Spaces:
Running
Running
chore: update something
Browse files- lightweight_embeddings/__init__.py +0 -21
- lightweight_embeddings/router.py +4 -23
- lightweight_embeddings/service.py +122 -154
lightweight_embeddings/__init__.py
CHANGED
@@ -1,24 +1,3 @@
|
|
1 |
-
# filename: __init__.py
|
2 |
-
|
3 |
-
"""
|
4 |
-
LightweightEmbeddings - FastAPI Application Entry Point
|
5 |
-
|
6 |
-
This application provides text and image embeddings using multiple text models and one image model.
|
7 |
-
|
8 |
-
Supported text model IDs:
|
9 |
-
- "multilingual-e5-small"
|
10 |
-
- "multilingual-e5-base"
|
11 |
-
- "multilingual-e5-large"
|
12 |
-
- "snowflake-arctic-embed-l-v2.0"
|
13 |
-
- "paraphrase-multilingual-MiniLM-L12-v2"
|
14 |
-
- "paraphrase-multilingual-mpnet-base-v2"
|
15 |
-
- "bge-m3"
|
16 |
-
- "gte-multilingual-base"
|
17 |
-
|
18 |
-
Supported image model ID:
|
19 |
-
- "siglip-base-patch16-256-multilingual"
|
20 |
-
"""
|
21 |
-
|
22 |
import gradio as gr
|
23 |
import requests
|
24 |
import json
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
import requests
|
3 |
import json
|
lightweight_embeddings/router.py
CHANGED
@@ -1,22 +1,3 @@
|
|
1 |
-
"""
|
2 |
-
FastAPI Router for Embeddings Service (Revised & Simplified)
|
3 |
-
|
4 |
-
Exposes the EmbeddingsService methods via a RESTful API.
|
5 |
-
|
6 |
-
Supported Text Model IDs:
|
7 |
-
- "multilingual-e5-small"
|
8 |
-
- "multilingual-e5-base"
|
9 |
-
- "multilingual-e5-large"
|
10 |
-
- "snowflake-arctic-embed-l-v2.0"
|
11 |
-
- "paraphrase-multilingual-MiniLM-L12-v2"
|
12 |
-
- "paraphrase-multilingual-mpnet-base-v2"
|
13 |
-
- "bge-m3"
|
14 |
-
- "gte-multilingual-base"
|
15 |
-
|
16 |
-
Supported Image Model IDs:
|
17 |
-
- "siglip-base-patch16-256-multilingual"
|
18 |
-
"""
|
19 |
-
|
20 |
from __future__ import annotations
|
21 |
|
22 |
import logging
|
@@ -158,10 +139,6 @@ async def create_embeddings(
|
|
158 |
},
|
159 |
}
|
160 |
|
161 |
-
background_tasks.add_task(
|
162 |
-
analytics.access, request.model, resp["usage"]["total_tokens"]
|
163 |
-
)
|
164 |
-
|
165 |
for idx, emb in enumerate(embeddings):
|
166 |
resp["data"].append(
|
167 |
{
|
@@ -171,6 +148,10 @@ async def create_embeddings(
|
|
171 |
}
|
172 |
)
|
173 |
|
|
|
|
|
|
|
|
|
174 |
return resp
|
175 |
|
176 |
except Exception as e:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from __future__ import annotations
|
2 |
|
3 |
import logging
|
|
|
139 |
},
|
140 |
}
|
141 |
|
|
|
|
|
|
|
|
|
142 |
for idx, emb in enumerate(embeddings):
|
143 |
resp["data"].append(
|
144 |
{
|
|
|
148 |
}
|
149 |
)
|
150 |
|
151 |
+
background_tasks.add_task(
|
152 |
+
analytics.access, request.model, resp["usage"]["total_tokens"]
|
153 |
+
)
|
154 |
+
|
155 |
return resp
|
156 |
|
157 |
except Exception as e:
|
lightweight_embeddings/service.py
CHANGED
@@ -1,29 +1,3 @@
|
|
1 |
-
"""
|
2 |
-
Lightweight Embeddings Service Module (Revised & Simplified)
|
3 |
-
|
4 |
-
This module provides a service for generating and comparing embeddings from text and images
|
5 |
-
using state-of-the-art transformer models. It supports both CPU and GPU inference.
