nan for mIOU

[niloofarAzari]

Executing the eval.py file using pre-trained model "res18_sfnet.pth", i get nan value for mIOU and also for the iou of all the classes. Although when looking at the output segmented images they seem really good. What may the reason be?

[lxtGH]

Could you show the detailed comand?

[TiesHendrickx]

I encountered the same issue. It seems like the loaded GT labels for the test set are all 255 (invalid). Running the eval script with "sliding" inference mode does not yield any results for me, therefore I used the following arguments: "--arch", "network.sfnet_resnet.DeepR18_SF_deeply", "--snapshot", "[...]/SFSegNets/pretrained_models/sfnet_r18_map.pth", "--cv_split", "0", "--split","test", "--dataset_dir","[...]/Cityscapes", "--inference_mode", "pooling"

[lxtGH]

I did not find this error during SFNet-lite pre-paration in these months. So I close this issue.