File size: 11,697 Bytes
42d88ae | 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 | import json
import time
import cv2
import numpy as np
import google.generativeai as genai
from pathlib import Path
from typing import Optional
from uuid import UUID
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession
from Backend.core.config import settings
from Backend.core.events import event_bus, IssueClassified, IssueCreated
from Backend.core.logging import get_logger
from Backend.core.schemas import ClassificationResult, DetectionBox, CLASS_ID_TO_CATEGORY, IssueCategory
from Backend.database.models import Classification, Issue, IssueImage, IssueEvent
from Backend.orchestration.base import BaseAgent
from Backend.utils.fuzzy_match import auto_validate_issue
from Backend.utils.storage import save_bytes, download_from_supabase, get_upload_url
logger = get_logger(__name__, agent_name="VisionAgent")
if settings.gemini_api_key:
genai.configure(api_key=settings.gemini_api_key)
class VisionAgent(BaseAgent):
_model = None
def __init__(self, db: Optional[AsyncSession] = None):
super().__init__("VisionAgent")
self.db = db
if settings.gemini_api_key:
self.gemini_model = genai.GenerativeModel('gemma-3-27b-it')
else:
self.gemini_model = None
@classmethod
def load_model(cls):
if cls._model is None:
from ultralytics import YOLO
model_path = settings.model_path
if not model_path.exists():
raise FileNotFoundError(f"Model not found: {model_path}")
cls._model = YOLO(str(model_path))
logger.info(f"YOLO model loaded from {model_path}")
return cls._model
@classmethod
def get_model(cls):
if cls._model is None:
cls.load_model()
return cls._model
async def download_image(self, remote_path: str) -> bytes:
return await download_from_supabase(remote_path)
async def save_annotated(self, results, original_path: str, subfolder: str) -> str:
im_array = results[0].plot()
original_name = Path(original_path).stem
annotated_filename = f"annotated_{original_name}.jpg"
_, buffer = cv2.imencode('.jpg', im_array, [cv2.IMWRITE_JPEG_QUALITY, 90])
image_bytes = buffer.tobytes()
remote_path = await save_bytes(image_bytes, annotated_filename, subfolder=subfolder)
return remote_path
async def run_inference(self, image_data: bytes) -> tuple[list, float]:
model = self.get_model()
nparr = np.frombuffer(image_data, np.uint8)
img = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
if img is None:
raise ValueError("Invalid image data")
start_time = time.perf_counter()
results = model.predict(
source=img,
conf=settings.model_confidence_threshold,
imgsz=settings.model_input_size,
verbose=False,
)
inference_time = (time.perf_counter() - start_time) * 1000
return results, inference_time
async def gemini_classify_image(
self,
image_data: bytes,
description: Optional[str] = None
) -> tuple[Optional[IssueCategory], float, Optional[str]]:
if not self.gemini_model:
return None, 0.0, None
allowed = [
{"class_id": k, "class_name": v.value}
for k, v in CLASS_ID_TO_CATEGORY.items()
]
prompt = (
"Classify the photo into exactly one of the allowed categories. "
"Return ONLY valid JSON with keys: class_id (int), confidence (0.0-1.0), reasoning (max 80 chars).\n\n"
f"Allowed categories: {json.dumps(allowed)}\n"
f"User description: {(description or '')[:200]}"
)
try:
response = self.gemini_model.generate_content(
[
{"text": prompt},
{
"inline_data": {
"mime_type": "image/jpeg",
"data": image_data,
}
},
]
)
text = (response.text or "").replace("```json", "").replace("```", "").strip()
data = json.loads(text)
class_id = data.get("class_id")
confidence = float(data.get("confidence", 0.0))
reasoning = data.get("reasoning")
if not isinstance(class_id, int):
return None, 0.0, None
category = CLASS_ID_TO_CATEGORY.get(class_id)
if not category:
return None, 0.0, None
confidence = max(0.0, min(1.0, confidence))
return category, confidence, reasoning
except Exception as e:
logger.error(f"Gemini vision classification failed: {e}")
return None, 0.0, None
def extract_detections(self, results) -> list[DetectionBox]:
detections = []
for result in results:
boxes = result.boxes
if boxes is not None:
for i in range(len(boxes)):
class_id = int(boxes.cls[i].item())
confidence = float(boxes.conf[i].item())
bbox = tuple(boxes.xyxy[i].tolist())
category = CLASS_ID_TO_CATEGORY.get(class_id)
if category:
detections.append(DetectionBox(
class_id=class_id,
class_name=category.value,