|
6 |
-
|
7 |
-
Features:
|
8 |
-
- Text and image embedding generation
|
9 |
-
- Cross-modal similarity ranking
|
10 |
-
- Batch processing support
|
11 |
-
- Asynchronous API support
|
12 |
-
|
13 |
-
Supported Text Model IDs:
|
14 |
-
- "multilingual-e5-small"
|
15 |
-
- "multilingual-e5-base"
|
16 |
-
- "multilingual-e5-large"
|
17 |
-
- "snowflake-arctic-embed-l-v2.0"
|
18 |
-
- "paraphrase-multilingual-MiniLM-L12-v2"
|
19 |
-
- "paraphrase-multilingual-mpnet-base-v2"
|
20 |
-
- "bge-m3"
|
21 |
-
- "gte-multilingual-base"
|
22 |
-
|
23 |
-
Supported Image Model IDs:
|
24 |
-
- "google/siglip-base-patch16-256-multilingual" (default, but extensible)
|
25 |
-
"""
|
26 |
-
|
27 |
from __future__ import annotations
|
28 |
|
29 |
import logging
|
@@ -49,7 +23,6 @@ logging.basicConfig(level=logging.INFO)
|
|
49 |
class TextModelType(str, Enum):
|
50 |
"""
|
51 |
Enumeration of supported text models.
|
52 |
-
Adjust as needed for your environment.
|
53 |
"""
|
54 |
|
55 |
MULTILINGUAL_E5_SMALL = "multilingual-e5-small"
|
@@ -72,7 +45,7 @@ class ImageModelType(str, Enum):
|
|
72 |
|
73 |
class ModelInfo(NamedTuple):
|
74 |
"""
|
75 |
-
|
76 |
- model_id: Hugging Face model ID (or local path)
|
77 |
- onnx_file: Path to ONNX file (if available)
|
78 |
"""
|
@@ -91,14 +64,12 @@ class ModelConfig:
|
|
91 |
image_model_type: ImageModelType = (
|
92 |
ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL
|
93 |
)
|
94 |
-
|
95 |
-
# If you need extra parameters like `logit_scale`, etc., keep them here
|
96 |
-
logit_scale: float = 4.60517
|
97 |
|
98 |
@property
|
99 |
def text_model_info(self) -> ModelInfo:
|
100 |
"""
|
101 |
-
|
102 |
"""
|
103 |
text_configs = {
|
104 |
TextModelType.MULTILINGUAL_E5_SMALL: ModelInfo(
|
@@ -139,7 +110,7 @@ class ModelConfig:
|
|
139 |
@property
|
140 |
def image_model_info(self) -> ModelInfo:
|
141 |
"""
|
142 |
-
|
143 |
"""
|
144 |
image_configs = {
|
145 |
ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL: ModelInfo(
|
@@ -156,8 +127,8 @@ class ModelKind(str, Enum):
|
|
156 |
|
157 |
def detect_model_kind(model_id: str) -> ModelKind:
|
158 |
"""
|
159 |
-
Detect whether model_id
|
160 |
-
Raises ValueError if
|
161 |
"""
|
162 |
if model_id in [m.value for m in TextModelType]:
|
163 |
return ModelKind.TEXT
|
@@ -173,21 +144,21 @@ def detect_model_kind(model_id: str) -> ModelKind:
|
|
173 |
|
174 |
class EmbeddingsService:
|
175 |
"""
|
176 |
-
Service for generating text/image embeddings and performing ranking.
|
|
|
177 |
"""
|
178 |
|
179 |
def __init__(self, config: Optional[ModelConfig] = None):
|
180 |
-
self.lru_cache = LRUCache(maxsize=10_000)
|
181 |
-
|
182 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
183 |
self.config = config or ModelConfig()
|
184 |
|
185 |
-
#
|
186 |
self.text_models: Dict[TextModelType, SentenceTransformer] = {}
|
187 |
self.image_models: Dict[ImageModelType, AutoModel] = {}
|
188 |
self.image_processors: Dict[ImageModelType, AutoProcessor] = {}
|
189 |
|
190 |
-
# Load all models
|
191 |
self._load_all_models()
|
192 |
|
193 |
def _load_all_models(self) -> None:
|
@@ -195,40 +166,37 @@ class EmbeddingsService:
|
|
195 |
Pre-load all known text and image models for quick switching.
