search

Lyken17 / SparseNet

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

the number of parameters dose not fits #2

Closed yueshiyi closed 6 years ago

yueshiyi commented 6 years ago

There is a small discrepancy in the number of parameters between the paper and this version of code. The number of parameters of the sparsenet-40-24 is 0.73m according to the code, and it should be 0.76m according to your paper. Sorry, I don’t known what’s going wrong, can you help me?

Lyken17 commented 6 years ago

Thanks for you interest in our work.

The small discrepancy is caused by my mistake in source code

In line ~161

         if fetch in ["linear", "dense"]:
             nChannels = 2 * growthRate
         elif fetch in ["exp", "sparse"]:
             nChannels = growthRate

This is one of our experimental settings, to explore how the initial block size affects model performance. In final release version, please change to 

             nChannels = 2 * growthRate

Then the parameters will be exactly same as listed in the paper.

Lyken17 commented 6 years ago

I will soon fix the mistake in next update, together with imagenet pretrained models.

Lyken17 commented 6 years ago

Closed for no update after a long time. Feel free to reopen it if necessary.