Closed windid closed 6 years ago
Thanks so much! I just find out the problem. The official TF implementation has a small difference against paper. I have modified my implementation accordingly and upload the new pretrained model. The statistics look identical now.
I've measured this implementation using this script: https://github.com/ShichenLiu/CondenseNet/blob/master/utils.py
and got the following result: FLOPs: 320.19M, Params: 3.51M
which is slightly higher than the paper. I just confusing about this.