File size: 8,469 Bytes
2b4f227 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 | """
ByteEncoder: byte-level transformer with cross-message statistics injection.
Input: (B, N, L) — B formats, N messages each, L bytes (padded to max_len)
Output: (B, N, L, D) per-byte representations, plus the last-layer cross-msg variance
The cross-message statistics injection is the key inductive bias:
after each self-attention block we compute mean / max / var across the N
messages in the same format and concatenate them to each token, then project
back to D. This makes per-offset correlations across messages directly
learnable with no extra overhead.
"""
from __future__ import annotations
import math
import torch
import torch.nn as nn
from einops import rearrange
# ── Special token indices ────────────────────────────────────────────────────
PAD_IDX = 256
BOS_IDX = 257
EOS_IDX = 258
VOCAB = 259 # 0-255 byte values + PAD + BOS + EOS
class CrossMsgTransformerLayer(nn.Module):
"""
One transformer layer whose post-attention step injects cross-message stats.
Sub-block order (pre-LayerNorm throughout for training stability):
1. LN → self-attention → residual
2. LN → cross-msg stats projection → residual (skipped if disable_cross_msg)
3. LN → FFN → residual
"""
def __init__(self, d_model: int = 128, n_heads: int = 4,
d_ff: int = 512, dropout: float = 0.1,
disable_cross_msg: bool = False) -> None:
super().__init__()
self.d_model = d_model
self.disable_cross_msg = disable_cross_msg
# — Self-attention —
self.norm1 = nn.LayerNorm(d_model)
self.self_attn = nn.MultiheadAttention(
d_model, n_heads, dropout=dropout, batch_first=True)
self.drop1 = nn.Dropout(dropout)
# — Cross-message statistics injection —
self.norm2 = nn.LayerNorm(d_model)
self.cross_msg_proj = nn.Linear(d_model * 4, d_model)
self.drop2 = nn.Dropout(dropout)
# — FFN —
self.norm3 = nn.LayerNorm(d_model)
self.ffn = nn.Sequential(
nn.Linear(d_model, d_ff),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(d_ff, d_model),
nn.Dropout(dropout),
)
def forward(
self,
h: torch.Tensor, # (B, N, L, D)
padding_mask: torch.Tensor | None, # (B, N, L) bool, True=pad
) -> tuple[torch.Tensor, torch.Tensor]:
"""Return updated h and the cross-msg variance (for boundary features)."""
B, N, L, D = h.shape
# ── 1. Self-attention ──────────────────────────────────────────────
h_norm = self.norm1(h)
h_flat = rearrange(h_norm, "b n l d -> (b n) l d")
attn_mask_flat: torch.Tensor | None = None
if padding_mask is not None:
# MultiheadAttention key_padding_mask: (B, L), True = ignore
attn_mask_flat = rearrange(padding_mask, "b n l -> (b n) l")
attn_out, _ = self.self_attn(
h_flat, h_flat, h_flat,
key_padding_mask=attn_mask_flat,
)
h = h + self.drop1(rearrange(attn_out, "(b n) l d -> b n l d", b=B, n=N))
# ── 2. Cross-message statistics injection ─────────────────────────
h_norm = self.norm2(h)
if self.disable_cross_msg:
# Ablation: skip cross-msg stats; pass zeros for mean/max/var slots
zeros = torch.zeros_like(h_norm)
combined = torch.cat([h_norm, zeros, zeros, zeros], dim=-1)
var_m = zeros
else:
# Zero out padding positions before aggregating across messages
if padding_mask is not None:
pad_mask_4d = padding_mask.unsqueeze(-1).float() # (B, N, L, 1)
h_agg = h_norm * (1.0 - pad_mask_4d) # zero padding
else:
h_agg = h_norm
mean_m = h_agg.mean(dim=1, keepdim=True).expand_as(h_norm)
max_m = h_agg.max(dim=1, keepdim=True).values.expand_as(h_norm)
var_m = h_agg.var(dim=1, unbiased=False, keepdim=True).expand_as(h_norm)
combined = torch.cat([h_norm, mean_m, max_m, var_m], dim=-1) # (B,N,L,4D)
h = h + self.drop2(self.cross_msg_proj(combined))
# ── 3. FFN ────────────────────────────────────────────────────────
h_norm = self.norm3(h)
h_flat = rearrange(h_norm, "b n l d -> (b n) l d")
ffn_out = self.ffn(h_flat)
h = h + rearrange(ffn_out, "(b n) l d -> b n l d", b=B, n=N)
# Return variance from this layer (used by boundary head if last layer)
return h, var_m # var_m: (B, N, L, D)
class ByteEncoder(nn.Module):
"""
Byte-level encoder with cross-message attention.
Accepts a batch of (B, N, L) byte index tensors and returns
per-byte representations h of shape (B, N, L, D) together with
the final-layer cross-message variance tensor for use in the
boundary prediction head.
"""
def __init__(
self,
d_model: int = 128,
n_heads: int = 4,
d_ff: int = 512,
n_layers: int = 4,
max_len: int = 512,
dropout: float = 0.1,
disable_cross_msg: bool = False,
) -> None:
super().__init__()
self.d_model = d_model
self.max_len = max_len
self.disable_cross_msg = disable_cross_msg
# Byte + special-token embedding
self.embedding = nn.Embedding(VOCAB, d_model, padding_idx=PAD_IDX)
# Learned positional encoding
self.pos_embedding = nn.Embedding(max_len, d_model)
self.embed_drop = nn.Dropout(dropout)
# Transformer layers with cross-message injection
self.layers = nn.ModuleList([
CrossMsgTransformerLayer(d_model, n_heads, d_ff, dropout,
disable_cross_msg=disable_cross_msg)
for _ in range(n_layers)
])
self.final_norm = nn.LayerNorm(d_model)
self._init_weights()
def _init_weights(self) -> None:
nn.init.normal_(self.embedding.weight, std=0.02)
nn.init.normal_(self.pos_embedding.weight, std=0.02)
for layer in self.layers:
nn.init.xavier_uniform_(layer.self_attn.in_proj_weight)
nn.init.xavier_uniform_(layer.cross_msg_proj.weight)
for ffn_layer in layer.ffn:
if isinstance(ffn_layer, nn.Linear):
nn.init.xavier_uniform_(ffn_layer.weight)
def make_padding_mask(self, x: torch.Tensor) -> torch.Tensor:
"""Return bool mask (B, N, L), True where x == PAD_IDX."""
return x == PAD_IDX
def forward(
self,
x: torch.Tensor, # (B, N, L) long
padding_mask: torch.Tensor | None = None, # (B, N, L) bool
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Returns
-------
h : (B, N, L, D) per-byte representations
cross_var : (B, N, L, D) final-layer cross-message variance
"""
B, N, L = x.shape
assert L <= self.max_len, f"Sequence length {L} > max_len {self.max_len}"
if padding_mask is None:
padding_mask = self.make_padding_mask(x)
# — Embeddings —
pos_ids = torch.arange(L, device=x.device).unsqueeze(0) # (1, L)
# Expand to (B*N, L) for pos lookup, then reshape
tok_emb = self.embedding(x) # (B, N, L, D)
pos_emb = self.pos_embedding(pos_ids) # (1, L, D) → broadcasts to (B,N,L,D)
h = self.embed_drop(tok_emb + pos_emb)
# — Transformer layers —
cross_var = torch.zeros_like(h)
for layer in self.layers:
h, cross_var = layer(h, padding_mask)
h = self.final_norm(h)
# Zero out padding positions in the output
if padding_mask is not None:
h = h * (~padding_mask).unsqueeze(-1).float()
return h, cross_var
|