Models without training are better than the trained one?

[LonglongaaaGo]

Hi @Haoyan-Guan ,

Thank you so much for your awesome work.
Recently, I tried your repo on my customer data; I found the interesting thing is that the model loaded from pre-trained ResNet18 and, without any training, can achieve the best performance. The more training, the worse the performance of the model.

Did you encounter a similar problem? I was wondering what the problem was behind that.

Thanks!