Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from fastapi import FastAPI, File, Form, UploadFile, HTTPException
|
| 2 |
+
from fastapi.responses import JSONResponse
|
| 3 |
+
from pydantic import BaseModel
|
| 4 |
+
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
|
| 5 |
+
from qwen_vl_utils import process_vision_info
|
| 6 |
+
from PIL import Image
|
| 7 |
+
import torch
|
| 8 |
+
import tempfile
|
| 9 |
+
import os
|
| 10 |
+
import logging
|
| 11 |
+
from datetime import datetime
|
| 12 |
+
|
| 13 |
+
# Configuration logging
|
| 14 |
+
logging.basicConfig(level=logging.INFO)
|
| 15 |
+
logger = logging.getLogger(__name__)
|
| 16 |
+
|
| 17 |
+
# Initialiser FastAPI
|
| 18 |
+
app = FastAPI(
|
| 19 |
+
title="Sparrow Qwen2-VL API",
|
| 20 |
+
description="API REST pour extraction de données depuis images via Qwen2-VL",
|
| 21 |
+
version="1.0.0"
|
| 22 |
+
)
|
| 23 |
+
|
| 24 |
+
# Charger le modèle au démarrage
|
| 25 |
+
logger.info("🔄 Chargement du modèle Qwen2-VL-7B-Instruct...")
|
| 26 |
+
try:
|
| 27 |
+
model = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 28 |
+
"Qwen/Qwen2-VL-7B-Instruct",
|
| 29 |
+
torch_dtype="auto",
|
| 30 |
+
device_map="auto"
|
| 31 |
+
)
|
| 32 |
+
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
|
| 33 |
+
logger.info("✅ Modèle chargé avec succès!")
|
| 34 |
+
except Exception as e:
|
| 35 |
+
logger.error(f"❌ Erreur chargement modèle: {e}")
|
| 36 |
+
raise
|
| 37 |
+
|
| 38 |
+
# Modèle de réponse
|
| 39 |
+
class ExtractionResponse(BaseModel):
|
| 40 |
+
result: str
|
| 41 |
+
status: str
|
| 42 |
+
timestamp: str
|
| 43 |
+
|
| 44 |
+
@app.post("/predict", response_model=ExtractionResponse)
|
| 45 |
+
async def predict(
|
| 46 |
+
image: UploadFile = File(..., description="Image à analyser"),
|
| 47 |
+
query: str = Form(..., description="Instruction d'extraction")
|
| 48 |
+
):
|
| 49 |
+
"""
|
| 50 |
+
Extraire des données d'une image selon la requête
|
| 51 |
+
"""
|
| 52 |
+
timestamp = datetime.now().isoformat()
|
| 53 |
+
temp_path = None
|
| 54 |
+
|
| 55 |
+
try:
|
| 56 |
+
# Validation du fichier
|
| 57 |
+
if not image.content_type.startswith('image/'):
|
| 58 |
+
raise HTTPException(status_code=400, detail="Fichier doit être une image")
|
| 59 |
+
|
| 60 |
+
# Sauvegarder temporairement
|
| 61 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmp_file:
|
| 62 |
+
content = await image.read()
|
| 63 |
+
tmp_file.write(content)
|
| 64 |
+
temp_path = tmp_file.name
|
| 65 |
+
|
| 66 |
+
logger.info(f"🖼️ Traitement image: {image.filename}")
|
| 67 |
+
logger.info(f"📝 Requête: {query}")
|
| 68 |
+
|
| 69 |
+
# Préparer l'image
|
| 70 |
+
img = Image.open(temp_path)
|
| 71 |
+
|
| 72 |
+
# Créer les messages pour le modèle
|
| 73 |
+
messages = [
|
| 74 |
+
{
|
| 75 |
+
"role": "user",
|
| 76 |
+
"content": [
|
| 77 |
+
{
|
| 78 |
+
"type": "image",
|
| 79 |
+
"image": temp_path
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"type": "text",
|
| 83 |
+
"text": query
|
| 84 |
+
}
|
| 85 |
+
]
|
| 86 |
+
}
|
| 87 |
+
]
|
| 88 |
+
|
| 89 |
+
# Appliquer le template de chat
|
| 90 |
+
text = processor.apply_chat_template(
|
| 91 |
+
messages, tokenize=False, add_generation_prompt=True
|
| 92 |
+
)
|
| 93 |
+
|
| 94 |
+
# Traiter les informations visuelles
|
| 95 |
+
image_inputs, video_inputs = process_vision_info(messages)
|
| 96 |
+
|
| 97 |
+
# Préparer les inputs
|
| 98 |
+
inputs = processor(
|
| 99 |
+
text=[text],
|
| 100 |
+
images=image_inputs,
|
| 101 |
+
videos=video_inputs,
|
| 102 |
+
padding=True,
|
| 103 |
+
return_tensors="pt",
|
| 104 |
+
)
|
| 105 |
+
|
| 106 |
+
# Déplacer sur le bon device
|
| 107 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 108 |
+
inputs = inputs.to(device)
|
| 109 |
+
|
| 110 |
+
# Générer la réponse
|
| 111 |
+
logger.info("🤖 Génération de la réponse...")
