Open SWHL opened 2 months ago
你好,我实际测试下来,发现CPU上rapidocr_onnxruntime性能比rapidocr_paddle好,但是这个差异主要是默认的cpu_math_library_num_threads没有设置导致的。具体原因是:以下这个分支里的设置线程数没有跑进去,导致paddle CPU推理的默认性能很差。
cpu_nums = os.cpu_count() infer_num_threads = config.get("cpu_math_library_num_threads", -1) if infer_num_threads != -1 and 1 <= infer_num_threads <= cpu_nums: infer_opts.set_cpu_math_library_num_threads(infer_num_threads)
希望把默认的infer_num_threads这块改一下。
Originally posted by @vivienfanghuagood in https://github.com/RapidAI/RapidOCR/discussions/196#discussioncomment-10405131
该问题查看源码,当前逻辑是:如果不显式指定infer_num_threads,则不通过inference.set_cpu_math_library_num_threads来设置,采用Config初始化的默认值。@vivienfanghuagood
infer_num_threads
inference.set_cpu_math_library_num_threads
你好,我实际测试下来,发现CPU上rapidocr_onnxruntime性能比rapidocr_paddle好,但是这个差异主要是默认的cpu_math_library_num_threads没有设置导致的。具体原因是:以下这个分支里的设置线程数没有跑进去,导致paddle CPU推理的默认性能很差。
希望把默认的infer_num_threads这块改一下。
Originally posted by @vivienfanghuagood in https://github.com/RapidAI/RapidOCR/discussions/196#discussioncomment-10405131