File size: 8,818 Bytes
42bcacc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# --- START OF FILE ALM.py ---

import torch
import torch.nn as nn
import torch.nn.functional as F
import math

class AttentionLinkedMemory(nn.Module):
    """
    Implements an Attention-Linked Memory Layer.

    This layer retrieves context from a structured memory based on a query.
    It uses a two-level attention mechanism:
    1. Within-Bucket Attention: Summarizes each memory bucket based on relevance
       to the query.
    2. Between-Bucket Attention: Weights the summarized buckets to find the most
       relevant buckets for the query, producing a final aggregated context.

    Args:
        query_dim (int): Dimension of the input query vector.
        memory_dim (int): Dimension of the memory item vectors.
        embed_dim (int): Internal embedding dimension for attention. Keys, Queries,
                         and Values will be projected to this dimension.
        num_heads (int): Number of attention heads for both levels. Must divide embed_dim.
        output_dim (int, optional): Dimension of the final output context vector.
                                   If None, defaults to embed_dim.
        dropout_rate (float, optional): Dropout probability. Defaults to 0.1.
    """
    def __init__(self,
                 query_dim: int,
                 memory_dim: int,
                 embed_dim: int,
                 num_heads: int,
                 output_dim: int = None,
                 dropout_rate: float = 0.1):
        super().__init__()

        if embed_dim % num_heads != 0:
            raise ValueError(f"embed_dim ({embed_dim}) must be divisible by num_heads ({num_heads})")

        self.query_dim = query_dim
        self.memory_dim = memory_dim
        self.embed_dim = embed_dim
        self.output_dim = output_dim if output_dim is not None else embed_dim
        self.num_heads = num_heads
        self.head_dim = embed_dim // num_heads
        self.scale = math.sqrt(self.head_dim)

        # --- Level 1 (Within-Bucket) Projections ---
        self.q_proj_l1 = nn.Linear(query_dim, embed_dim)
        self.k_proj_l1 = nn.Linear(memory_dim, embed_dim)
        self.v_proj_l1 = nn.Linear(memory_dim, embed_dim)
        self.dropout_l1 = nn.Dropout(dropout_rate)
        self.norm_l1_out = nn.LayerNorm(embed_dim)

        # --- Level 2 (Between-Bucket) Projections ---
        self.q_proj_l2 = nn.Linear(query_dim, embed_dim)
        self.k_proj_l2 = nn.Linear(embed_dim, embed_dim)
        self.v_proj_l2 = nn.Linear(embed_dim, embed_dim)
        self.dropout_l2 = nn.Dropout(dropout_rate)

        # --- Final Output Projection ---
        self.out_proj = nn.Linear(embed_dim, self.output_dim)
        self.norm_final = nn.LayerNorm(self.output_dim)

    def forward(self,
                query: torch.Tensor,
                memory_buckets: torch.Tensor,
                memory_mask: torch.Tensor = None):
        """
        Forward pass through the Attention-Linked Memory layer.

        Args:
            query (torch.Tensor): Input query tensor.
                Shape: (batch_size, query_dim)
            memory_buckets (torch.Tensor): Tensor containing memory items organized
                into buckets. Padded for consistent size.
                Shape: (batch_size, num_buckets, max_items_per_bucket, memory_dim)
            memory_mask (torch.Tensor, optional): Boolean mask indicating valid
                memory items (True) vs padding (False).
                Shape: (batch_size, num_buckets, max_items_per_bucket)

        Returns:
            Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
                - aggregated_context (torch.Tensor): The final context vector.
                  Shape: (batch_size, output_dim)
                - bucket_attention_weights_l2 (torch.Tensor): Attention weights for buckets (L2).
                  Shape: (batch_size, num_buckets)
                - item_attention_weights_l1 (torch.Tensor): Attention weights for items within buckets (L1, averaged over heads).
                  Shape: (batch_size, num_buckets, max_items_per_bucket)
        """
        batch_size, num_buckets, max_items, _ = memory_buckets.shape

