cost more training time per epoch than centernet

[Libaishun]

Using single GPU, same network architecture, I found the training time per epoch of ttfnet is about 1.6 times than centernet.
I'm not sure whether I did the comparison right , has anyone made such comparison?

[enhenghengheng]

I also found the same problem. Using dla34 for training, the training time of centernet is faster than that of ttfnet, and the accuracy of centernet is also higher than that of ttfnet when training for 12 epochs.