Spaces:
Sleeping
Sleeping
File size: 12,036 Bytes
543e824 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 |
# models/text_models.py
import logging
from typing import Dict, List, Optional, Tuple, Union, Any
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
from sentence_transformers import SentenceTransformer
class TextModelManager:
def __init__(self, token_manager=None, cache_manager=None, metrics_calculator=None):
"""Initialize the TextModelManager with optional utilities."""
self.logger = logging.getLogger(__name__)
self.token_manager = token_manager
self.cache_manager = cache_manager
self.metrics_calculator = metrics_calculator
# Model instances
self.embedding_model = None
self.understanding_model = None
self.embedding_tokenizer = None
self.understanding_tokenizer = None
# Model names
self.embedding_model_name = "sentence-transformers/all-MiniLM-L6-v2"
self.understanding_model_name = "microsoft/deberta-v3-small"
# Track initialization state
self.initialized = {
"embedding": False,
"understanding": False
}
def initialize_embedding_model(self):
"""Initialize the embedding model for topic-document relevance."""
if self.initialized["embedding"]:
return
try:
# Register with token manager if available
if self.token_manager:
self.token_manager.register_model(
self.embedding_model_name, "embedding")
# Load model
self.logger.info(f"Loading embedding model: {self.embedding_model_name}")
self.embedding_model = SentenceTransformer(self.embedding_model_name)
# Also load tokenizer separately for token counting
self.embedding_tokenizer = AutoTokenizer.from_pretrained(
self.embedding_model_name)
self.initialized["embedding"] = True
self.logger.info("Embedding model initialized successfully")
except Exception as e:
self.logger.error(f"Failed to initialize embedding model: {e}")
raise
def initialize_understanding_model(self):
"""Initialize the understanding model for document analysis."""
if self.initialized["understanding"]:
return
try:
model_name = self.understanding_model_name
# Register with token manager if available
if self.token_manager:
self.token_manager.register_model(
self.understanding_model_name, "understanding")
# Load model and tokenizer
self.logger.info(f"Loading understanding model: {self.understanding_model_name}")
self.understanding_tokenizer = AutoTokenizer.from_pretrained(
self.understanding_model_name)
self.understanding_model = AutoModel.from_pretrained(
self.understanding_model_name)
self.initialized["understanding"] = True
self.logger.info("Understanding model initialized successfully")
except Exception as e:
self.logger.error(f"Failed to initialize understanding model: {e}")
# Try fallback model if primary fails
try:
fallback_model = "distilbert-base-uncased"
self.logger.info(f"Trying fallback model: {fallback_model}")
self.understanding_tokenizer = AutoTokenizer.from_pretrained(fallback_model)
self.understanding_model = AutoModel.from_pretrained(fallback_model)
self.understanding_model_name = fallback_model
self.initialized["understanding"] = True
self.logger.info("Fallback understanding model initialized successfully")
except Exception as fallback_error:
self.logger.error(f"Failed to initialize fallback model: {fallback_error}")
raise
def get_embeddings(self, texts: Union[str, List[str]],
agent_name: str = "text_analysis") -> np.ndarray:
"""
Generate embeddings for the given texts.
Optimized with caching and token tracking.
