How to convert a segmentation model to tensorrt engine that can output probability map ?

[yuanyuangoo]

Hi, Now the tensorrt engine can output result that has been passed through argmax function, how to also get output of the seg_logit before that?

[RunningLeon]

@yuanyuangoo Hi, a simple way is to return the seg_logit in the rewritings function, just like here.

[yuanyuangoo]

Hi, Thanks. It works to get seg_logit. Then How can I modify csrc/mmdeploy/codebase/mmseg/segment.cpp to accommodate the new output size? It gives me error like "[2022-10-10 08:50:36.869] [mmdeploy] [error] [segment.cpp:35] unsupported output tensor, shape: [1, 8, 512, 1024]"

[RunningLeon]

@yuanyuangoo Hi, if you want to support this in sdk, you have to add another post_process with resize and argmax.

[AllentDan]

#1379 might be what you need. @yuanyuangoo