|
196 |
"""
|
197 |
try:
|
|
|
198 |
for t_model_type in TextModelType:
|
199 |
info = ModelConfig(text_model_type=t_model_type).text_model_info
|
200 |
logger.info("Loading text model: %s", info.model_id)
|
201 |
|
202 |
-
# If you have an ONNX file AND your SentenceTransformer supports ONNX
|
203 |
if info.onnx_file:
|
204 |
logger.info("Using ONNX file: %s", info.onnx_file)
|
205 |
-
# The following 'backend' & 'model_kwargs' parameters
|
206 |
-
# are recognized only in special/certain distributions of SentenceTransformer
|
207 |
self.text_models[t_model_type] = SentenceTransformer(
|
208 |
info.model_id,
|
209 |
device=self.device,
|
210 |
-
backend="onnx",
|
211 |
model_kwargs={
|
212 |
-
"provider": "CPUExecutionProvider",
|
213 |
"file_name": info.onnx_file,
|
214 |
},
|
215 |
trust_remote_code=True,
|
216 |
)
|
217 |
else:
|
218 |
-
# Fallback: standard HF loading
|
219 |
self.text_models[t_model_type] = SentenceTransformer(
|
220 |
info.model_id,
|
221 |
device=self.device,
|
222 |
trust_remote_code=True,
|
223 |
)
|
224 |
|
|
|
225 |
for i_model_type in ImageModelType:
|
226 |
model_id = ModelConfig(
|
227 |
image_model_type=i_model_type
|
228 |
).image_model_info.model_id
|
229 |
logger.info("Loading image model: %s", model_id)
|
230 |
|
231 |
-
# Typically, for CLIP-like models:
|
232 |
model = AutoModel.from_pretrained(model_id).to(self.device)
|
233 |
processor = AutoProcessor.from_pretrained(model_id)
|
234 |
|
@@ -242,9 +210,10 @@ class EmbeddingsService:
|
|
242 |
raise RuntimeError(msg) from e
|
243 |
|
244 |
@staticmethod
|
245 |
-
def
|
246 |
"""
|
247 |
-
|
|
|
248 |
"""
|
249 |
if isinstance(input_text, str):
|
250 |
if not input_text.strip():
|
@@ -262,27 +231,42 @@ class EmbeddingsService:
|
|
262 |
return input_text
|
263 |
|
264 |
@staticmethod
|
265 |
-
def
|
266 |
-
|
267 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
268 |
|
269 |
-
def _process_image(self, path_or_url:
|
270 |
"""
|
271 |
-
|
|
|
272 |
"""
|
273 |
try:
|
274 |
-
if
|
275 |
-
# Local file path
|
276 |
-
img = Image.open(path_or_url).convert("RGB")
|
277 |
-
else:
|
278 |
-
# URL
|
279 |
resp = requests.get(path_or_url, timeout=10)
|
280 |
resp.raise_for_status()
|
281 |
img = Image.open(BytesIO(resp.content)).convert("RGB")
|
|
|
|
|
282 |
|
283 |
-
|
284 |
-
|
285 |
-
return
|
286 |
except Exception as e:
|
287 |
raise ValueError(f"Error processing image '{path_or_url}': {str(e)}") from e
|
288 |
|
@@ -292,145 +276,125 @@ class EmbeddingsService:
|
|
292 |
texts: List[str],
|
293 |
) -> np.ndarray:
|
294 |
"""
|
295 |
-
|
296 |
-
|
297 |
"""
|
298 |
try:
|
299 |
if len(texts) == 1:
|
300 |
-
|
|
|
301 |
if key in self.lru_cache:
|
302 |
return self.lru_cache[key]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
303 |
model = self.text_models[model_id]
|
304 |
-
|
305 |
|
306 |
-
if len(texts) == 1:
|
307 |
-
self.lru_cache[key] = embeddings
|
308 |
-
return embeddings
|
309 |
except Exception as e:
|
310 |
raise RuntimeError(
|
311 |
-
f"Error generating text embeddings
|
312 |
) from e
|
313 |
|
314 |
def _generate_image_embeddings(
|
315 |
self,
|
316 |
model_id: ImageModelType,
|
317 |
-
images:
|
318 |
-
