调整模型结构及参数以优化性能

src/trainer/model.py

@@ -89,8 +89,8 @@ class Expert(nn.Module):
         dropout_prob=0.1,
     ):
         """
-        input_dim : BERT 输出的 hidden_size（如 312/768）
-        d_model   : 专家内部维度（固定 1024）
+        input_dim : 输入维度
+        d_model   : 专家内部维度
         output_multiplier : 输出维度 = input_dim * output_multiplier
         dropout_prob      : 残差块内 Dropout
         """
@@ -154,15 +154,16 @@ class MoEModel(nn.Module):
             batch_first=True,
         )
         self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=4)
-        self.pooler = nn.AdaptiveAvgPool1d(1)
         self.res_blocks = nn.ModuleList([ResidualBlock(self.hidden_size) for _ in range(4)])
 
+        self.pooler = nn.AdaptiveAvgPool1d(2)
+
         self.total_experts = 20
         self.experts = nn.ModuleList()
 
         for i in range(self.total_experts):
             expert = Expert(
-                input_dim=self.hidden_size,
+                input_dim=self.hidden_size * 2,
                 d_model=self.experts_dim[i],
                 num_resblocks=num_resblocks,
                 output_multiplier=self.output_multiplier,  # 输出维度 = 2 * hidden_size
@@ -171,24 +172,24 @@ class MoEModel(nn.Module):
             self.experts.append(expert)
 
         self.expert_bias = nn.Embedding(
-            self.total_experts, self.output_multiplier * self.hidden_size
+            self.total_experts, self.output_multiplier * self.hidden_size * 2
         )
 
         # 4. 分类头
         self.classifier = nn.Sequential(
-            nn.LayerNorm(self.output_multiplier * self.hidden_size),
+            nn.LayerNorm(self.output_multiplier * self.hidden_size * 2),
             nn.Linear(
-                self.output_multiplier * self.hidden_size,
-                self.output_multiplier * self.hidden_size,
-            ),
-            nn.ReLU(inplace=True),
-            nn.Linear(
-                self.output_multiplier * self.hidden_size,
+                self.output_multiplier * self.hidden_size * 2,
                 self.output_multiplier * self.hidden_size * 2,
             ),
             nn.ReLU(inplace=True),
-            nn.Dropout(0.2),
-            nn.Linear(self.output_multiplier * self.hidden_size * 2, num_classes),
+            nn.Linear(
+                self.output_multiplier * self.hidden_size * 2,
+                self.output_multiplier * self.hidden_size * 4,
+            ),
+            nn.ReLU(inplace=True),
+            nn.Dropout(0.1),
+            nn.Linear(self.output_multiplier * self.hidden_size * 4, num_classes),
         )
         # 可选：为领域专家和共享专家设置不同权重衰减（通过优化器实现，此处不处理）
 
@@ -213,11 +214,12 @@ class MoEModel(nn.Module):
             embeddings, src_key_padding_mask=padding_mask
         )  # [B, S, H]
 
-        # ----- 3. 池化量 -----
-        pooled = self.pooler(encoded.transpose(1, 2)).squeeze(-1)
-
         for block in self.res_blocks:
-            pooled = block(pooled)
+            pooled = block(encoded)
+
+        # ----- 3. 池化量 -----
+        pooled = self.pooler(encoded.transpose(1, 2))   # [B, H, 2]
+        pooled = pooled.flatten(1)                      # [B, H*2]
 
         # ----- 4. 专家路由（硬路由）-----
         if torch.jit.is_tracing():
@@ -466,7 +468,7 @@ class MoEModel(nn.Module):
         train_dataloader,
         eval_dataloader=None,
         monitor: Optional[TrainingMonitor] = None,
-        criterion=nn.CrossEntropyLoss(),
+        criterion=None,
         optimizer=None,
         num_epochs=1,
         stop_batch=1e6,
@@ -475,7 +477,7 @@ class MoEModel(nn.Module):
         clip_grad_norm=1.0,
         mixed_precision=False,
         loss_weight=None,
-        lr=1e-4,
+        lr=6e-5,
         lr_schedule=None,  # 新增：可选的自定义学习率调度函数
     ):
         """
@@ -512,17 +514,19 @@ class MoEModel(nn.Module):
         if self.device is None:
             self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
             self.to(self.device)
-        if loss_weight:
-            loss_weight = 1 / torch.sqrt(torch.sqrt(torch.tensor(loss_weight)))
-            self.loss_weight = loss_weight.to(self.device)
-            criterion.weight = self.loss_weight
 
         # 切换到训练模式
         super().train()
 
         # 默认优化器
         if optimizer is None:
-            optimizer = optim.AdamW(self.parameters(), lr=lr)  # 初始学习率 1e-4
+            optimizer = optim.AdamW(self.parameters(), lr=lr)
+        
+        if criterion is None:
+            if loss_weight is not None:
+                criterion = nn.CrossEntropyLoss(weight=loss_weight)
+            else:
+                criterion = nn.CrossEntropyLoss()
 
         # 混合精度缩放器
         scaler = amp.GradScaler(enabled=mixed_precision)
@@ -560,7 +564,7 @@ class MoEModel(nn.Module):
                 scaler.scale(loss).backward()
 
                 # 梯度累积
-                if (batch_idx + 1) % grad_accum_steps == 0:
+                if (batch_idx) % grad_accum_steps == 0:
                     scaler.unscale_(optimizer)
                     torch.nn.utils.clip_grad_norm_(self.parameters(), clip_grad_norm)
                     scaler.step(optimizer)