Update evo_model.py

evo_model.py

@@ -3,30 +3,33 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 class EvoEncoder(nn.Module):
-    def __init__(self, d_model=384, nhead=6, dim_feedforward=1024, num_layers=6):
+    def __init__(self, d_model=512, num_heads=8, ffn_dim=1024, num_layers=6, memory_enabled=True):
         super().__init__()
-        self.embedding = nn.Embedding(30522, d_model)  # BERT-base vocab size
-        encoder_layer = nn.TransformerEncoderLayer(
-            d_model=d_model,
-            nhead=nhead,
-            dim_feedforward=dim_feedforward,
-            batch_first=True,
-        )
+        self.embedding = nn.Embedding(30522, d_model)
+        self.positional_encoding = nn.Parameter(torch.zeros(1, 512, d_model))
+        encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=num_heads, dim_feedforward=ffn_dim, batch_first=True)
         self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
-        self.memory_proj = nn.Linear(d_model, d_model)
+        self.norm = nn.LayerNorm(d_model)
+        self.memory_enabled = memory_enabled
+        if memory_enabled:
+            self.memory_proj = nn.Linear(d_model, d_model)
+            self.memory_token = nn.Parameter(torch.zeros(1, 1, d_model))
 
     def forward(self, input_ids):
-        x = self.embedding(input_ids)
+        x = self.embedding(input_ids) + self.positional_encoding[:, :input_ids.size(1), :]
+        if self.memory_enabled:
+            mem = self.memory_token.expand(x.size(0), -1, -1)
+            x = torch.cat([mem, x], dim=1)
         x = self.transformer(x)
-        x = x.mean(dim=1)
-        return self.memory_proj(x)
+        x = self.norm(x)
+        return x[:, 0]  # Return [CLS]-like token (memory or first token)
 
 class EvoTransformer(nn.Module):
-    def __init__(self, d_model=384):
+    def __init__(self, d_model=512, num_heads=8, ffn_dim=1024, num_layers=6, num_classes=1, memory_enabled=True):
         super().__init__()
-        self.encoder = EvoEncoder(d_model=d_model)
-        self.classifier = nn.Linear(d_model, 2)
+        self.encoder = EvoEncoder(d_model, num_heads, ffn_dim, num_layers, memory_enabled)
+        self.classifier = nn.Linear(d_model, num_classes)
 
     def forward(self, input_ids):
         x = self.encoder(input_ids)
-        return x
+        return self.classifier(x)