confidence=confidence,
bbox=bbox,
))
return detections
async def classify_image(
self,
image_path: str,
subfolder: str = "",
description: Optional[str] = None
) -> tuple[list[DetectionBox], str, Optional[IssueCategory], float, Optional[str]]:
image_data = await self.download_image(image_path)
results, inference_time = await self.run_inference(image_data)
annotated_path = await self.save_annotated(results, image_path, subfolder)
detections = self.extract_detections(results)
gemini_category = None
gemini_confidence = 0.0
gemini_reasoning = None
if self.gemini_model and (not detections or max(d.confidence for d in detections) < 0.5):
gemini_category, gemini_confidence, gemini_reasoning = await self.gemini_classify_image(
image_data=image_data,
description=description
)
logger.info(f"Inference completed in {inference_time:.2f}ms, {len(detections)} detections")
return detections, annotated_path, gemini_category, gemini_confidence, gemini_reasoning
async def process_issue(
self,
issue_id: UUID,
image_paths: list[str],
description: Optional[str] = None
) -> ClassificationResult:
all_detections = []
annotated_paths = []
total_time = 0.0
subfolder = str(issue_id)
gemini_best_category = None
gemini_best_confidence = 0.0
gemini_best_reasoning = None
for path in image_paths:
start = time.perf_counter()
detections, annotated_path, gemini_category, gemini_confidence, gemini_reasoning = await self.classify_image(
path,
subfolder=subfolder,
description=description
)
total_time += (time.perf_counter() - start) * 1000
all_detections.extend(detections)
annotated_paths.append(annotated_path)
if gemini_category and gemini_confidence > gemini_best_confidence:
gemini_best_category = gemini_category
gemini_best_confidence = gemini_confidence
gemini_best_reasoning = gemini_reasoning
if self.db:
query = select(IssueImage).where(IssueImage.file_path == path)
result = await self.db.execute(query)
image_record = result.scalar_one_or_none()
if image_record:
image_record.annotated_path = annotated_path
result = ClassificationResult(
issue_id=issue_id,
detections=all_detections,
annotated_urls=[get_upload_url(p) for p in annotated_paths],
inference_time_ms=total_time,
)
if gemini_best_category and (not result.primary_category or result.primary_confidence < 0.5):
result.primary_category = gemini_best_category
result.primary_confidence = gemini_best_confidence
detected_categories = list(set(d.class_name for d in all_detections))
auto_validated, validation_reason = auto_validate_issue(description, detected_categories)
validation_source = "auto" if auto_validated else "pending_manual"
new_state = "validated" if auto_validated else "reported"
self.log_decision(
issue_id=issue_id,
decision=f"Validation: {validation_source}",
reasoning=validation_reason
)
if self.db:
classification = Classification(
issue_id=issue_id,
primary_category=result.primary_category.value if result.primary_category else None,
primary_confidence=result.primary_confidence,
detections_json=json.dumps([d.model_dump() for d in all_detections]),
inference_time_ms=total_time,
)
self.db.add(classification)
issue = await self.db.get(Issue, issue_id)
if issue:
issue.state = new_state
issue.validation_source = validation_source
issue.validation_reason = validation_reason
event_record = IssueEvent(
issue_id=issue_id,
event_type="classified",
agent_name=self.name,
event_data=json.dumps({
"category": result.primary_category.value if result.primary_category else None,
"confidence": result.primary_confidence,
"detections_count": len(all_detections),
"validation_source": validation_source,
"annotated_images": annotated_paths,
"gemini_category": gemini_best_category.value if gemini_best_category else None,
"gemini_confidence": gemini_best_confidence,
"gemini_reasoning": gemini_best_reasoning,
})
)
self.db.add(event_record)
await self.db.flush()
if result.primary_category:
event = IssueClassified(
issue_id=issue_id,
category=result.primary_category.value,
confidence=result.primary_confidence,
detections_count=len(all_detections),
metadata={
"validation_source": validation_source,
"validation_reason": validation_reason,
"annotated_images": [get_upload_url(p) for p in annotated_paths],
}
)
await event_bus.publish(event)
return result
async def handle(self, event: IssueCreated) -> None:
await self.process_issue(
event.issue_id,
event.image_paths,
event.description
)
|