Update app.py
Browse files
app.py
CHANGED
|
@@ -1,9 +1,12 @@
|
|
| 1 |
import os
|
| 2 |
import io
|
|
|
|
|
|
|
| 3 |
import numpy as np
|
| 4 |
from PIL import Image
|
| 5 |
from fastapi import FastAPI, UploadFile, File
|
| 6 |
from fastapi.responses import HTMLResponse
|
|
|
|
| 7 |
from huggingface_hub import hf_hub_download
|
| 8 |
|
| 9 |
# Ограничиваем треды ONNX Runtime (на HF Spaces CPU Basic всего 2 vCPU)
|
|
@@ -13,6 +16,12 @@ os.environ["MKL_NUM_THREADS"] = "2"
|
|
| 13 |
app = FastAPI()
|
| 14 |
MODEL_DIR = os.path.join(os.path.dirname(__file__), "models")
|
| 15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
def download_file(repo_id, filename, local_dir):
|
| 17 |
"""Скачивает файл с HF Hub, если его нет локально."""
|
| 18 |
local_path = os.path.join(local_dir, filename)
|
|
@@ -27,14 +36,24 @@ def download_file(repo_id, filename, local_dir):
|
|
| 27 |
print(f" Failed: {e}")
|
| 28 |
return None
|
| 29 |
|
|
|
|
| 30 |
def ensure_models():
|
| 31 |
-
"""
|
|
|
|
|
|
|
|
|
|
| 32 |
os.makedirs(MODEL_DIR, exist_ok=True)
|
| 33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
rec = download_file("monkt/paddleocr-onnx", "languages/eslav/rec.onnx", MODEL_DIR)
|
| 35 |
dict_path = download_file("monkt/paddleocr-onnx", "languages/eslav/dict.txt", MODEL_DIR)
|
| 36 |
return det, rec, dict_path
|
| 37 |
|
|
|
|
| 38 |
det_path, rec_path, dict_path = ensure_models()
|
| 39 |
|
| 40 |
from rapidocr_onnxruntime import RapidOCR
|
|
@@ -61,7 +80,10 @@ try:
|
|
| 61 |
ocr.text_detector.postprocess_op.thresh = 0.5
|
| 62 |
ocr.text_detector.postprocess_op.use_dilation = False
|
| 63 |
ocr.text_detector.postprocess_op.score_mode = "fast"
|
| 64 |
-
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
if hasattr(ocr, "text_recognizer"):
|
| 67 |
ocr.text_recognizer.rec_batch_num = 1
|
|
@@ -71,7 +93,6 @@ except TypeError:
|
|
| 71 |
# === Попытка 2: старая версия, нужен полный config.yaml ===
|
| 72 |
print("kwargs not supported, using full config.yaml")
|
| 73 |
|
| 74 |
-
# Скачиваем cls модель из нескольких возможных источников
|
| 75 |
cls_path = None
|
| 76 |
for repo, path in [
|
| 77 |
("RapidAI/RapidOCR", "onnx/PP-OCRv4/cls/ch_ppocr_mobile_v2.0_cls_infer.onnx"),
|
|
@@ -82,7 +103,6 @@ except TypeError:
|
|
| 82 |
if cls_path:
|
| 83 |
break
|
| 84 |
|
| 85 |
-
# Если не нашли — запускаем стандартный RapidOCR, чтобы он скачал модели в кэш
|
| 86 |
if not cls_path:
|
| 87 |
print("Downloading standard models via RapidOCR()...")
