinkaVQVAE 的 d_z 一致）
             map_size:    地图 token 总数（H * W）
             nhead:       Cross-Attention 的注意力头数（d_z 须能被 nhead 整除）
+            dim_ff:      FFN 隐层维度
+            num_layers:  解码层数（建议与编码器 num_layers 相同）
         """
         super().__init__()
 
         # 每个格子一个可学习位置查询
         self.pos_queries = nn.Parameter(torch.randn(1, map_size, d_z) * 0.02)
 
-        # Cross-Attention：query=位置查询，key/value=z_q
-        self.cross_attn = nn.MultiheadAttention(
-            embed_dim=d_z,
-            num_heads=nhead,
-            batch_first=True,
-            dropout=0.0,
-        )
-        self.norm = nn.LayerNorm(d_z)
+        # 堆叠多层解码块
+        self.layers = nn.ModuleList([
+            _DecodeLayer(d_z, nhead, dim_ff) for _ in range(num_layers)
+        ])
 
-        # 最终分类头
+        self.norm_out = nn.LayerNorm(d_z)
         self.classifier = nn.Linear(d_z, num_classes)
 
     def forward(self, z_q: torch.Tensor) -> torch.Tensor:
@@ -58,10 +82,11 @@ class VQDecodeHead(nn.Module):
             logits: [B, map_size, num_classes]
         """
         B = z_q.shape[0]
-        q = self.pos_queries.expand(B, -1, -1)          # [B, map_size, d_z]
-        out, _ = self.cross_attn(q, z_q, z_q)           # [B, map_size, d_z]
-        out = self.norm(out)
-        return self.classifier(out)                      # [B, map_size, num_classes]
+        x = self.pos_queries.expand(B, -1, -1)   # [B, map_size, d_z]
+        for layer in self.layers:
+            x = layer(x, z_q)
+        x = self.norm_out(x)
+        return self.classifier(x)                # [B, map_size, num_classes]
 
 
 class GinkaVQVAE(nn.Module):