Spaces:
Running
Running
chore: update something
Browse files
lightweight_embeddings/service.py
CHANGED
@@ -4,6 +4,7 @@ from __future__ import annotations
|
|
4 |
|
5 |
import asyncio
|
6 |
import logging
|
|
|
7 |
from enum import Enum
|
8 |
from typing import List, Union, Dict, Optional, NamedTuple, Any
|
9 |
from dataclasses import dataclass
|
@@ -27,7 +28,6 @@ class TextModelType(str, Enum):
|
|
27 |
"""
|
28 |
Enumeration of supported text models.
|
29 |
"""
|
30 |
-
|
31 |
MULTILINGUAL_E5_SMALL = "multilingual-e5-small"
|
32 |
MULTILINGUAL_E5_BASE = "multilingual-e5-base"
|
33 |
MULTILINGUAL_E5_LARGE = "multilingual-e5-large"
|
@@ -42,7 +42,6 @@ class ImageModelType(str, Enum):
|
|
42 |
"""
|
43 |
Enumeration of supported image models.
|
44 |
"""
|
45 |
-
|
46 |
SIGLIP_BASE_PATCH16_256_MULTILINGUAL = "siglip-base-patch16-256-multilingual"
|
47 |
|
48 |
|
@@ -50,7 +49,6 @@ class MaxModelLength(str, Enum):
|
|
50 |
"""
|
51 |
Enumeration of maximum token lengths for supported text models.
|
52 |
"""
|
53 |
-
|
54 |
MULTILINGUAL_E5_SMALL = 512
|
55 |
MULTILINGUAL_E5_BASE = 512
|
56 |
MULTILINGUAL_E5_LARGE = 512
|
@@ -65,7 +63,6 @@ class ModelInfo(NamedTuple):
|
|
65 |
"""
|
66 |
Container mapping a model type to its model identifier and optional ONNX file.
|
67 |
"""
|
68 |
-
|
69 |
model_id: str
|
70 |
onnx_file: Optional[str] = None
|
71 |
|
@@ -75,11 +72,8 @@ class ModelConfig:
|
|
75 |
"""
|
76 |
Configuration for text and image models.
|
77 |
"""
|
78 |
-
|
79 |
text_model_type: TextModelType = TextModelType.MULTILINGUAL_E5_SMALL
|
80 |
-
image_model_type: ImageModelType =
|
81 |
-
ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL
|
82 |
-
)
|
83 |
logit_scale: float = 4.60517 # Example scale used in cross-modal similarity
|
84 |
|
85 |
@property
|
@@ -140,7 +134,6 @@ class ModelKind(str, Enum):
|
|
140 |
"""
|
141 |
Indicates the type of model: text or image.
|
142 |
"""
|
143 |
-
|
144 |
TEXT = "text"
|
145 |
IMAGE = "image"
|
146 |
|
@@ -184,6 +177,11 @@ class EmbeddingsService:
|
|
184 |
self.image_models: Dict[ImageModelType, AutoModel] = {}
|
185 |
self.image_processors: Dict[ImageModelType, AutoProcessor] = {}
|
186 |
|
|
|
|
|
|
|
|
|
|
|
187 |
# Create a persistent asynchronous HTTP client.
|
188 |
self.async_http_client = httpx.AsyncClient(timeout=10)
|
189 |
|
@@ -220,17 +218,11 @@ class EmbeddingsService:
|
|
220 |
# Set maximum sequence length based on configuration.
|
221 |
max_length = int(MaxModelLength[t_model_type.name].value)
|
222 |
self.text_models[t_model_type].max_seq_length = max_length
|
223 |
-
logger.info(
|
224 |
-
"Set max_seq_length=%d for text model: %s",
|
225 |
-
max_length,
|
226 |
-
info.model_id,
|
227 |
-
)
|
228 |
|
229 |
# Preload image models.
|
230 |
for i_model_type in ImageModelType:
|
231 |
-
model_id = ModelConfig(
|
232 |
-
image_model_type=i_model_type
|
233 |
-
).image_model_info.model_id
|
234 |
logger.info("Loading image model: %s", model_id)
|
235 |
model = AutoModel.from_pretrained(model_id).to(self.device)
|
236 |
model.eval() # Set the model to evaluation mode.