|
| 88 |
temp_ocr = RapidOCR()
|
|
@@ -95,7 +115,6 @@ except TypeError:
|
|
| 95 |
break
|
| 96 |
if cls_path:
|
| 97 |
break
|
| 98 |
-
# Ищем в кэше
|
| 99 |
if not cls_path:
|
| 100 |
cache_dir = os.path.expanduser("~/.cache/rapidocr_onnxruntime")
|
| 101 |
if os.path.exists(cache_dir):
|
|
@@ -107,7 +126,6 @@ except TypeError:
|
|
| 107 |
if cls_path:
|
| 108 |
break
|
| 109 |
|
| 110 |
-
# Создаём полный config.yaml со ВСЕМИ обязательными полями
|
| 111 |
config_path = os.path.join(os.path.dirname(__file__), "config.yaml")
|
| 112 |
with open(config_path, "w") as f:
|
| 113 |
f.write(f"""
|
|
@@ -138,7 +156,7 @@ Det:
|
|
| 138 |
post_process:
|
| 139 |
thresh: 0.5
|
| 140 |
box_thresh: 0.5
|
| 141 |
-
max_candidates:
|
| 142 |
unclip_ratio: 1.6
|
| 143 |
use_dilation: false
|
| 144 |
score_mode: fast
|
|
@@ -214,10 +232,21 @@ HTML_FORM = """<!DOCTYPE html>
|
|
| 214 |
</body>
|
| 215 |
</html>"""
|
| 216 |
|
|
|
|
| 217 |
@app.get("/", response_class=HTMLResponse)
|
| 218 |
def read_root():
|
| 219 |
return HTML_FORM
|
| 220 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 221 |
@app.post("/predict")
|
| 222 |
async def predict(file: UploadFile = File(...)):
|
| 223 |
try:
|
|
@@ -228,7 +257,10 @@ async def predict(file: UploadFile = File(...)):
|
|
| 228 |
if image.width > 480 or image.height > 480:
|
| 229 |
image.thumbnail((480, 480))
|
| 230 |
img_array = np.array(image)
|
| 231 |
-
|
|
|
|
|
|
|
|
|
|
| 232 |
if not results:
|
| 233 |
return {"text": ""}
|
| 234 |
text = " ".join([item[1] for item in results]).strip()
|
|
|
|
| 1 |
import os
|
| 2 |
import io
|
| 3 |
+
import time
|
| 4 |
+
import asyncio
|
| 5 |
import numpy as np
|
| 6 |
from PIL import Image
|
| 7 |
from fastapi import FastAPI, UploadFile, File
|
| 8 |
from fastapi.responses import HTMLResponse
|
| 9 |
+
from starlette.concurrency import run_in_threadpool
|
| 10 |
from huggingface_hub import hf_hub_download
|
| 11 |
|
| 12 |
# Ограничиваем треды ONNX Runtime (на HF Spaces CPU Basic всего 2 vCPU)
|
|
|
|
| 16 |
app = FastAPI()
|
| 17 |
MODEL_DIR = os.path.join(os.path.dirname(__file__), "models")
|
| 18 |
|
| 19 |
+
# Ограничиваем реальную конкурентность OCR-вызовов под число vCPU.
|
| 20 |
+
# На 2 vCPU параллельный запуск 2+ OCR одновременно делит ядра между ними
|
| 21 |
+
# и может быть МЕДЛЕННЕЕ, чем строгая очередь — поэтому 1, не 2.
|
| 22 |
+
ocr_semaphore = asyncio.Semaphore(1)
|
| 23 |
+
|
| 24 |
+
|
| 25 |
def download_file(repo_id, filename, local_dir):
|
| 26 |
"""Скачивает файл с HF Hub, если его нет локально."""
|
| 27 |
local_path = os.path.join(local_dir, filename)
|
|
|
|
| 36 |
print(f" Failed: {e}")
|
| 37 |
return None
|
| 38 |
|
| 39 |
+
|
| 40 |
def ensure_models():
|
| 41 |
+
"""
|
| 42 |
+
Скачивает det (mobile, лёгкий ~4.8 МБ вместо server-варианта ~88 МБ),
|
| 43 |
+
rec и dict для русского+английского (eslav покрывает кириллицу).
|
| 44 |
+
"""
|
| 45 |
os.makedirs(MODEL_DIR, exist_ok=True)
