Low MIOU for inference

[Hide-A-Pumpkin]

I saw you are keeping the dropout during the testing phase in test_pretrained_SemanticKITTI.py. When I trained the model, the results turned out well, reaching 75% MIOU. I went on to validate the model using test_pretrained_SemanticKITTI.py and found the MIOU went down to nearly 55%. Also, each inference generates a quite different answer; over 30% of labels were changed each time.

Is it supposed to be that drastic? Should I change the dropout rate to 0 when inferencing?