|
@@ -257,9 +249,7 @@ class EmbeddingsService:
|
|
257 |
raise ValueError("Text input cannot be empty.")
|
258 |
return [input_text]
|
259 |
|
260 |
-
if not isinstance(input_text, list) or not all(
|
261 |
-
isinstance(x, str) for x in input_text
|
262 |
-
):
|
263 |
raise ValueError("Text input must be a string or a list of strings.")
|
264 |
|
265 |
if len(input_text) == 0:
|
@@ -280,9 +270,7 @@ class EmbeddingsService:
|
|
280 |
raise ValueError("Image input cannot be empty.")
|
281 |
return [input_images]
|
282 |
|
283 |
-
if not isinstance(input_images, list) or not all(
|
284 |
-
isinstance(x, str) for x in input_images
|
285 |
-
):
|
286 |
raise ValueError("Image input must be a string or a list of strings.")
|
287 |
|
288 |
if len(input_images) == 0:
|
@@ -305,17 +293,15 @@ class EmbeddingsService:
|
|
305 |
try:
|
306 |
# Attempt to get the tokenizer from the first module of the SentenceTransformer.
|
307 |
module = model._first_module()
|
308 |
-
if not hasattr(module,
|
309 |
return text
|
310 |
tokenizer = module.tokenizer
|
311 |
# Tokenize without truncation.
|
312 |
encoded = tokenizer(text, add_special_tokens=True, truncation=False)
|
313 |
max_length = model.max_seq_length
|
314 |
-
if len(encoded[
|
315 |
-
truncated_ids = encoded[
|
316 |
-
truncated_text = tokenizer.decode(
|
317 |
-
truncated_ids, skip_special_tokens=True
|
318 |
-
)
|
319 |
return truncated_text
|
320 |
except Exception as e:
|
321 |
logger.warning("Error during text truncation: %s", str(e))
|
@@ -367,9 +353,7 @@ class EmbeddingsService:
|
|
367 |
processed_data = processor(images=img, return_tensors="pt").to(self.device)
|
368 |
return processed_data
|
369 |
|
370 |
-
def _generate_text_embeddings(
|
371 |
-
self, model_id: TextModelType, texts: List[str]
|
372 |
-
) -> np.ndarray:
|
373 |
"""
|
374 |
Generate text embeddings using the SentenceTransformer model.
|
375 |
Single-text requests are cached using an LRU cache.
|
@@ -385,26 +369,25 @@ class EmbeddingsService:
|
|
385 |
RuntimeError: If text embedding generation fails.
|
386 |
"""
|
387 |
try:
|
388 |
-
if len(texts) == 1:
|
389 |
-
single_text = texts[0]
|
390 |
-
key = md5(f"{model_id}:{single_text}".encode("utf-8")).hexdigest()[:8]
|
391 |
-
if key in self.lru_cache:
|
392 |
-
return self.lru_cache[key]
|
393 |
-
model = self.text_models[model_id]
|
394 |
-
emb = model.encode([single_text])
|
395 |
-
self.lru_cache[key] = emb
|
396 |
-
return emb
|
397 |
-
|
398 |
model = self.text_models[model_id]
|
399 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
400 |
except Exception as e:
|
401 |
raise RuntimeError(
|
402 |
f"Error generating text embeddings with model '{model_id}': {e}"
|
403 |
) from e
|
404 |
|
405 |
-
async def _async_generate_image_embeddings(
|
406 |
-
self, model_id: ImageModelType, images: List[str]
|
407 |
-
) -> np.ndarray:
|
408 |
"""
|
409 |
Asynchronously generate image embeddings.