|
| 46 |
+
|
| 47 |
+
# Детекция language-agnostic — просто находит текстовые блоки,
|
| 48 |
+
# поэтому mobile-вариант той же PP-OCRv5 серии работает для любого языка.
|
| 49 |
+
det = download_file("ilaylow/PP_OCRv5_mobile_onnx", "ppocrv5_det.onnx", MODEL_DIR)
|
| 50 |
+
|
| 51 |
+
# Рекогнишн оставляем eslav — обучена именно под кириллицу + латиницу
|
| 52 |
rec = download_file("monkt/paddleocr-onnx", "languages/eslav/rec.onnx", MODEL_DIR)
|
| 53 |
dict_path = download_file("monkt/paddleocr-onnx", "languages/eslav/dict.txt", MODEL_DIR)
|
| 54 |
return det, rec, dict_path
|
| 55 |
|
| 56 |
+
|
| 57 |
det_path, rec_path, dict_path = ensure_models()
|
| 58 |
|
| 59 |
from rapidocr_onnxruntime import RapidOCR
|
|
|
|
| 80 |
ocr.text_detector.postprocess_op.thresh = 0.5
|
| 81 |
ocr.text_detector.postprocess_op.use_dilation = False
|
| 82 |
ocr.text_detector.postprocess_op.score_mode = "fast"
|
| 83 |
+
# Меньше кандидатов для NMS — на скрине редко >10-15 текстовых блоков
|
| 84 |
+
if hasattr(ocr.text_detector.postprocess_op, "max_candidates"):
|
| 85 |
+
ocr.text_detector.postprocess_op.max_candidates = 100
|
| 86 |
+
print(" Patched: thresh=0.5, use_dilation=False, score_mode=fast, max_candidates=100")
|
| 87 |
|
| 88 |
if hasattr(ocr, "text_recognizer"):
|
| 89 |
ocr.text_recognizer.rec_batch_num = 1
|
|
|
|
| 93 |
# === Попытка 2: старая версия, нужен полный config.yaml ===
|
| 94 |
print("kwargs not supported, using full config.yaml")
|
| 95 |
|
|
|
|
| 96 |
cls_path = None
|
| 97 |
for repo, path in [
|
| 98 |
("RapidAI/RapidOCR", "onnx/PP-OCRv4/cls/ch_ppocr_mobile_v2.0_cls_infer.onnx"),
|
|
|
|
| 103 |
if cls_path:
|
| 104 |
break
|
| 105 |
|
|
|
|
| 106 |
if not cls_path:
|
| 107 |
print("Downloading standard models via RapidOCR()...")
|
| 108 |
temp_ocr = RapidOCR()
|
|
|
|
| 115 |
break
|
| 116 |
if cls_path:
|
| 117 |
break
|
|
|
|
| 118 |
if not cls_path:
|
| 119 |
cache_dir = os.path.expanduser("~/.cache/rapidocr_onnxruntime")
|
| 120 |
if os.path.exists(cache_dir):
|
|
|
|
| 126 |
if cls_path:
|
| 127 |
break
|
| 128 |
|
|
|
|
| 129 |
config_path = os.path.join(os.path.dirname(__file__), "config.yaml")
|
| 130 |
with open(config_path, "w") as f:
|
| 131 |
f.write(f"""
|
|
|
|
| 156 |
post_process:
|
| 157 |
thresh: 0.5
|
| 158 |
box_thresh: 0.5
|
| 159 |
+
max_candidates: 100
|
| 160 |
unclip_ratio: 1.6
|
| 161 |
use_dilation: false
|
| 162 |
score_mode: fast
|
|
|
|
| 232 |
</body>
|
| 233 |
</html>"""
|
| 234 |
|
| 235 |
+
|
| 236 |
@app.get("/", response_class=HTMLResponse)
|
| 237 |
def read_root():
|
| 238 |
return HTML_FORM
|
| 239 |
|
| 240 |
+
|
| 241 |
+
def _run_ocr_sync(img_array):
|
| 242 |
+
"""Синхронный вызов OCR — выполняет��я в threadpool, не блокирует event loop."""
|
| 243 |
+
t0 = time.time()
|
| 244 |
+
results, elapse_list = ocr(img_array)
|
| 245 |
+
total = time.time() - t0
|
| 246 |
+
print(f"[TIMING] total={total:.2f}s breakdown(det,cls,rec)={elapse_list}")
|
| 247 |
+
return results
|
| 248 |
+
|
| 249 |
+
|
| 250 |
@app.post("/predict")
|
| 251 |
async def predict(file: UploadFile = File(...)):
|
| 252 |
try:
|
|
|
|
| 257 |
if image.width > 480 or image.height > 480:
|
| 258 |
image.thumbnail((480, 480))
|
| 259 |
img_array = np.array(image)
|
| 260 |
+
|
| 261 |
+
async with ocr_semaphore:
|
| 262 |
+
results = await run_in_threadpool(_run_ocr_sync, img_array)
|
| 263 |
+
|
| 264 |
if not results:
|
| 265 |
return {"text": ""}
|
| 266 |
text = " ".join([item[1] for item in results]).strip()
|