Mismatch in reported testing accuracy and actual testing accuracy of the model on the sutd-traffic dataset

[bennycortese]

In the paper associated with the model and the pre-trained checkpoint, reported testing accuracy is 46% but when you run the training loop on the model from this repository, testing accuracy by default on the sutd-traffic dataset is 45.1%. I am not sure why this discrepancy occurs, thank you!

[CHENGY12]

Thanks for your questions! It seems because of the environmental randomness. I just re-trained using another independent device and obtained 45.7% accuracy. We have released the checkpoint that was used in the paper here. In light of your suggestion, we will re-run the code 5 times and report an average performance and coresspooning std as clarification. Hope it can help.