|
410 |
|
@@ -428,15 +411,11 @@ class EmbeddingsService:
|
|
428 |
)
|
429 |
# Assume all processed outputs have the same keys.
|
430 |
keys = processed_tensors[0].keys()
|
431 |
-
combined = {
|
432 |
-
k: torch.cat([pt[k] for pt in processed_tensors], dim=0) for k in keys
|
433 |
-
}
|
434 |
|
435 |
def infer():
|
436 |
with torch.no_grad():
|
437 |
-
embeddings = self.image_models[model_id].get_image_features(
|
438 |
-
**combined
|
439 |
-
)
|
440 |
return embeddings.cpu().numpy()
|
441 |
|
442 |
return await asyncio.to_thread(infer)
|
@@ -445,9 +424,7 @@ class EmbeddingsService:
|
|
445 |
f"Error generating image embeddings with model '{model_id}': {e}"
|
446 |
) from e
|
447 |
|
448 |
-
async def generate_embeddings(
|
449 |
-
self, model: str, inputs: Union[str, List[str]]
|
450 |
-
) -> np.ndarray:
|
451 |
"""
|
452 |
Asynchronously generate embeddings for text or image inputs based on model type.
|
453 |
|
@@ -463,26 +440,19 @@ class EmbeddingsService:
|
|
463 |
text_model_enum = TextModelType(model)
|
464 |
text_list = self._validate_text_list(inputs)
|
465 |
model_instance = self.text_models[text_model_enum]
|
466 |
-
|
467 |
-
|
468 |
-
|
469 |
-
|
470 |
return await asyncio.to_thread(
|
471 |
self._generate_text_embeddings, text_model_enum, truncated_texts
|
472 |
)
|
473 |
elif modality == ModelKind.IMAGE:
|
474 |
image_model_enum = ImageModelType(model)
|
475 |
image_list = self._validate_image_list(inputs)
|
476 |
-
return await self._async_generate_image_embeddings(
|
477 |
-
image_model_enum, image_list
|
478 |
-
)
|
479 |
|
480 |
-
async def rank(
|
481 |
-
self,
|
482 |
-
model: str,
|
483 |
-
queries: Union[str, List[str]],
|
484 |
-
candidates: Union[str, List[str]],
|
485 |
-
) -> Dict[str, Any]:
|
486 |
"""
|
487 |
Asynchronously rank candidate texts/images against the provided queries.
|
488 |
Embeddings for queries and candidates are generated concurrently.
|
@@ -503,12 +473,8 @@ class EmbeddingsService:
|
|
503 |
|
504 |
# Concurrently generate embeddings.
|
505 |
query_task = asyncio.create_task(self.generate_embeddings(model, queries))
|
506 |
-
candidate_task = asyncio.create_task(
|
507 |
-
|
508 |
-
)
|
509 |
-
query_embeds, candidate_embeds = await asyncio.gather(
|
510 |
-
query_task, candidate_task
|
511 |
-
)
|
512 |
|
513 |
# Compute cosine similarity.
|
514 |
sim_matrix = self.cosine_similarity(query_embeds, candidate_embeds)
|
|
|
4 |
|
5 |
import asyncio
|
6 |
import logging
|
7 |
+
import threading
|
8 |
from enum import Enum
|
9 |
from typing import List, Union, Dict, Optional, NamedTuple, Any
|
10 |
from dataclasses import dataclass
|
|
|
28 |
"""
|
29 |
Enumeration of supported text models.
|
30 |
"""
|
|
|
31 |
MULTILINGUAL_E5_SMALL = "multilingual-e5-small"
|
32 |
MULTILINGUAL_E5_BASE = "multilingual-e5-base"
|
33 |
MULTILINGUAL_E5_LARGE = "multilingual-e5-large"
|
|
|
42 |
"""
|
43 |
Enumeration of supported image models.
|
44 |
"""
|
|
|
45 |
SIGLIP_BASE_PATCH16_256_MULTILINGUAL = "siglip-base-patch16-256-multilingual"
|
46 |
|
47 |
|
|
|
49 |
"""
|
50 |
Enumeration of maximum token lengths for supported text models.
|
51 |
"""
|
|
|
52 |
MULTILINGUAL_E5_SMALL = 512
|
53 |
MULTILINGUAL_E5_BASE = 512
|
54 |
MULTILINGUAL_E5_LARGE = 512
|
|
|
63 |
"""
|
64 |
Container mapping a model type to its model identifier and optional ONNX file.
