Spaces:
Running
Running
File size: 14,493 Bytes
64935ac | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 | """
Image encoder microservice.
This service stays lightweight by default:
- lazy-loads the vision encoder instead of preloading on boot
- prefers the vision-only CLIP path when available
- constrains CPU thread usage
- optionally applies dynamic int8 quantization on CPU
- keeps self-ping disabled unless explicitly enabled
"""
from __future__ import annotations
import asyncio
import base64
import io
import os
import threading
from typing import Dict, List
import httpx
import numpy as np
from fastapi import BackgroundTasks, FastAPI, File, HTTPException, UploadFile
from fastapi.middleware.cors import CORSMiddleware
from loguru import logger
from PIL import Image
from pydantic import BaseModel
def _env_bool(name: str, default: bool) -> bool:
return os.getenv(name, str(default).lower()).strip().lower() == "true"
def _recommended_threads() -> int:
cpu_count = max(1, os.cpu_count() or 1)
return min(4, cpu_count)
COSMO_AI_URL = os.getenv("COSMO_AI_URL", "https://shubhjn-cosmo-ai.hf.space").strip().rstrip("/")
IMAGE_ENCODER_MODEL_ID = os.getenv("IMAGE_ENCODER_MODEL_ID", "openai/clip-vit-base-patch32").strip()
IMAGE_ENCODER_DEVICE = os.getenv("IMAGE_ENCODER_DEVICE", "cpu").strip() or "cpu"
IMAGE_ENCODER_THREADS = max(1, int(os.getenv("IMAGE_ENCODER_THREADS", str(_recommended_threads()))))
IMAGE_ENCODER_MAX_IMAGE_DIM = max(224, int(os.getenv("IMAGE_ENCODER_MAX_IMAGE_DIM", "384")))
IMAGE_ENCODER_PRELOAD = _env_bool("IMAGE_ENCODER_PRELOAD", False)
IMAGE_ENCODER_KEEPALIVE = _env_bool("IMAGE_ENCODER_KEEPALIVE", False)
IMAGE_ENCODER_QUANTIZE = _env_bool("IMAGE_ENCODER_QUANTIZE", IMAGE_ENCODER_DEVICE == "cpu")
IMAGE_ENCODER_KEEPALIVE_INTERVAL_SECONDS = max(
300,
int(os.getenv("IMAGE_ENCODER_KEEPALIVE_INTERVAL_SECONDS", str(30 * 60))),
)
def apply_local_tuning() -> None:
os.environ.setdefault("OMP_NUM_THREADS", str(IMAGE_ENCODER_THREADS))
os.environ.setdefault("OPENBLAS_NUM_THREADS", str(IMAGE_ENCODER_THREADS))
os.environ.setdefault("MKL_NUM_THREADS", str(IMAGE_ENCODER_THREADS))
os.environ.setdefault("NUMEXPR_NUM_THREADS", str(IMAGE_ENCODER_THREADS))
os.environ.setdefault("VECLIB_MAXIMUM_THREADS", str(IMAGE_ENCODER_THREADS))
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
apply_local_tuning()
# Try to import a leaner vision-only CLIP path first.
try:
import torch
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
CLIP_VISION_ONLY_AVAILABLE = True
except ImportError:
torch = None
CLIPImageProcessor = None
CLIPVisionModelWithProjection = None
CLIP_VISION_ONLY_AVAILABLE = False
try:
if torch is None:
import torch
except ImportError:
torch = None
try:
from transformers import CLIPModel, CLIPProcessor
CLIP_FULL_AVAILABLE = True
except ImportError:
CLIP_FULL_AVAILABLE = False
CLIPModel = None
CLIPProcessor = None
CLIP_AVAILABLE = torch is not None and (CLIP_VISION_ONLY_AVAILABLE or CLIP_FULL_AVAILABLE)
if not CLIP_AVAILABLE:
logger.warning("transformers/torch vision encoder dependencies are unavailable")
if CLIP_AVAILABLE and torch is not None:
try:
torch.set_num_threads(IMAGE_ENCODER_THREADS)
except Exception:
pass
try:
torch.set_num_interop_threads(min(2, IMAGE_ENCODER_THREADS))
except Exception:
pass
app = FastAPI(
title="Image Encoder Service",
description="Lightweight image encoding service for Cosmo AI",
version="1.1.0",
)
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
class EncodeRequest(BaseModel):
image_base64: str
send_to_cosmo: bool = False
class EncodeResponse(BaseModel):
embedding: List[float]
text_representation: str
dimension: int
sent_to_cosmo: bool = False
_MODEL_LOCK = threading.Lock()
_clip_model = None
_clip_processor = None
_clip_backend = "unloaded"
def _resize_for_encoding(image: Image.Image) -> Image.Image:
image = image.convert("RGB")
resampling = getattr(Image, "Resampling", Image).LANCZOS
image.thumbnail((IMAGE_ENCODER_MAX_IMAGE_DIM, IMAGE_ENCODER_MAX_IMAGE_DIM), resampling)
return image
def _apply_dynamic_quantization(model):
if not (CLIP_AVAILABLE and torch is not None and IMAGE_ENCODER_DEVICE == "cpu" and IMAGE_ENCODER_QUANTIZE):
