Not up to the official performance.

[hellohaley]

Hello, I use your code for training, but the final added performance is only 86.00, which is far from the paper data (89.86) and lower than the value provided by github (91.36). All training steps are run according to the readme file, using lm data set and Tesla K80.

Can you tell me how to improve the performance until the add is the same as that provided by the paper data or github?

[LZGMatrix]

Hi, can you provide more training information such as your training log files? Since we have provided the log files of the training process, it can help you to find the detailed difference in the training by comparing the logs infos.

[hellohaley]

Hello, by comparing the log files, I found that I only used the model (i.e., model_cpu.pth) of the last epoch in the previous stage, instead of the epoch with the highest ADD. I wonder if this is the main reason for the gap.