batch_size: Optional[int] = None,
|
319 |
) -> np.ndarray:
|
320 |
"""
|
321 |
-
|
322 |
-
|
323 |
"""
|
324 |
try:
|
325 |
model = self.image_models[model_id]
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
|
339 |
-
|
340 |
-
|
341 |
-
keys = tensors[0].keys()
|
342 |
-
combined = {k: torch.cat([t[k] for t in tensors], dim=0) for k in keys}
|
343 |
-
with torch.no_grad():
|
344 |
-
emb = model.get_image_features(**combined)
|
345 |
-
return emb.cpu().numpy()
|
346 |
-
|
347 |
-
# Process in smaller batches
|
348 |
-
all_embeddings = []
|
349 |
-
for i in range(0, len(images), batch_size):
|
350 |
-
batch_images = images[i : i + batch_size]
|
351 |
-
# Process each sub-batch
|
352 |
-
tensors = []
|
353 |
-
for img_path in batch_images:
|
354 |
-
tensors.append(self._process_image(img_path))
|
355 |
-
keys = tensors[0].keys()
|
356 |
-
combined = {k: torch.cat([t[k] for t in tensors], dim=0) for k in keys}
|
357 |
-
|
358 |
-
with torch.no_grad():
|
359 |
-
emb = model.get_image_features(**combined)
|
360 |
-
all_embeddings.append(emb.cpu().numpy())
|
361 |
-
|
362 |
-
return np.vstack(all_embeddings)
|
363 |
|
364 |
except Exception as e:
|
365 |
raise RuntimeError(
|
366 |
-
f"Error generating image embeddings
|
367 |
) from e
|
368 |
|
369 |
async def generate_embeddings(
|
370 |
self,
|
371 |
model: str,
|
372 |
inputs: Union[str, List[str]],
|
373 |
-
batch_size: Optional[int] = None,
|
374 |
) -> np.ndarray:
|
375 |
"""
|
376 |
-
Asynchronously
|
377 |
"""
|
378 |
-
# Determine if it's text or image
|
379 |
modality = detect_model_kind(model)
|
380 |
-
model_id = (
|
381 |
-
TextModelType(model)
|
382 |
-
if modality == ModelKind.TEXT
|
383 |
-
else ImageModelType(model)
|
384 |
-
)
|
385 |
|
386 |
-
|
387 |
-
|
388 |
-
text_list = self.
|
389 |
-
return self._generate_text_embeddings(
|
390 |
-
|
391 |
-
|
392 |
-
|
393 |
-
)
|
|
|
394 |
|
395 |
async def rank(
|
396 |
self,
|
397 |
model: str,
|
398 |
queries: Union[str, List[str]],
|
399 |
-
candidates: List[str],
|
400 |
-
batch_size: Optional[int] = None,
|
401 |
) -> Dict[str, Any]:
|
402 |
"""
|
403 |
-
|
404 |
-
|
|
|
|
|
|
|
405 |
"""
|
406 |
-
# Determine if it's text or image
|
407 |
modality = detect_model_kind(model)
|
408 |
-
|
409 |
-
|
410 |
-
|
411 |
-
|
412 |
-
|
|
|
413 |
|
414 |
# 1) Generate embeddings for queries
|
415 |
-
query_embeds = await self.generate_embeddings(
|
416 |
-
|
417 |
-
)
|
418 |
-
#
|
419 |
-
candidate_embeds = await self.generate_embeddings(
|
420 |
-
model=model_id, inputs=candidates, batch_size=batch_size
|
421 |
-
)
|
422 |
|
423 |
# 3) Compute cosine similarity
|
424 |
sim_matrix = self.cosine_similarity(query_embeds, candidate_embeds)
|
425 |
|
426 |
-
# 4) Apply logit scale + softmax
|
427 |
scaled = np.exp(self.config.logit_scale) * sim_matrix
|
428 |
probs = self.softmax(scaled)
|
429 |
|
430 |
-
# 5)
|
431 |
-
|
432 |
-
|
433 |
-
|
|
|
|
|
|
|
|
|
434 |
usage = {
|
435 |
"prompt_tokens": total_tokens,
|
436 |
"total_tokens": total_tokens,
|
@@ -444,27 +408,31 @@ class EmbeddingsService:
|
|
444 |
|
445 |
def estimate_tokens(self, input_data: Union[str, List[str]]) -> int:
|
446 |
"""
|
447 |
-
|
|
|
448 |
"""
|
449 |
-
texts = self.
