File size: 6,996 Bytes
924e4e0 |
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 |
"""
Simplified Query Auto-Completion Model
Uses CNN+Transformer for prefix/candidate encoding (IE module)
Optionally uses pretrained ByT5 embeddings
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import T5EncoderModel
class CNNLocalEncoder(nn.Module):
"""Multi-scale CNN for local pattern extraction"""
def __init__(self, embed_dim=128, num_filters=64, filter_sizes=[3, 4, 5]):
super().__init__()
self.convs = nn.ModuleList(
[
nn.Conv1d(embed_dim, num_filters, fs, padding=fs // 2)
for fs in filter_sizes
]
)
self.layer_norm = nn.LayerNorm(num_filters * len(filter_sizes))
self._init_weights()
def _init_weights(self):
for conv in self.convs:
nn.init.kaiming_normal_(conv.weight, mode="fan_out", nonlinearity="relu")
nn.init.zeros_(conv.bias)
def forward(self, x):
x = x.transpose(1, 2)
conv_outs = [F.relu(conv(x)) for conv in self.convs]
pooled = [
(
F.max_pool1d(out, out.size(2)).squeeze(2)
if out.size(2) > 1
else out.squeeze(2)
)
for out in conv_outs
]
out = torch.cat(pooled, dim=1)
return self.layer_norm(out)
class PrefixEncoder(nn.Module):
"""CNN + Transformer encoder for prefix"""
def __init__(self, embed_dim=128, num_filters=64, num_heads=4, num_layers=2):
super().__init__()
self.cnn = CNNLocalEncoder(embed_dim, num_filters)
cnn_out_dim = num_filters * 3
self.transformer = nn.TransformerEncoder(
nn.TransformerEncoderLayer(
d_model=cnn_out_dim,
nhead=num_heads,
dim_feedforward=cnn_out_dim * 4,
dropout=0.1,
batch_first=True,
activation="gelu",
layer_norm_eps=1e-6,
norm_first=True,
),
num_layers=num_layers,
)
self.proj = nn.Linear(cnn_out_dim, embed_dim)
self.layer_norm = nn.LayerNorm(embed_dim)
self._init_weights()
def _init_weights(self):
nn.init.xavier_uniform_(self.proj.weight, gain=0.5)
nn.init.zeros_(self.proj.bias)
def forward(self, prefix_embed):
cnn_out = self.cnn(prefix_embed).unsqueeze(1)
transformer_out = self.transformer(cnn_out).squeeze(1)
return self.layer_norm(self.proj(transformer_out))
class CandidateEncoder(nn.Module):
"""Transformer encoder for candidate (no CNN)"""
def __init__(self, embed_dim=128, num_heads=4, num_layers=2):
super().__init__()
self.transformer = nn.TransformerEncoder(
nn.TransformerEncoderLayer(
d_model=embed_dim,
nhead=num_heads,
dim_feedforward=embed_dim * 4,
dropout=0.1,
batch_first=True,
activation="gelu",
layer_norm_eps=1e-6,
norm_first=True,
),
num_layers=num_layers,
)
self.layer_norm = nn.LayerNorm(embed_dim)
def forward(self, candidate_embed):
transformer_out = self.transformer(candidate_embed)
pooled = torch.max(transformer_out, dim=1)[0]
return self.layer_norm(pooled)
class QueryCompletionModel(nn.Module):
"""Query auto-completion: CNN+Transformer for prefix, Transformer for candidate"""
def __init__(
self,
vocab_size=10000,
embed_dim=128,
num_filters=64,
num_heads=4,
num_transformer_layers=2,
use_pretrained_embeddings=False,
pretrained_model_name="google/byt5-small",
):
super().__init__()
self.use_pretrained_embeddings = use_pretrained_embeddings
if use_pretrained_embeddings:
# Load pretrained ByT5 and use its embeddings
print(f"Loading pretrained embeddings from {pretrained_model_name}...")
byt5_model = T5EncoderModel.from_pretrained(pretrained_model_name)
pretrained_embed_dim = byt5_model.config.d_model
# Share the pretrained embedding for both prefix and candidate
self.shared_embedding = byt5_model.shared
self.shared_embedding.requires_grad_(True) # Fine-tune embeddings
# Project to target embed_dim if different
if pretrained_embed_dim != embed_dim:
self.embed_proj = nn.Linear(pretrained_embed_dim, embed_dim)
nn.init.xavier_uniform_(self.embed_proj.weight, gain=0.5)
else:
self.embed_proj = nn.Identity()
print(
f" ✓ Using pretrained embeddings: {pretrained_embed_dim}D → {embed_dim}D"
)
else:
# Use separate learned embeddings (original behavior)
self.prefix_embedding = nn.Embedding(vocab_size, embed_dim)
self.candidate_embedding = nn.Embedding(vocab_size, embed_dim)
self._init_embeddings()
self.prefix_encoder = PrefixEncoder(
embed_dim, num_filters, num_heads, num_transformer_layers
)
self.candidate_encoder = CandidateEncoder(
embed_dim, num_heads, num_transformer_layers
)
self.match_predictor = nn.Sequential(
nn.LayerNorm(embed_dim * 2),
nn.Linear(embed_dim * 2, embed_dim * 2),
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(embed_dim * 2, embed_dim),
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(embed_dim, 1),
)
self._init_predictor()
def _init_embeddings(self):
if not self.use_pretrained_embeddings:
nn.init.normal_(self.prefix_embedding.weight, std=0.02)
nn.init.normal_(self.candidate_embedding.weight, std=0.02)
def _init_predictor(self):
for module in self.match_predictor:
if isinstance(module, nn.Linear):
nn.init.xavier_uniform_(module.weight, gain=0.5)
nn.init.zeros_(module.bias)
def forward(self, prefix_ids, candidate_ids):
if self.use_pretrained_embeddings:
# Use shared pretrained embeddings for both
prefix_embed = self.embed_proj(self.shared_embedding(prefix_ids))
candidate_embed = self.embed_proj(self.shared_embedding(candidate_ids))
else:
# Use separate learned embeddings
prefix_embed = self.prefix_embedding(prefix_ids)
candidate_embed = self.candidate_embedding(candidate_ids)
prefix_intention = self.prefix_encoder(prefix_embed)
candidate_intention = self.candidate_encoder(candidate_embed)
combined = torch.cat([prefix_intention, candidate_intention], dim=-1)
logits = self.match_predictor(combined)
return torch.sigmoid(logits)
|