DenseNet: Densely Connected Convolutional Networks in Caffe?

[Evan Weiner]

DenseNet ("Densely Connected Convolutional Networks") is apparently the state-of-art architecture now for deep nets. It is implemented in Torch. Anyone know if a fork exists to run on Caffe? If not, anyone planning on doing so?

[Zhuang Liu]

Hi,

check https://github.com/liuzhuang13/DenseNetCaffe

Thanks for your interest!

[nila...@gmail.com]

It seems there is no version pretrained on ImageNet yet?

Regards
Nilas

[nila...@gmail.com]

So there actually is an imagenet trained model available at https://github.com/liuzhuang13/DenseNet/issues/5 . These models can be converted to caffe using https://zhanghang1989.github.io/Torch2CaffeConverter/ . I am currently trying to train a FCN model based on this, however I cannot make the model converge. I found that sgd works better than the nesterov solver that is used for the original implementation, but still I cannot get a useful model this way. Is there a systematic way to choose the learn rate? Also does anyone know if mean BGR value substraction is done in Torch in the same way as it is in caffe?

[Zhuang Liu]

Hi, there was a recent work adapting DenseNets to FCNs (https://arxiv.org/abs/1611.09326). But they didn't use pretrained models.

For RGB subtraction, check https://github.com/facebook/fb.resnet.torch. We trained our models under the same setting as theirs. Specifically, check line 93- 99 of https://github.com/facebook/fb.resnet.torch/blob/master/datasets/imagenet.lua. They should be the input processing for test time.

We released a new set of pretrained models (described in our updated paper) here https://github.com/liuzhuang13/DenseNet#imagenet-and-pretrained-models.

If you can confirm the conversion to caffe is successful, could you please share the caffe models with us? Thanks in advance. I tried the converter but can't succeed :(

[nila...@gmail.com]

I didn't have time to do further experiments and I did not test the converted models on imagenet, nevertheless here are my files.