search

Lyken17 / SparseNet

[ECCV 2018] Sparsely Aggreagated Convolutional Networks https://arxiv.org/abs/1801.05895
MIT License
125 stars 25 forks source link

SparseNet depth_40_gr_12 validation error is 7.31% on cifar10 #7

Closed Sirius083 closed 5 years ago

Sirius083 commented 5 years ago

the tensorflow implementation on SparseNet depth=40 growth_rate=12 (no bottleneck) on cifar10 best validation error is 7.31% while the paper reported error is 5.47% what reason could probably cause the difference? (change nothing excepth for the training log directory) tensorflow version 1.12 tensorpack version 0.9.1 python version 3.6.6

Looking forward to your reply

Lyken17 commented 5 years ago

I think I know what is happening... This bug is related to certain CUDA and CUDNN version. Can you share your version?

Sirius083 commented 5 years ago

thanks for your reply, my cuda version: 9.0 cudnn version: 7.0

Sirius083 commented 5 years ago

I changed the model setting to depth_40_k_24 on cifar10, then the results (min-validation error is 0.566) is almost same as presented in the paper :blush:

Lyken17 commented 5 years ago

Happy to see you solve the problem. I will fix the typo and prepare an updated version.