"""
# Initialize model if needed
if not self.initialized["embedding"]:
self.initialize_embedding_model()
# Handle single text
if isinstance(texts, str):
texts = [texts]
results = []
cache_hits = 0
tokens_used = 0
for text in texts:
# Check cache if available
if self.cache_manager:
cache_hit, cached_embedding = self.cache_manager.get(
text, namespace="embeddings")
if cache_hit:
results.append(cached_embedding)
cache_hits += 1
continue
# Request token budget if available
if self.token_manager:
approved, reason = self.token_manager.request_tokens(
agent_name, "embedding", text, self.embedding_model_name)
if not approved:
self.logger.warning(f"Token budget exceeded: {reason}")
# Return zeros as fallback
results.append(np.zeros(384)) # Default embedding dimension
continue
# Generate embedding
with torch.no_grad():
embedding = self.embedding_model.encode(text)
# Store in cache if available
if self.cache_manager:
self.cache_manager.put(text, embedding, namespace="embeddings",
embedding=embedding)
# Log token usage if available
if self.token_manager:
token_count = len(self.embedding_tokenizer.encode(text))
self.token_manager.log_usage(
agent_name, "embedding", token_count, self.embedding_model_name)
tokens_used += token_count
# Log energy usage if metrics calculator is available
if self.metrics_calculator:
energy_usage = self.token_manager.calculate_energy_usage(
token_count, self.embedding_model_name)
self.metrics_calculator.log_energy_usage(
energy_usage, self.embedding_model_name,
agent_name, "embedding")
results.append(embedding)
# Update cache metrics if available
if self.cache_manager and self.metrics_calculator:
# Estimate energy saved through cache hits
if cache_hits > 0 and tokens_used > 0:
avg_tokens_per_text = tokens_used / (len(texts) - cache_hits)
estimated_tokens_saved = avg_tokens_per_text * cache_hits
if self.token_manager:
energy_saved = self.token_manager.calculate_energy_usage(
estimated_tokens_saved, self.embedding_model_name)
self.metrics_calculator.update_cache_metrics(
cache_hits, len(texts) - cache_hits, energy_saved)
# Return single embedding or list based on input
if len(results) == 1 and isinstance(texts, str):
return results[0]
return np.array(results)
def compute_similarity(self, topic: str, documents: List[str],
agent_name: str = "text_analysis") -> List[float]:
"""
Compute semantic similarity between topic and documents.
Returns list of similarity scores (0-1).
"""
# Get embeddings
topic_embedding = self.get_embeddings(topic, agent_name)
doc_embeddings = self.get_embeddings(documents, agent_name)
# Compute similarities
similarities = []
for doc_embedding in doc_embeddings:
# Cosine similarity
similarity = np.dot(topic_embedding, doc_embedding) / (
np.linalg.norm(topic_embedding) * np.linalg.norm(doc_embedding))
similarities.append(float(similarity))
return similarities
def analyze_document(self, document: str, query: str = None,
agent_name: str = "text_analysis") -> Dict[str, Any]:
"""
Analyze document content using the understanding model.
If query is provided, focuses analysis on that query.
"""
# Initialize model if needed
if not self.initialized["understanding"]:
self.initialize_understanding_model()
# Check cache if available
cache_key = f"{document}::{query}" if query else document
if self.cache_manager:
cache_hit, cached_result = self.cache_manager.get(
cache_key, namespace="document_analysis")
if cache_hit:
# Update metrics if available
if self.metrics_calculator:
self.metrics_calculator.update_cache_metrics(1, 0, 0.005) # Estimated energy saving
return cached_result
# Prepare input
if query:
input_text = f"Query: {query}\nDocument: {document}"
else:
input_text = document
# Request token budget if available
if self.token_manager:
approved, reason = self.token_manager.request_tokens(
agent_name, "understanding", input_text, self.understanding_model_name)
if not approved:
self.logger.warning(f"Token budget exceeded: {reason}")
return {"error": reason, "summary": "Token budget exceeded"}
# Tokenize
inputs = self.understanding_tokenizer(
input_text, return_tensors="pt", truncation=True, max_length=512)
# Get model outputs
with torch.no_grad():
outputs = self.understanding_model(**inputs)
# Process outputs - using last hidden state for analysis
last_hidden_state = outputs.last_hidden_state
# Extract key information (simplified for demonstration)
# In a real implementation, we'd use more sophisticated analysis
mean_embedding = torch.mean(last_hidden_state, dim=1).squeeze().numpy()
# Create analysis result
result = {
"document_length": len(document.split()),
"embedding_norm": float(np.linalg.norm(mean_embedding)),
"content_vector": mean_embedding.tolist()[:10] # First 10 dims as sample
}
# Log token usage if available
if self.token_manager:
token_count = len(inputs.input_ids[0])
self.token_manager.log_usage(
agent_name, "understanding", token_count, self.understanding_model_name)
# Log energy usage if metrics calculator is available
if self.metrics_calculator:
energy_usage = self.token_manager.calculate_energy_usage(
token_count, self.understanding_model_name)
self.metrics_calculator.log_energy_usage(
energy_usage, self.understanding_model_name,
agent_name, "understanding")
# Store in cache if available
if self.cache_manager:
self.cache_manager.put(cache_key, result, namespace="document_analysis")
return result
|