Lower resolution for faster inference?

[antithing]

Hi, and thanks for making this code available!

Is it possible to run at low resolution, eg 512x512, and reach faster inference speeds?

I am aiming for 20ms per image.

[leon0514]

I don't know how to do it, I tried to export 512x512 onnx model but it failed. But it is possible to export onnx models with dynamic width and height, before I had problems reasoning about models with dynamic width and height with onnxruntime reasoning about them @antithing

[leon0514]

你可以看看这个pr apple/ml-depth-pro#45 ，里面有一个脚本convert_to_coreml.py，我成功了一部分，但是我改动的太多了，导致onnx的结构乱了。现在也没有精力弄了 @antithing

[leon0514]

我成功改成了768 x 768，3090时间在60ms，但是结果完全不对

[luvwinnie]

How can I convert it to smaller image...? i want to test it on jeston agx orin 64GB for faster inference

[raphael-yona]

Hello @luvwinnie , I would like to test the model on a jetson orin too. Have you found anything yet ? Thank you!