|
450 |
model = self.text_models[self.config.text_model_type]
|
451 |
tokenized = model.tokenize(texts)
|
|
|
452 |
return sum(len(ids) for ids in tokenized["input_ids"])
|
453 |
|
454 |
@staticmethod
|
455 |
def softmax(scores: np.ndarray) -> np.ndarray:
|
456 |
"""
|
457 |
-
|
458 |
"""
|
|
|
459 |
exps = np.exp(scores - np.max(scores, axis=-1, keepdims=True))
|
460 |
return exps / np.sum(exps, axis=-1, keepdims=True)
|
461 |
|
462 |
@staticmethod
|
463 |
def cosine_similarity(a: np.ndarray, b: np.ndarray) -> np.ndarray:
|
464 |
"""
|
|
|
465 |
a: (N, D)
|
466 |
b: (M, D)
|
467 |
-
Return: (N, M) of
|
468 |
"""
|
469 |
a_norm = a / (np.linalg.norm(a, axis=1, keepdims=True) + 1e-9)
|
470 |
b_norm = b / (np.linalg.norm(b, axis=1, keepdims=True) + 1e-9)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from __future__ import annotations
|
2 |
|
3 |
import logging
|
|
|
23 |
class TextModelType(str, Enum):
|
24 |
"""
|
25 |
Enumeration of supported text models.
|
|
|
26 |
"""
|
27 |
|
28 |
MULTILINGUAL_E5_SMALL = "multilingual-e5-small"
|
|
|
45 |
|
46 |
class ModelInfo(NamedTuple):
|
47 |
"""
|
48 |
+
This container maps an enum to:
|
49 |
- model_id: Hugging Face model ID (or local path)
|
50 |
- onnx_file: Path to ONNX file (if available)
|
51 |
"""
|
|
|
64 |
image_model_type: ImageModelType = (
|
65 |
ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL
|
66 |
)
|
67 |
+
logit_scale: float = 4.60517 # Example scale used in cross-modal similarity
|
|
|
|
|
68 |
|
69 |
@property
|
70 |
def text_model_info(self) -> ModelInfo:
|
71 |
"""
|
72 |
+
Returns ModelInfo for the configured text_model_type.
|
73 |
"""
|
74 |
text_configs = {
|
75 |
TextModelType.MULTILINGUAL_E5_SMALL: ModelInfo(
|
|
|
110 |
@property
|
111 |
def image_model_info(self) -> ModelInfo:
|
112 |
"""
|
113 |
+
Returns ModelInfo for the configured image_model_type.
|
114 |
"""
|
115 |
image_configs = {
|
116 |
ImageModelType.SIGLIP_BASE_PATCH16_256_MULTILINGUAL: ModelInfo(
|
|
|
127 |
|
128 |
def detect_model_kind(model_id: str) -> ModelKind:
|
129 |
"""
|
130 |
+
Detect whether model_id belongs to a text or an image model.
|
131 |
+
Raises ValueError if the model is not recognized.
|
132 |
"""
|
133 |
if model_id in [m.value for m in TextModelType]:
|
134 |
return ModelKind.TEXT
|
|
|
144 |
|
145 |
class EmbeddingsService:
|
146 |
"""
|
147 |
+
Service for generating text/image embeddings and performing similarity ranking.
|
148 |
+
Batch size has been removed. Single or multiple inputs are handled uniformly.
