how to export onnx

[hywchina]

How to Convert a Trained .pth File to an .onnx File

thanks

[Wishrem]

Hi, I have done this work. I'm glad to share the exported .onnx file. Here is the download link powered by Google Drive.

The exported model have been tested by evaluate function in inference_rsvg.py, and the runtime arguments file is attached to this comment.

Result:

[email protected]: 0.8737	[email protected]: 0.8257	[email protected]: 0.7234	[email protected]: 0.5110	[email protected]: 0.1443	meanIoU: 0.7155	cumuIoU: 0.7006

The exported model ONLY accepts (due to static computing graph) the following inputs with specific names, shapes and dtypes:

img (1, 3, 1024, 1024) float32
img_attn_mask (1, 1024, 1024) bool
input_ids (1, dynamic shape) int64
input_ids_attn_mask (1, dynamic shape) int64

the outputs of this exported model are raw outputs_class and output_coord, which you can refer to models/LQVG.py:LQVG.forward.

args.txt