File size: 7,134 Bytes
3451ca0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
# modules.py
import torch
import torch.nn as nn
import torch.nn.functional as F
import config

class CrossAttentionDelta(nn.Module):
    """

    Enhanced version of CrossAttentionDelta that computes the update delta (Δ) using cross-attention.

    Improvements:

    1. Pre-norm architecture (layer norm before attention)

    2. More sophisticated attention patterns

    3. Ability to incorporate reasoning trace

    """
    def __init__(self, hidden_dim, num_heads=8, dropout=0.1):
        super().__init__()
        self.hidden_dim = hidden_dim
        self.num_heads = num_heads
        
        # Pre-norm layer normalization (applied before attention)
        self.pre_norm = nn.LayerNorm(hidden_dim)
        
        # Cross-attention mechanism
        self.cross_attn = nn.MultiheadAttention(
            embed_dim=hidden_dim,
            num_heads=num_heads,
            dropout=dropout,
            batch_first=True
        )
        
        # Post-attention layer normalization
        self.post_norm = nn.LayerNorm(hidden_dim)
        
        # Trace integration module (to incorporate reasoning trace T)
        self.trace_integration = nn.Sequential(
            nn.Linear(hidden_dim * 2, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, hidden_dim)
        )
        
        # Enhanced MLP for delta computation
        self.delta_mlp = nn.Sequential(
            nn.Linear(hidden_dim * 2, hidden_dim * 4),  # Larger intermediate expansion
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim * 4, hidden_dim * 2),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim * 2, hidden_dim)
        )
        
        # Final layer normalization
        self.final_norm = nn.LayerNorm(hidden_dim)

    def forward(self, h0, reasoning_trace=None):
        """

        Args:

            h0 (torch.Tensor): Initial hidden states (batch_size, seq_len, hidden_dim).

            reasoning_trace (tuple of torch.Tensor, optional): Reasoning trace from base model.

                Each tensor has shape (batch_size, seq_len, hidden_dim).



        Returns:

            delta (torch.Tensor): The computed update delta (batch_size, seq_len, hidden_dim).

        """
        batch_size, seq_len, _ = h0.shape
        
        # --- Pre-norm Architecture ---
        # Apply layer normalization before attention (pre-norm)
        h0_norm = self.pre_norm(h0)
        
        # --- Enhanced Cross-Attention ---
        # Get attention weights to visualize attention patterns
        attn_output, attn_weights = self.cross_attn(
            query=h0_norm, 
            key=h0_norm, 
            value=h0_norm, 
            need_weights=True
        )
        
        # Residual connection and post-norm
        c = self.post_norm(h0 + attn_output)
        
        # --- Reasoning Trace Integration (if provided) ---
        if reasoning_trace is not None and len(reasoning_trace) > 0:
            # Use the last layer from the reasoning trace (most semantic)
            last_layer = reasoning_trace[-1]
            
            # Integrate the reasoning trace with the current context
            trace_info = self.trace_integration(
                torch.cat([c, last_layer], dim=-1)
            )
            
            # Add the trace information to the context
            c = c + trace_info
        
        # --- Enhanced MLP for Delta ---
        # Concatenate original h0 with context c
        mlp_input = torch.cat((h0, c), dim=-1)
        
        # Compute delta through enhanced MLP
        delta = self.delta_mlp(mlp_input)
        
        # Apply final normalization
        delta = self.final_norm(delta)
        
        return delta, attn_weights

class GatingMechanism(nn.Module):
    """

    Gating mechanism to selectively apply updates.

    Learns when to apply the delta update based on the hidden state and delta.

    """
    def __init__(self, hidden_dim, dropout=0.1):
        super().__init__()
        self.gate_network = nn.Sequential(
            nn.Linear(hidden_dim * 2, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, 1),
            nn.Sigmoid()  # Output between 0 and 1
        )
    
    def forward(self, h0, delta):
        """

        Args:

            h0 (torch.Tensor): Initial hidden states (batch_size, seq_len, hidden_dim).

            delta (torch.Tensor): Computed delta (batch_size, seq_len, hidden_dim).

            

        Returns:

            gate (torch.Tensor): Gate values between 0 and 1 (batch_size, seq_len, 1).

        """
        # Concatenate h0 and delta
        gate_input = torch.cat([h0, delta], dim=-1)
        
        # Compute gate values
        gate = self.gate_network(gate_input)
        
        return gate

class EnhancedQAHead(nn.Module):
    """

    Enhanced Question Answering head with deeper architecture and bilinear scoring.

    """
    def __init__(self, hidden_dim, dropout=0.1):
        super().__init__()
        
        # Deeper representation before prediction
        self.start_transform = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, hidden_dim)
        )
        
        self.end_transform = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, hidden_dim)
        )
        
        # Bilinear layer for start position scoring
        self.start_bilinear = nn.Bilinear(hidden_dim, hidden_dim, 1)
        
        # Bilinear layer for end position scoring
        self.end_bilinear = nn.Bilinear(hidden_dim, hidden_dim, 1)
        
        # Global representation for bilinear scoring
        self.global_rep = nn.Parameter(torch.randn(hidden_dim))

    def forward(self, hidden_states):
        """

        Args:

            hidden_states (torch.Tensor): Hidden states (batch_size, seq_len, hidden_dim).



        Returns:

            dict: Dictionary with start_logits and end_logits.

        """
        batch_size, seq_len, hidden_dim = hidden_states.shape
        
        # Transform hidden states
        start_rep = self.start_transform(hidden_states)
        end_rep = self.end_transform(hidden_states)
        
        # Expand global representation for batch processing
        global_rep = self.global_rep.expand(batch_size, seq_len, -1)
        
        # Compute start and end logits using bilinear scoring
        start_logits = self.start_bilinear(start_rep, global_rep).squeeze(-1)
        end_logits = self.end_bilinear(end_rep, global_rep).squeeze(-1)
        
        return {"start_logits": start_logits, "end_logits": end_logits}