|
149 |
"""
|
150 |
|
151 |
def __init__(self, config: Optional[ModelConfig] = None):
|
152 |
+
self.lru_cache = LRUCache(maxsize=10_000)
|
|
|
153 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
154 |
self.config = config or ModelConfig()
|
155 |
|
156 |
+
# Dictionaries to hold preloaded models
|
157 |
self.text_models: Dict[TextModelType, SentenceTransformer] = {}
|
158 |
self.image_models: Dict[ImageModelType, AutoModel] = {}
|
159 |
self.image_processors: Dict[ImageModelType, AutoProcessor] = {}
|
160 |
|
161 |
+
# Load all relevant models on init
|
162 |
self._load_all_models()
|
163 |
|
164 |
def _load_all_models(self) -> None:
|
|
|
166 |
Pre-load all known text and image models for quick switching.
|
167 |
"""
|
168 |
try:
|
169 |
+
# Preload text models
|
170 |
for t_model_type in TextModelType:
|
171 |
info = ModelConfig(text_model_type=t_model_type).text_model_info
|
172 |
logger.info("Loading text model: %s", info.model_id)
|
173 |
|
|
|
174 |
if info.onnx_file:
|
175 |
logger.info("Using ONNX file: %s", info.onnx_file)
|
|
|
|
|
176 |
self.text_models[t_model_type] = SentenceTransformer(
|
177 |
info.model_id,
|
178 |
device=self.device,
|
179 |
+
backend="onnx",
|
180 |
model_kwargs={
|
181 |
+
"provider": "CPUExecutionProvider",
|
182 |
"file_name": info.onnx_file,
|
183 |
},
|
184 |
trust_remote_code=True,
|
185 |
)
|
186 |
else:
|
|
|
187 |
self.text_models[t_model_type] = SentenceTransformer(
|
188 |
info.model_id,
|
189 |
device=self.device,
|
190 |
trust_remote_code=True,
|
191 |
)
|
192 |
|
193 |
+
# Preload image models
|
194 |
for i_model_type in ImageModelType:
|
195 |
model_id = ModelConfig(
|
196 |
image_model_type=i_model_type
|
197 |
).image_model_info.model_id
|
198 |
logger.info("Loading image model: %s", model_id)
|
199 |
|
|
|
200 |
model = AutoModel.from_pretrained(model_id).to(self.device)
|
201 |
processor = AutoProcessor.from_pretrained(model_id)
|
202 |
|
|
|
210 |
raise RuntimeError(msg) from e
|
211 |
|
212 |
@staticmethod
|
213 |
+
def _validate_text_list(input_text: Union[str, List[str]]) -> List[str]:
|
214 |
"""
|
215 |
+
Convert text input into a non-empty list of strings.
|
216 |
+
Raises ValueError if the input is invalid.
|
217 |
"""
|
218 |
if isinstance(input_text, str):
|
219 |
if not input_text.strip():
|
|
|
231 |
return input_text
|
232 |
|
233 |
@staticmethod
|
234 |
+
def _validate_image_list(input_images: Union[str, List[str]]) -> List[str]:
|
235 |
+
"""
|
236 |
+
Convert image input into a non-empty list of image paths/URLs.
|
237 |
+
Raises ValueError if the input is invalid.
|
238 |
+
"""
|
239 |
+
if isinstance(input_images, str):
|
240 |
+
if not input_images.strip():
|
241 |
+
raise ValueError("Image input cannot be empty.")
|
242 |
+
return [input_images]
|
243 |
+
|
244 |
+
if not isinstance(input_images, list) or not all(
|
245 |
+
isinstance(x, str) for x in input_images
|
246 |
+
):
|
247 |
+
raise ValueError("Image input must be a string or a list of strings.")
|
248 |
+
|
249 |
+
if len(input_images) == 0:
|
250 |
+
raise ValueError("Image input list cannot be empty.")
|
251 |
+
|
252 |
+
return input_images
|
253 |
|
254 |
+
def _process_image(self, path_or_url: str) -> Dict[str, torch.Tensor]:
|
255 |
"""
|
256 |
+
Loads and processes a single image from local path or URL.
|
257 |
+
Returns a dictionary of tensors ready for the model.
