search

chengtaipu / lowrankcnn

Low-rank convolutional neural networks
96 stars 44 forks source link

GoogLeNet discrepancies in paper #12

Closed ghost closed 6 years ago

ghost commented 6 years ago

I have found that in your paper, it is reported that GoogLeNet's inception module will contain one 3x3 Convolution followed by 2 consecutive 3x3 convolution layer. Whereas in GoogLeNet's original paper I couldn't find those 2 3x3 consecutive convolution layers.

Do you have a reason for that or are you referring to some other form of GoogLeNet architecture?

ghost commented 6 years ago

I went through more of your code under cifar directory and found that you have implemented your own GoogLeNet where 5x5 convolutions are replaced by 2 3x3 convolutions, and also you have an extra inception module - inception_3c. Would have been nice to mention some of these details original paper. Leaving this note here for any other user confused by the original paper.