Performance of ASPP module and ONNX conversion feasibility
Closed this issue · 2 comments
I'm truly grateful and amazed by your model. However, I have a question. As mentioned in this issue (#26), the deform_conv2d is not working well.
According to the ONNX documentation, deform_conv2d seems to be supported from version 22 onwards.
So, my question is: In your code, there is a line that says 'config.dec_att == 'ASPP''. Do you know how well this ASPP module performs? I couldn't find any ablation or comparison experiments in the paper where something other than Deconv was used.
If the performance drop with ASPP is not significant, I would like to try converting the model to ONNX.
It could be not too much degradation in performance. Those with deform_conv2d should work a little better, but I don't know how much it is. However, I still personally suggest keeping the deform_conv2d. If you have the computation resources, you can have a try on training it.
Thank you for the response. I will give it a try