return model, False
try:
quantized_model = torch.ao.quantization.quantize_dynamic(model, {torch.nn.Linear}, dtype=torch.qint8)
return quantized_model, True
except Exception as exc:
logger.warning("Dynamic quantization skipped: {}", exc)
return model, False
def load_clip_model():
global _clip_model, _clip_processor, _clip_backend
if not CLIP_AVAILABLE:
raise RuntimeError("CLIP encoder dependencies are not installed")
if _clip_model is not None and _clip_processor is not None:
return _clip_model, _clip_processor, _clip_backend
with _MODEL_LOCK:
if _clip_model is not None and _clip_processor is not None:
return _clip_model, _clip_processor, _clip_backend
logger.info(
"Loading image encoder model={} device={} preload={} quantize={} max_dim={}",
IMAGE_ENCODER_MODEL_ID,
IMAGE_ENCODER_DEVICE,
IMAGE_ENCODER_PRELOAD,
IMAGE_ENCODER_QUANTIZE,
IMAGE_ENCODER_MAX_IMAGE_DIM,
)
backend_name = "clip_full"
if CLIP_VISION_ONLY_AVAILABLE:
model = CLIPVisionModelWithProjection.from_pretrained(IMAGE_ENCODER_MODEL_ID)
processor = CLIPImageProcessor.from_pretrained(IMAGE_ENCODER_MODEL_ID)
backend_name = "clip_vision_projection"
elif CLIP_FULL_AVAILABLE:
model = CLIPModel.from_pretrained(IMAGE_ENCODER_MODEL_ID)
processor = CLIPProcessor.from_pretrained(IMAGE_ENCODER_MODEL_ID)
else:
raise RuntimeError("No compatible CLIP image encoder backend is available")
if IMAGE_ENCODER_DEVICE != "cpu":
model = model.to(IMAGE_ENCODER_DEVICE)
model.eval()
model, quantized = _apply_dynamic_quantization(model)
if quantized:
backend_name = f"{backend_name}+int8"
_clip_model = model
_clip_processor = processor
_clip_backend = backend_name
logger.info("Image encoder ready with backend={}", _clip_backend)
return _clip_model, _clip_processor, _clip_backend
def _encode_with_loaded_model(image: Image.Image) -> Dict[str, object]:
model, processor, backend = load_clip_model()
prepared_image = _resize_for_encoding(image)
if backend.startswith("clip_vision_projection"):
inputs = processor(images=prepared_image, return_tensors="pt")
pixel_values = inputs["pixel_values"]
if IMAGE_ENCODER_DEVICE != "cpu":
pixel_values = pixel_values.to(IMAGE_ENCODER_DEVICE)
with torch.inference_mode():
outputs = model(pixel_values=pixel_values)
image_features = outputs.image_embeds
else:
inputs = processor(images=prepared_image, return_tensors="pt")
if IMAGE_ENCODER_DEVICE != "cpu":
inputs = {key: value.to(IMAGE_ENCODER_DEVICE) for key, value in inputs.items()}
with torch.inference_mode():
image_features = model.get_image_features(**inputs)
embedding = image_features[0].detach().cpu().numpy().astype(np.float32)
norm = np.linalg.norm(embedding)
if norm > 0:
embedding = embedding / norm
text_repr = create_text_representation(prepared_image, embedding)
return {
"embedding": embedding.tolist(),
"text_representation": text_repr,
"dimension": int(embedding.shape[0]),
"backend": backend,
}
def encode_image(image: Image.Image) -> Dict[str, object]:
if not CLIP_AVAILABLE:
raise RuntimeError("CLIP encoder is unavailable")
return _encode_with_loaded_model(image)
def create_text_representation(image: Image.Image, embedding: np.ndarray) -> str:
rgb = np.asarray(image.convert("RGB"), dtype=np.float32) / 255.0
luminance = float(rgb.mean())
channel_means = rgb.mean(axis=(0, 1))
channel_variance = float(rgb.var())
aspect_ratio = image.width / max(1, image.height)
chunks = np.array_split(embedding, 12)
lines = [
"<vision_input>",
f"Image size: {image.width}x{image.height}",
f"Aspect ratio: {aspect_ratio:.2f}",
f"Brightness: {luminance:.3f}",
f"Color balance rgb=({channel_means[0]:.3f}, {channel_means[1]:.3f}, {channel_means[2]:.3f})",
f"Visual variance: {channel_variance:.4f}",
"Embedding groups:",
]
for index, chunk in enumerate(chunks, start=1):
lines.append(
f"- Group {index}: mean={float(chunk.mean()):.4f}, std={float(chunk.std()):.4f}, peak={float(np.max(np.abs(chunk))):.4f}"
)
lines.append("</vision_input>")
return "\n".join(lines)
async def send_to_cosmo_ai(embedding: List[float], text: str) -> bool:
try:
async with httpx.AsyncClient(timeout=30.0) as client:
response = await client.post(
f"{COSMO_AI_URL}/api/feed/vision",
json={
