Update evo_model.py

evo_model.py

@@ -3,32 +3,37 @@ import torch.nn as nn
 from torch.nn import TransformerEncoder, TransformerEncoderLayer
 
 class EvoEncoder(nn.Module):
-    def __init__(self, d_model=512, nhead=8, num_layers=6, dim_feedforward=1024, dropout=0.1):
+    def __init__(self, vocab_size=30522, d_model=512, nhead=8, num_layers=6, dim_feedforward=1024, dropout=0.1):
         super(EvoEncoder, self).__init__()
-        self.positional_encoding = nn.Parameter(torch.zeros(1, 512, d_model))  # Assuming max seq length = 512
+        self.embedding = nn.Embedding(vocab_size, d_model)
+        self.memory_token = nn.Parameter(torch.zeros(1, 1, d_model))
+        self.positional_encoding = nn.Parameter(torch.zeros(1, 512, d_model))
 
         encoder_layer = TransformerEncoderLayer(d_model=d_model, nhead=nhead,
                                                 dim_feedforward=dim_feedforward, dropout=dropout)
         self.transformer = TransformerEncoder(encoder_layer, num_layers=num_layers)
         self.norm = nn.LayerNorm(d_model)
-        self.memory_proj = nn.Linear(d_model, d_model)
 
-    def forward(self, x):
+    def forward(self, input_ids):
+        x = self.embedding(input_ids)
+        bsz = x.size(0)
+
+        # Add memory token
+        mem_token = self.memory_token.expand(bsz, -1, -1)  # [B, 1, D]
+        x = torch.cat([mem_token, x], dim=1)
+
         x = x + self.positional_encoding[:, :x.size(1), :]
         x = self.transformer(x)
         x = self.norm(x)
-        memory_output = self.memory_proj(x[:, 0])  # Use first token
-        return memory_output
+        return x[:, 0]  # return memory token output
 
 class EvoTransformer(nn.Module):
     def __init__(self, vocab_size=30522, d_model=512, nhead=8, num_layers=6,
                  dim_feedforward=1024, dropout=0.1):
         super(EvoTransformer, self).__init__()
-        self.embedding = nn.Embedding(vocab_size, d_model)
-        self.encoder = EvoEncoder(d_model, nhead, num_layers, dim_feedforward, dropout)
+        self.encoder = EvoEncoder(vocab_size, d_model, nhead, num_layers, dim_feedforward, dropout)
         self.classifier = nn.Linear(d_model, 1)
 
     def forward(self, input_ids):
-        x = self.embedding(input_ids)
-        memory_output = self.encoder(x)
-        return memory_output
+        x = self.encoder(input_ids)
+        return x  # shape: [batch, d_model]