|
65 |
"""
|
|
|
66 |
model_id: str
|
67 |
onnx_file: Optional[str] = None
|
68 |
|
|
|
72 |
"""
|
73 |
Configuration for text and image models.
|
74 |
"""
|
|
|
75 |
text_model_type: TextModelType = TextModelType.MULTILINGUAL_E5_SMALL
|
76 |
+
image_model_type: ImageModelType = ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL
|
|
|
|
|
77 |
logit_scale: float = 4.60517 # Example scale used in cross-modal similarity
|
78 |
|
79 |
@property
|
|
|
134 |
"""
|
135 |
Indicates the type of model: text or image.
|
136 |
"""
|
|
|
137 |
TEXT = "text"
|
138 |
IMAGE = "image"
|
139 |
|
|
|
177 |
self.image_models: Dict[ImageModelType, AutoModel] = {}
|
178 |
self.image_processors: Dict[ImageModelType, AutoProcessor] = {}
|
179 |
|
180 |
+
# Create reentrant locks for each text model to ensure thread safety.
|
181 |
+
self.text_model_locks: Dict[TextModelType, threading.RLock] = {
|
182 |
+
t: threading.RLock() for t in TextModelType
|
183 |
+
}
|
184 |
+
|
185 |
# Create a persistent asynchronous HTTP client.
|
186 |
self.async_http_client = httpx.AsyncClient(timeout=10)
|
187 |
|
|
|
218 |
# Set maximum sequence length based on configuration.
|
219 |
max_length = int(MaxModelLength[t_model_type.name].value)
|
220 |
self.text_models[t_model_type].max_seq_length = max_length
|
221 |
+
logger.info("Set max_seq_length=%d for text model: %s", max_length, info.model_id)
|
|
|
|
|
|
|
|
|
222 |
|
223 |
# Preload image models.
|
224 |
for i_model_type in ImageModelType:
|
225 |
+
model_id = ModelConfig(image_model_type=i_model_type).image_model_info.model_id
|
|
|
|
|
226 |
logger.info("Loading image model: %s", model_id)
|
227 |
model = AutoModel.from_pretrained(model_id).to(self.device)
|
228 |
model.eval() # Set the model to evaluation mode.
|
|
|
249 |
raise ValueError("Text input cannot be empty.")
|
250 |
return [input_text]
|
251 |
|
252 |
+
if not isinstance(input_text, list) or not all(isinstance(x, str) for x in input_text):
|
|
|
|
|
253 |
raise ValueError("Text input must be a string or a list of strings.")
|
254 |
|
255 |
if len(input_text) == 0:
|
|
|
270 |
raise ValueError("Image input cannot be empty.")
|
271 |
return [input_images]
|
272 |
|
273 |
+
if not isinstance(input_images, list) or not all(isinstance(x, str) for x in input_images):
|
|
|
|
|
274 |
raise ValueError("Image input must be a string or a list of strings.")
|
275 |
|
276 |
if len(input_images) == 0:
|
|
|
293 |
try:
|
294 |
# Attempt to get the tokenizer from the first module of the SentenceTransformer.
|
295 |
module = model._first_module()
|
296 |
+
if not hasattr(module, 'tokenizer'):
|
297 |
return text
|
298 |
tokenizer = module.tokenizer
|
299 |
# Tokenize without truncation.
|
300 |
encoded = tokenizer(text, add_special_tokens=True, truncation=False)
|
301 |
max_length = model.max_seq_length
|
302 |
+
if len(encoded['input_ids']) > max_length:
|
303 |
+
truncated_ids = encoded['input_ids'][:max_length]
|
304 |
+
truncated_text = tokenizer.decode(truncated_ids, skip_special_tokens=True)
|
|
|
|
|
305 |
return truncated_text
|
306 |
except Exception as e:
|
307 |
logger.warning("Error during text truncation: %s", str(e))
|
|
|
353 |
processed_data = processor(images=img, return_tensors="pt").to(self.device)
|
354 |
return processed_data
|
355 |
|
356 |
+
def _generate_text_embeddings(self, model_id: TextModelType, texts: List[str]) -> np.ndarray:
|
|
|
|
|
357 |
"""
|
358 |
Generate text embeddings using the SentenceTransformer model.