|
258 |
"""
|
259 |
try:
|
260 |
+
if path_or_url.startswith("http"):
|
|
|
|
|
|
|
|
|
261 |
resp = requests.get(path_or_url, timeout=10)
|
262 |
resp.raise_for_status()
|
263 |
img = Image.open(BytesIO(resp.content)).convert("RGB")
|
264 |
+
else:
|
265 |
+
img = Image.open(Path(path_or_url)).convert("RGB")
|
266 |
|
267 |
+
processor = self.image_processors[self.config.image_model_type]
|
268 |
+
processed_data = processor(images=img, return_tensors="pt").to(self.device)
|
269 |
+
return processed_data
|
270 |
except Exception as e:
|
271 |
raise ValueError(f"Error processing image '{path_or_url}': {str(e)}") from e
|
272 |
|
|
|
276 |
texts: List[str],
|
277 |
) -> np.ndarray:
|
278 |
"""
|
279 |
+
Generates text embeddings using the SentenceTransformer-based model.
|
280 |
+
Utilizes an LRU cache for single-input scenarios.
|
281 |
"""
|
282 |
try:
|
283 |
if len(texts) == 1:
|
284 |
+
single_text = texts[0]
|
285 |
+
key = md5(single_text.encode("utf-8")).hexdigest()
|
286 |
if key in self.lru_cache:
|
287 |
return self.lru_cache[key]
|
288 |
+
|
289 |
+
model = self.text_models[model_id]
|
290 |
+
emb = model.encode([single_text])
|
291 |
+
self.lru_cache[key] = emb
|
292 |
+
return emb
|
293 |
+
|
294 |
+
# For multiple texts, no LRU cache is used
|
295 |
model = self.text_models[model_id]
|
296 |
+
return model.encode(texts)
|
297 |
|
|
|
|
|
|
|
298 |
except Exception as e:
|
299 |
raise RuntimeError(
|
300 |
+
f"Error generating text embeddings with model '{model_id}': {e}"
|
301 |
) from e
|
302 |
|
303 |
def _generate_image_embeddings(
|
304 |
self,
|
305 |
model_id: ImageModelType,
|
306 |
+
images: List[str],
|
|
|
307 |
) -> np.ndarray:
|
308 |
"""
|
309 |
+
Generates image embeddings using the CLIP-like transformer model.
|
310 |
+
Handles single or multiple images uniformly (no batch size parameter).
|
311 |
"""
|
312 |
try:
|
313 |
model = self.image_models[model_id]
|
314 |
+
# Collect processed inputs in a single batch
|
315 |
+
processed_tensors = []
|
316 |
+
for img_path in images:
|
317 |
+
processed_tensors.append(self._process_image(img_path))
|
318 |
+
|
319 |
+
# Keys should be the same for all processed outputs
|
320 |
+
keys = processed_tensors[0].keys()
|
321 |
+
# Concatenate along the batch dimension
|
322 |
+
combined = {
|
323 |
+
k: torch.cat([pt[k] for pt in processed_tensors], dim=0) for k in keys
|
324 |
+
}
|
325 |
+
|
326 |
+
with torch.no_grad():
|
327 |
+
embeddings = model.get_image_features(**combined)
|
328 |
+
return embeddings.cpu().numpy()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
329 |
|
330 |
except Exception as e:
|
331 |
raise RuntimeError(
|
332 |
+
f"Error generating image embeddings with model '{model_id}': {e}"
|
333 |
) from e
|
334 |
|
335 |
async def generate_embeddings(
|
336 |
self,
|
337 |
model: str,
|
338 |
inputs: Union[str, List[str]],
|
|
|
339 |
) -> np.ndarray:
|
340 |
"""
|
341 |
+
Asynchronously generates embeddings for either text or image based on the model type.
|
342 |
"""
|
|
|
343 |
modality = detect_model_kind(model)
|
|
|
|
|
|
|
|
|
|
|
344 |
|
345 |
+
if modality == ModelKind.TEXT:
|
346 |
+
text_model_id = TextModelType(model)
|
347 |
+
text_list = self._validate_text_list(inputs)
|
348 |
+
return self._generate_text_embeddings(text_model_id, text_list)
|
349 |
+
|
350 |
+
elif modality == ModelKind.IMAGE:
|
351 |
+
image_model_id = ImageModelType(model)
|
352 |
+
image_list = self._validate_image_list(inputs)
|
353 |
+
return self._generate_image_embeddings(image_model_id, image_list)
|
354 |
|
355 |
async def rank(
|
356 |
self,
|
357 |
model: str,
|
358 |
queries: Union[str, List[str]],
|
359 |
+
candidates: Union[str, List[str]],
|
|
|
360 |
) -> Dict[str, Any]:
|
361 |
"""
|
362 |
+
Ranks text `candidates` given `queries`, which can be text or images.