        # ========================= Level 1: Within-Bucket Attention =========================
        q1 = self.q_proj_l1(query).view(batch_size, self.num_heads, 1, 1, self.head_dim)
        k1 = self.k_proj_l1(memory_buckets)
        v1 = self.v_proj_l1(memory_buckets)

        k1 = k1.view(batch_size, num_buckets, max_items, self.num_heads, self.head_dim).permute(0, 3, 1, 2, 4)
        v1 = v1.view(batch_size, num_buckets, max_items, self.num_heads, self.head_dim).permute(0, 3, 1, 2, 4)

        attn_scores_l1 = torch.einsum('bhqid,bhnkd->bhnqk', q1, k1) / self.scale
        attn_scores_l1 = attn_scores_l1.squeeze(3) # (B, H, N_b, N_i)

        if memory_mask is not None:
            expanded_mask_l1 = memory_mask.unsqueeze(1) # (B, 1, N_b, N_i)
            attn_scores_l1 = attn_scores_l1.masked_fill(expanded_mask_l1 == 0, -1e9)

        attn_weights_l1_raw = F.softmax(attn_scores_l1, dim=-1) # (B, H, N_b, N_i)
        # For returning, average over heads
        item_attention_weights_l1_for_output = attn_weights_l1_raw.mean(dim=1) # (B, N_b, N_i)
        
        attn_weights_l1_dropout = self.dropout_l1(attn_weights_l1_raw)

        bucket_summaries = torch.einsum('bhnk,bhnkd->bhnd', attn_weights_l1_dropout, v1)
        bucket_summaries = bucket_summaries.permute(0, 2, 1, 3).contiguous().view(batch_size, num_buckets, self.embed_dim)
        bucket_summaries = self.norm_l1_out(bucket_summaries)

        # ========================= Level 2: Between-Bucket Attention ========================
        q2 = self.q_proj_l2(query).view(batch_size, self.num_heads, 1, self.head_dim)
        k2 = self.k_proj_l2(bucket_summaries)
        v2 = self.v_proj_l2(bucket_summaries)

        # Corrected typo: self.head__dim -> self.head_dim
        k2 = k2.view(batch_size, num_buckets, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        v2 = v2.view(batch_size, num_buckets, self.num_heads, self.head_dim).permute(0, 2, 1, 3)

        attn_scores_l2 = torch.einsum('bhqd,bhnd->bhqn', q2, k2) / self.scale
        attn_scores_l2 = attn_scores_l2.squeeze(2) # (B, H, N_b)
        
        # Optional: Mask entire buckets if they were all padding in L1 (advanced)
        # For now, assume valid buckets have non-zero summaries
        # Or if memory_mask indicates all items in a bucket are false, that bucket's summary might be ~0
        # and L2 attention should naturally give it low weight.

        attn_weights_l2_raw = F.softmax(attn_scores_l2, dim=-1) # (B, H, N_b)
        bucket_attention_weights_l2_for_output = attn_weights_l2_raw.mean(dim=1) # (B, N_b)

        attn_weights_l2_dropout = self.dropout_l2(attn_weights_l2_raw)

        aggregated_context_heads = torch.einsum('bhn,bhnd->bhd', attn_weights_l2_dropout, v2)
        aggregated_context = aggregated_context_heads.contiguous().view(batch_size, self.embed_dim)

        # ========================= Final Output Projection ========================
        final_output = self.out_proj(aggregated_context)
        final_output = self.norm_final(final_output)

        return final_output, bucket_attention_weights_l2_for_output, item_attention_weights_l1_for_output

# ========================= Example Usage (ALM Standalone) =========================
if __name__ == '__main__':
    # Test the modified ALM layer
    _batch_size = 2
    _query_dim = 32
    _memory_dim = 24
    _embed_dim = 64
    _num_heads = 4
    _output_dim = 32
    _num_buckets = 3
    _max_items_per_bucket = 5

    alm_layer_test = AttentionLinkedMemory(
        query_dim=_query_dim, memory_dim=_memory_dim, embed_dim=_embed_dim,
        num_heads=_num_heads, output_dim=_output_dim
    )

    query_test = torch.randn(_batch_size, _query_dim)
    memory_buckets_test = torch.randn(_batch_size, _num_buckets, _max_items_per_bucket, _memory_dim)
    memory_mask_test = torch.ones(_batch_size, _num_buckets, _max_items_per_bucket, dtype=torch.bool)
    memory_mask_test[:, :, -1] = 0 # Mask last item in each bucket

    agg_ctx, buck_att, item_att = alm_layer_test(query_test, memory_buckets_test, memory_mask_test)

    print("--- ALM Standalone Test ---")
    print(f"Aggregated Context Shape: {agg_ctx.shape}") # Expected: (B, output_dim)
    print(f"Bucket Attention (L2) Shape: {buck_att.shape}") # Expected: (B, num_buckets)
    print(f"Item Attention (L1) Shape: {item_att.shape}") # Expected: (B, num_buckets, max_items_per_bucket)
    print("Item Attention for first batch, first bucket:\n", item_att[0, 0, :])
    print("Sum of item attentions (should be ~1.0 for unmasked items):", item_att[0,0,:-1].sum())
    print("Attention on masked item (should be ~0.0):", item_att[0,0,-1])


# --- END OF FILE ALM.py ---