search

liuzhuang13 / DenseNet

Densely Connected Convolutional Networks, In CVPR 2017 (Best Paper Award).
BSD 3-Clause "New" or "Revised" License
4.69k stars 1.06k forks source link

Purpose using first convolution #29

Closed John1231983 closed 6 years ago

John1231983 commented 6 years ago

In your network architecture for CIFAR and imagenet dataset, what does purpose use the first convolution (before pooling-denseblock1)? In the imagenet, you use two convolutions block before entering the dense-block, while CIFAR just one, any reason? Thanks

liuzhuang13 commented 6 years ago

Hi @John1231983 We follow ResNet's design on both datasets. https://github.com/facebook/fb.resnet.torch

John1231983 commented 6 years ago

Thanks. I knew it. In your opinion, do you have any explanation?

liuzhuang13 commented 6 years ago

In ImageNet we need to downsample the image twice (so that the image becomes of 56x56 size) before entering the first dense block, this is through one stride convolution and one max pooling.

In CIFAR dataset, the input image and the first dense block feature maps are of the same size (32x32), the first convolution is just a conventional design so that we don't include raw input image in the first dense block.