Giters

x-zho14 / ProbMask-official

Implementation of Effective Sparsification of Neural Networks with Global Sparsity Constraint

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ResNet width

vanderschuea opened this issue · comments

commented

Hi,

The width of your ResNet is not correct, the first convolution should output a width of 16 not 32, and all other values should be divided by 2 too. Was this the network structure used to generate the results in the paper? Because if this is the case, the published results are wrong and not comparable to other papers (although I suppose the comparison to STR was re-run and thus the relative comparison inside the paper is nonetheless correct).

Thanks in advance for your response!

commented

Actually, this configuration of width of ResNet-32 on CIFAR-10/100 is described in appendix, following the setting of [1]. The results of STR are run under the same setting.

[1] Picking Winning Tickets Before Training by Preserving Gradient Flow

commented

Indeed, I read the version of the paper w/o the appendix at first and missed this information, thanks for your reply