|
363 |
+
Always returns a dictionary of { probabilities, cosine_similarities, usage }.
|
364 |
+
|
365 |
+
Note: This implementation uses the same model for both queries and candidates.
|
366 |
+
For true cross-modal ranking, you might need separate models or a shared model.
|
367 |
"""
|
|
|
368 |
modality = detect_model_kind(model)
|
369 |
+
|
370 |
+
# Convert the string model to the appropriate enum
|
371 |
+
if modality == ModelKind.TEXT:
|
372 |
+
model_enum = TextModelType(model)
|
373 |
+
else:
|
374 |
+
model_enum = ImageModelType(model)
|
375 |
|
376 |
# 1) Generate embeddings for queries
|
377 |
+
query_embeds = await self.generate_embeddings(model_enum.value, queries)
|
378 |
+
|
379 |
+
# 2) Generate embeddings for candidates (assumed text if queries are text;
|
380 |
+
# or if queries are images, also use the image model for candidates).
|
381 |
+
candidate_embeds = await self.generate_embeddings(model_enum.value, candidates)
|
|
|
|
|
382 |
|
383 |
# 3) Compute cosine similarity
|
384 |
sim_matrix = self.cosine_similarity(query_embeds, candidate_embeds)
|
385 |
|
386 |
+
# 4) Apply logit scale + softmax to obtain probabilities
|
387 |
scaled = np.exp(self.config.logit_scale) * sim_matrix
|
388 |
probs = self.softmax(scaled)
|
389 |
|
390 |
+
# 5) Estimate token usage if we're dealing with text
|
391 |
+
if modality == ModelKind.TEXT:
|
392 |
+
query_tokens = self.estimate_tokens(queries)
|
393 |
+
candidate_tokens = self.estimate_tokens(candidates)
|
394 |
+
total_tokens = query_tokens + candidate_tokens
|
395 |
+
else:
|
396 |
+
total_tokens = 0
|
397 |
+
|
398 |
usage = {
|
399 |
"prompt_tokens": total_tokens,
|
400 |
"total_tokens": total_tokens,
|
|
|
408 |
|
409 |
def estimate_tokens(self, input_data: Union[str, List[str]]) -> int:
|
410 |
"""
|
411 |
+
Estimates token count using the SentenceTransformer tokenizer.
|
412 |
+
Only applicable if the current configured model is a text model.
|
413 |
"""
|
414 |
+
texts = self._validate_text_list(input_data)
|
415 |
model = self.text_models[self.config.text_model_type]
|
416 |
tokenized = model.tokenize(texts)
|
417 |
+
# Summing over the lengths of input_ids for each example
|
418 |
return sum(len(ids) for ids in tokenized["input_ids"])
|
419 |
|
420 |
@staticmethod
|
421 |
def softmax(scores: np.ndarray) -> np.ndarray:
|
422 |
"""
|
423 |
+
Applies the standard softmax function along the last dimension.
|
424 |
"""
|
425 |
+
# Stabilize scores by subtracting max
|
426 |
exps = np.exp(scores - np.max(scores, axis=-1, keepdims=True))
|
427 |
return exps / np.sum(exps, axis=-1, keepdims=True)
|
428 |
|
429 |
@staticmethod
|
430 |
def cosine_similarity(a: np.ndarray, b: np.ndarray) -> np.ndarray:
|
431 |
"""
|
432 |
+
Computes the pairwise cosine similarity between all rows of a and b.
|
433 |
a: (N, D)
|
434 |
b: (M, D)
|
435 |
+
Return: (N, M) matrix of cosine similarities
|
436 |
"""
|
437 |
a_norm = a / (np.linalg.norm(a, axis=1, keepdims=True) + 1e-9)
|
438 |
b_norm = b / (np.linalg.norm(b, axis=1, keepdims=True) + 1e-9)
|