|
| 112 |
+
generated_ids = model.generate(**inputs, max_new_tokens=4096)
|
| 113 |
+
|
| 114 |
+
# Nettoyer les tokens
|
| 115 |
+
generated_ids_trimmed = [
|
| 116 |
+
out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
| 117 |
+
]
|
| 118 |
+
|
| 119 |
+
# Décoder le résultat
|
| 120 |
+
output = processor.batch_decode(
|
| 121 |
+
generated_ids_trimmed,
|
| 122 |
+
skip_special_tokens=True,
|
| 123 |
+
clean_up_tokenization_spaces=True
|
| 124 |
+
)[0]
|
| 125 |
+
|
| 126 |
+
logger.info(f"✅ Extraction réussie: {len(output)} caractères")
|
| 127 |
+
|
| 128 |
+
return ExtractionResponse(
|
| 129 |
+
result=output,
|
| 130 |
+
status="success",
|
| 131 |
+
timestamp=timestamp
|
| 132 |
+
)
|
| 133 |
+
|
| 134 |
+
except Exception as e:
|
| 135 |
+
logger.error(f"❌ Erreur traitement: {str(e)}")
|
| 136 |
+
raise HTTPException(status_code=500, detail=str(e))
|
| 137 |
+
|
| 138 |
+
finally:
|
| 139 |
+
# Nettoyer le fichier temporaire
|
| 140 |
+
if temp_path and os.path.exists(temp_path):
|
| 141 |
+
os.remove(temp_path)
|
| 142 |
+
logger.info("🧹 Fichier temporaire nettoyé")
|
| 143 |
+
|
| 144 |
+
@app.get("/health")
|
| 145 |
+
def health_check():
|
| 146 |
+
"""
|
| 147 |
+
Vérifier que l'API fonctionne
|
| 148 |
+
"""
|
| 149 |
+
return {
|
| 150 |
+
"status": "healthy",
|
| 151 |
+
"model": "Qwen2-VL-7B-Instruct",
|
| 152 |
+
"device": "cuda" if torch.cuda.is_available() else "cpu",
|
| 153 |
+
"timestamp": datetime.now().isoformat()
|
| 154 |
+
}
|
| 155 |
+
|
| 156 |
+
@app.get("/info")
|
| 157 |
+
def api_info():
|
| 158 |
+
"""
|
| 159 |
+
Informations sur l'API
|
| 160 |
+
"""
|
| 161 |
+
return {
|
| 162 |
+
"name": "Sparrow Qwen2-VL API",
|
| 163 |
+
"version": "1.0.0",
|
| 164 |
+
"endpoints": {
|
| 165 |
+
"predict": "/predict",
|
| 166 |
+
"health": "/health",
|
| 167 |
+
"info": "/info"
|
| 168 |
+
},
|
| 169 |
+
"model": "Qwen/Qwen2-VL-7B-Instruct"
|
| 170 |
+
}
|
| 171 |
+
|
| 172 |
+
# Pour compatibilité avec Gradio (optionnel)
|
| 173 |
+
@app.get("/")
|
| 174 |
+
def root():
|
| 175 |
+
return JSONResponse({
|
| 176 |
+
"message": "Sparrow Qwen2-VL API is running",
|
| 177 |
+
"docs": "/docs",
|
| 178 |
+
"health": "/health",
|
| 179 |
+
"predict": "/predict"
|
| 180 |
+
})
|
| 181 |
+
|
| 182 |
+
# Lancer le serveur si exécuté directement
|
| 183 |
+
if __name__ == "__main__":
|
| 184 |
+
import uvicorn
|
| 185 |
+
uvicorn.run(app, host="0.0.0.0", port=7860)
|