Parallel calls to '.ocr' function in paddle_ocr result in OOM even when the ocr object is initialized once.
Code to replicate:
paddle_ocr = PaddleOCR(lang='en',use_gpu=False)
def do_ocr(image):
paddle_lines = paddle_ocr.ocr(image)
paddle_lines = [line[1][0] for line in paddle_lines]
print(paddle_lines)
//Called via endpoint
image=
do_ocr(image)
Here the memory size increases linearly with the no of parallel http calls even though paddle_ocr was initialized at the module level.
version: tested on both 2.1.0 and 2.2.0
2条答案
按热度按时间6tr1vspr1#
Sorry I can not reproduce the results you mention above, could you provide a
pythonscript that can reproduce your problem?qpgpyjmq2#
I am using the code via a fastAPI endpoint and as such cannot copy the exact code. However I am able to duplicate it using the threadpoolexecutor framework in python as follows:
from paddleocr import PaddleOCR
import imageio
paddle_ocr = PaddleOCR(lang='en', gpu=False, use_gpu=False)
from concurrent.futures import ThreadPoolExecutor as Pool
executors = Pool(40)
def do_ocr(image):
paddle_lines = paddle_ocr.ocr(image)
paddle_lines = [line[1][0] for line in paddle_lines]
print(paddle_lines)
load any random image
image = imageio.imread(' https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.3/doc/imgs_results/multi_lang/img_01.jpg ')
jobs = [image for i in range(500)]
executors.map(do_ocr,jobs)