|
359 |
Single-text requests are cached using an LRU cache.
|
|
|
369 |
RuntimeError: If text embedding generation fails.
|
370 |
"""
|
371 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
372 |
model = self.text_models[model_id]
|
373 |
+
lock = self.text_model_locks[model_id]
|
374 |
+
with lock:
|
375 |
+
if len(texts) == 1:
|
376 |
+
single_text = texts[0]
|
377 |
+
key = md5(f"{model_id}:{single_text}".encode("utf-8")).hexdigest()[:8]
|
378 |
+
if key in self.lru_cache:
|
379 |
+
return self.lru_cache[key]
|
380 |
+
emb = model.encode([single_text])
|
381 |
+
self.lru_cache[key] = emb
|
382 |
+
return emb
|
383 |
+
|
384 |
+
return model.encode(texts)
|
385 |
except Exception as e:
|
386 |
raise RuntimeError(
|
387 |
f"Error generating text embeddings with model '{model_id}': {e}"
|
388 |
) from e
|
389 |
|
390 |
+
async def _async_generate_image_embeddings(self, model_id: ImageModelType, images: List[str]) -> np.ndarray:
|
|
|
|
|
391 |
"""
|
392 |
Asynchronously generate image embeddings.
|
393 |
|
|
|
411 |
)
|
412 |
# Assume all processed outputs have the same keys.
|
413 |
keys = processed_tensors[0].keys()
|
414 |
+
combined = {k: torch.cat([pt[k] for pt in processed_tensors], dim=0) for k in keys}
|
|
|
|
|
415 |
|
416 |
def infer():
|
417 |
with torch.no_grad():
|
418 |
+
embeddings = self.image_models[model_id].get_image_features(**combined)
|
|
|
|
|
419 |
return embeddings.cpu().numpy()
|
420 |
|
421 |
return await asyncio.to_thread(infer)
|
|
|
424 |
f"Error generating image embeddings with model '{model_id}': {e}"
|
425 |
) from e
|
426 |
|
427 |
+
async def generate_embeddings(self, model: str, inputs: Union[str, List[str]]) -> np.ndarray:
|
|
|
|
|
428 |
"""
|
429 |
Asynchronously generate embeddings for text or image inputs based on model type.
|
430 |
|
|
|
440 |
text_model_enum = TextModelType(model)
|
441 |
text_list = self._validate_text_list(inputs)
|
442 |
model_instance = self.text_models[text_model_enum]
|
443 |
+
lock = self.text_model_locks[text_model_enum]
|
444 |
+
with lock:
|
445 |
+
# Truncate each text if it exceeds the maximum allowed token length.
|
446 |
+
truncated_texts = [self._truncate_text(text, model_instance) for text in text_list]
|
447 |
return await asyncio.to_thread(
|
448 |
self._generate_text_embeddings, text_model_enum, truncated_texts
|
449 |
)
|
450 |
elif modality == ModelKind.IMAGE:
|
451 |
image_model_enum = ImageModelType(model)
|
452 |
image_list = self._validate_image_list(inputs)
|
453 |
+
return await self._async_generate_image_embeddings(image_model_enum, image_list)
|
|
|
|
|
454 |
|
455 |
+
async def rank(self, model: str, queries: Union[str, List[str]], candidates: Union[str, List[str]]) -> Dict[str, Any]:
|
|
|
|
|
|
|
|
|
|
|
456 |
"""
|
457 |
Asynchronously rank candidate texts/images against the provided queries.
|
458 |
Embeddings for queries and candidates are generated concurrently.
|
|
|
473 |
|
474 |
# Concurrently generate embeddings.
|
475 |
query_task = asyncio.create_task(self.generate_embeddings(model, queries))
|
476 |
+
candidate_task = asyncio.create_task(self.generate_embeddings(model, candidates))
|
477 |
+
query_embeds, candidate_embeds = await asyncio.gather(query_task, candidate_task)
|
|
|
|
|
|
|
|
|
478 |
|
479 |
# Compute cosine similarity.
|
480 |
sim_matrix = self.cosine_similarity(query_embeds, candidate_embeds)
|