Skip to content

Commit

Permalink
Update docs
Browse files Browse the repository at this point in the history
  • Loading branch information
SWHL committed Jan 30, 2024
1 parent a263cb4 commit 2c649c7
Showing 1 changed file with 3 additions and 1 deletion.
4 changes: 3 additions & 1 deletion content/docs/faq.md
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,9 @@ toc: true
---

#### Q: 为什么我的模型在ONNXRuntime GPU版上比在CPU上还要慢?
**A:** : 因为OCR任务中输入图像Shape是动态的。每次GPU上都需要重新清空上一次不同Shape的缓存结果。如果输入图像Shape不变的情况下,ONNXRuntime GPU版一般都要比CPU快的。该问题已经提了相关issue([issue #13198](https://github.com/microsoft/onnxruntime/issues/13198))。推荐CPU端推理用`rapidocr_onnxruntime`或者`rapidocr_openvino`,GPU端用`rapidocr_paddle`
**A:** 因为OCR任务中输入图像Shape是动态的。每次GPU上都需要重新清空上一次不同Shape的缓存结果。如果输入图像Shape不变的情况下,ONNXRuntime GPU版一般都要比CPU快的。该问题已经提了相关issue([issue #13198](https://github.com/microsoft/onnxruntime/issues/13198))。

推荐CPU端推理用`rapidocr_onnxruntime`或者`rapidocr_openvino`,GPU端用`rapidocr_paddle`。关于`rapidocr_onnxruntime``rapidocr_paddle`两者之间推理,可参见:[docs](./install_usage/rapidocr_paddle.md#推理速度比较)

#### Q: 请问这个能在32位C#中用嘛?
**A:** C#可以32位,要用32位的dll,但nuget上的onnxruntime不支持win7。
Expand Down

0 comments on commit 2c649c7

Please sign in to comment.