"embedding": embedding,
"text_representation": text,
"source": "image-encoder",
},
)
return response.status_code == 200
except Exception as exc:
logger.error("Failed to send image embedding to cosmo-ai: {}", exc)
return False
def _service_url() -> str:
service_url = os.getenv("SPACE_HOST", "http://localhost:7860")
if not service_url.startswith("http"):
service_url = f"https://{service_url}"
return service_url.rstrip("/")
async def keepalive_loop():
await asyncio.sleep(120)
service_url = _service_url()
logger.info(
"Image encoder keepalive enabled; pinging {} every {}s",
service_url,
IMAGE_ENCODER_KEEPALIVE_INTERVAL_SECONDS,
)
while True:
try:
await asyncio.sleep(IMAGE_ENCODER_KEEPALIVE_INTERVAL_SECONDS)
async with httpx.AsyncClient(timeout=10.0) as client:
response = await client.get(f"{service_url}/ping")
if response.status_code == 200:
logger.info("Image encoder keepalive ping succeeded")
else:
logger.warning("Image encoder keepalive ping returned {}", response.status_code)
except asyncio.CancelledError:
raise
except Exception as exc:
logger.error("Image encoder keepalive failed: {}", exc)
@app.on_event("startup")
async def startup():
logger.info(
"Image Encoder Service starting model={} device={} preload={} keepalive={}",
IMAGE_ENCODER_MODEL_ID,
IMAGE_ENCODER_DEVICE,
IMAGE_ENCODER_PRELOAD,
IMAGE_ENCODER_KEEPALIVE,
)
if IMAGE_ENCODER_PRELOAD and CLIP_AVAILABLE:
load_clip_model()
if IMAGE_ENCODER_KEEPALIVE:
asyncio.create_task(keepalive_loop())
else:
logger.info("Image encoder keepalive disabled by configuration")
@app.get("/")
async def root():
return {
"service": "image-encoder",
"version": "1.1.0",
"status": "running",
"model": IMAGE_ENCODER_MODEL_ID if CLIP_AVAILABLE else "unavailable",
"backend": _clip_backend,
"clip_available": CLIP_AVAILABLE,
"model_loaded": _clip_model is not None,
"device": IMAGE_ENCODER_DEVICE,
"threads": IMAGE_ENCODER_THREADS,
"max_image_dim": IMAGE_ENCODER_MAX_IMAGE_DIM,
"dynamic_quantization": IMAGE_ENCODER_QUANTIZE and IMAGE_ENCODER_DEVICE == "cpu",
"cosmo_ai": COSMO_AI_URL,
}
@app.get("/health")
async def health():
return {
"status": "healthy",
"clip_available": CLIP_AVAILABLE,
"clip_backend": _clip_backend,
"model_loaded": _clip_model is not None,
"model_id": IMAGE_ENCODER_MODEL_ID,
"device": IMAGE_ENCODER_DEVICE,
"threads": IMAGE_ENCODER_THREADS,
"keepalive_enabled": IMAGE_ENCODER_KEEPALIVE,
}
@app.post("/encode", response_model=EncodeResponse)
async def encode_endpoint(request: EncodeRequest, background_tasks: BackgroundTasks):
try:
image_data = base64.b64decode(request.image_base64)
image = Image.open(io.BytesIO(image_data))
result = encode_image(image)
sent = False
if request.send_to_cosmo:
background_tasks.add_task(
send_to_cosmo_ai,
result["embedding"],
result["text_representation"],
)
sent = True
return EncodeResponse(
embedding=result["embedding"],
text_representation=result["text_representation"],
dimension=result["dimension"],
sent_to_cosmo=sent,
)
except Exception as exc:
logger.error("Encoding failed: {}", exc)
raise HTTPException(status_code=500, detail=str(exc))
@app.post("/encode/upload")
async def encode_upload(file: UploadFile = File(...), send_to_cosmo: bool = False):
try:
contents = await file.read()
image = Image.open(io.BytesIO(contents))
result = encode_image(image)
if send_to_cosmo:
await send_to_cosmo_ai(result["embedding"], result["text_representation"])
return result
except Exception as exc:
logger.error("Upload encoding failed: {}", exc)
raise HTTPException(status_code=500, detail=str(exc))
@app.get("/ping")
async def ping():
return {"status": "alive", "service": "image-encoder"}
@app.post("/keepalive/trigger")
async def trigger_keepalive():
results = {"self": False, "cosmo_ai": False}
try:
async with httpx.AsyncClient(timeout=10.0) as client:
response = await client.get(f"{_service_url()}/ping")
results["self"] = response.status_code == 200
response = await client.get(f"{COSMO_AI_URL}/api/ping")
results["cosmo_ai"] = response.status_code == 200
except Exception as exc:
logger.error("Keepalive trigger failed: {}", exc)
return {
"triggered": True,
"results": results,
"message": "Keepalive pings sent",
}
|