Prediction results of pre-trained weights

[BenPinhasov]

Hello,

I use this implementation for mesoNet with pre-trained weights that you provided. When I test it on real video from FF++ dataset, the accuracy of the model is very low. It predict the videos (and the frames from the video) most of the time as fake. When I test in on fake video from the same dataset the accuracy near to accuracy that declared in the original paper. When I combined the real and fake datasets the accuracy 50% I did something wrong with my tests?