Infer with one single image question

[Datou98]

I use the model_final_v2_ade20k.pth that you provided, using the command:

python tools/export_onnx.py --config configs/bisenetv2_city.py --weight-path /home/datou/BiSeNet/BiSeNet-master/model/model_final_v2_city.pth --outpath ./model_v2_city.onnx --aux-mode eval

Convert pth to onnx and use the command： ./segment complie ../../model_v2_city.onnx ../../saved_v2city.trt --f16

Finally get trt, use the following command to test： ./segment run ../../saved_v2city.trt ../../example.png ../../output_v2city.jpg

But the output result is not correct, the result is as follows, I want to ask what is the reason?

[CoinCheung]

Again, I observe no such problem on my own platform. What model are you using? Is the pytorch demo.py works correctly?

[Datou98]

Again, I observe no such problem on my own platform. What model are you using? Is the pytorch demo.py works correctly?

yes,I am using the model model_final_v2_city.pth provided by your github, which can run normally in model.py when i commend ./segment run ../../saved_v2_city.trt ../../example.png ../../output_v2city.jpg

this warning :TensorRT was linked against cuBLAS/cuBLASLt 11.6.5 but loaded cuBLAS/cuBLASLt 11.4.2 Will it affect the result? do you have this problem

[CoinCheung]

I pulled a new repo and run step by step, but I still cannot reproduce this problem.

[Datou98]

I pulled a new repo and run step by step, but I still cannot reproduce this problem.  I just dropped the tensorrt version from 8.2.5.1 to 8.0.1.6, and the result of the segmentation can be displayed normally.

I tested it, the gpu is 3060, the fps is 25.6845, is this normal?

[CoinCheung]

I am not familiar with 3060, and since this thread is not activate for a long time(my bad), let me close this. If you still want more discussion, please open new issues.