pspnet101 model pretrained on cityscapes from repository Google-Drive different test results #92
Description
Hello I wanted to evaluate the provided (from Google-Drive) pspnet-101 model pretrained (on cityscapes train set) on cityscapes val set with the provided code. But I got strange results from running test.sh:
System specs:
I used pytorch 1.8.2 with LTS support in conda environment with cuda 11.1 on Linux with Nvidia A100 GPU with Driver Version: 550.54.15 and CUDA Version: 12.4
I used the gray predictions created from test script to evaluate them (also an val set) on cityscapesScripts Github-Repository evalPixelLevelSemanticLabeling.py and I got an mIoU of 0.729 (ss) which is much lower than what the model should get by this repository (0.7863 mIoU).
The model seems to be a bit better when using 2049x1025 images yielding and mIoU of 0.7347 but still much worse.
Am I doing something wrong?