我使用onnxruntime对齐官方的推理，推理速度比使用paddlepaddle框架推理快很多（速度快5，6倍） #13195
-
问题描述 / Problem Description我使用onnxruntime对齐官方的推理，推理速度比使用paddlepaddle框架推理快很多（速度快5，6倍），是框架有问题吗 运行环境 / Runtime Environment
复现代码 / Reproduction Code代码：https://github.com/jingsongliujing/OnnxOCR.git 完整报错 / Complete Error Message可能解决方案 / Possible solutions附件 / Appendix |
Beta Was this translation helpful? Give feedback.
Replies: 9 comments 3 replies
-
|
是比 paddle inference 快吗 |
Beta Was this translation helpful? Give feedback.
-
是的，要快很多 |
Beta Was this translation helpful? Give feedback.
-
|
特别是v4版本 |
Beta Was this translation helpful? Give feedback.
-
|
其他几个模型不明显，就v4 mobile 比较明显 |
Beta Was this translation helpful? Give feedback.
-
|
用 |
Beta Was this translation helpful? Give feedback.
-
|
没试过，但是官方框架它本身会加载很多配置，估计推理速度也会比自己拆出来慢一些 |
Beta Was this translation helpful? Give feedback.
-
|
我目前只做了检测，分类，识别三个模型的串联，表格，抽取啥的没考虑 |
Beta Was this translation helpful? Give feedback.
-
|
请问效果和资源占用上有差异吗 |
Beta Was this translation helpful? Give feedback.
-
|
这和我们做的RapidOCR似乎是一样的 |
Beta Was this translation helpful? Give feedback.
-
嗯，看了一下，确实差不多，但是我看仓库只更新到了v3的onnx版本，v4是否有更新呢 |
Beta Was this translation helpful? Give feedback.
-
|
有的，目前v1.3.22里面集成的就是v4模型。你看到的应该是目录名是v3的。 |
Beta Was this translation helpful? Give feedback.
-
|
onnx c++，gpu 服务端模型v4，推理速度 7ms 识别单行文本10个字，3060 |
Beta Was this translation helpful? Give feedback.
有的，目前v1.3.22里面集成的就是v4模型。你看到的应该是目录名是v3的。