Is it possible to parser/convert model with dynamic shape [-1, 3, -1, -1]?

General Discussion
, , ,
1

For OCR detection and recognition model, the model will be applied in various/uncertain scenerios which users is uncertain. So it is necessary for the OCR model (similiar to ppocr) with dynamic shape. And the ppocr onnx model is dynamic shape too. But it seems that `hailo parser` in the 1st step only supports the certain shape (fixed W and H).
Is there any other ways to allow this kind of dynamic shape for OCR to be converted in Hailo NPU?

About dynamic shape, it is similar to this issue: Can not parse my onnx model which trained used Faster-RCNN - #4 by omria

Thanks in advanced.

2

Hey @Kevin_Chen,

We already have an OCR model and OCR application available, and we’ll be adding one to the hailo-apps repository soon! You can check out the current implementation here: Hailo-Application-Code-Examples/runtime/hailo-8/python/paddle_ocr at main · hailo-ai/Hailo-Application-Code-Examples · GitHub

Regarding dynamic shapes: The Hailo Dataflow Compiler (Hailo DFC) currently does not support dynamic input shapes during the parsing phase. Only fixed input dimensions are supported at this time."

Hope this helps!

3

@omria Get it. Dynamic shape not supported. It means that most of the OCR model cannot be flexibly deployed with Hailo. It is a pity.
I’ve seen this demo before. It is a fixed input shape model which cannot be applied in uncertain shape scenerios.
Thanks anyway.

4

Hi @Kevin_Chen
Hailo does not support dynamic resolution for the network.
But you can definitely use it in an application with different input sizes. You’ll just have to resize the image before sending it to hailo. Do an aspect ratio preserving resize.
We will release an OCR example soon.
I can assure you it is possible :grinning_face_with_smiling_eyes: