Architecture of CondenseNet{light-160*, 182*, light-94, 84}

[geevi]

Hi,
The paper mentions CondenseNet{light-160*, 182*, light-94, 84} for CIFAR, though is not clear about the details of the architecture. Could you share the architectures and how those results can be reproduced?

[Coderx7]

@ShichenLiu : Thanks for your great work. would you please have a look here and help us with the architectures? its greatly appreciated

[gaohuang]

Sorry for the late reply.

CondenseNet^{light} refers to the network that simply applies learned group convolution to the original DenseNet network, and CondenseNet further introduces two architecture changes: full dense connection and increasing growth rate. The legend of Figure 6 in the paper explains these in a more intuitive way.

For the network configurations, CondenseNet^{light} always has 3x2xN+4 layers, where N is the number of densenet layers (one 1x1 conv plus on 3x3 conv), 3 corresponds to the three dense blocks, and 2 corresponds to the 2 conv layers in each densenet layer. The number 4 counts the very first conv layer, the two transition layers, and the final FC layer. CondenseNet always has 3x2xN+2 layers, because it does not have the two transition layers. The growth rate for these networks are set differently, such that the resulting network has comparable parameters or flops as some of the baseline networks. I paste the command lines to reproduce the reported results below (please correct me if I'm wrong @ShichenLiu ):

CondenseNet-86
python main.py --model condensenet -b 64 -j 2 cifar10 --epochs 300 --stages 14-14-14 --growth 8-16-32

CondenseNet-182*
python main.py --model condensenet -b 64 -j 2 cifar10 --epochs 600 --stages 30-30-30 --growth 12-24-48

CondenseNet-light-94
python main.py --model densenet -b 64 -j 2 cifar10 --epochs 300 --stages 15-15-15 --growth 16-16-16

CondenseNet-light-160*
python main.py --model densenet -b 64 -j 2 cifar10 --epochs 600 --stages 26-26-26 --growth 32-32-32

[ShichenLiu]

Hi @geevi @Coderx7 ,

Sorry for the late reply. The command that could reproduce the results are:

CondenseNet-86
python main.py --model condensenet -b 64 -j 2 cifar10 --epochs 300 --stages 14-14-14 --growth 8-16-32

CondenseNet-182*
python main.py --model condensenet -b 64 -j 2 cifar10 --epochs 600 --stages 30-30-30 --growth 12-24-48

CondenseNet-light-94
python main.py --model densenet_LGC -b 64 -j 2 cifar10 --epochs 300 --stages 15-15-15 --growth 16-16-16

CondenseNet-light-160*
python main.py --model densenet_LGC -b 64 -j 2 cifar10 --epochs 600 --stages 26-26-26 --growth 32-32-32

[Coderx7]

@gaohuang and @ShichenLiu : Thank you very much, guys. its really appreciated ;)

[lizhenstat]

Hi, @ShichenLiu did you set group-lasso-lambda to 1e-5 on cifar100 dataset
the paper noted group-lasso-lambda=1e-5 on ImageNet Dataset
the default value for the parameter is 0.

Looking forward to your reply

[ShichenLiu]

Hi,

The group-lasso-lambda makes no conspicuous difference on CIFAR dataset. However, we set it to 1e-5 on ImageNet dataset.

[lizhenstat]

@ShichenLiu Hi, does group lasso make any difference on ImageNet. Since the paper seems only gives the results with group lasso on ImageNet, right? What is the result if not including this term? Thanks

[emergencyd]

What is the converted model for densenet_LGC?

[xiaohe725]

What is the specific network structure configuration of Condensenetv2 on the CIFAR dataset of Condensenetv2-110 and Condensenetv2-146

[ShichenLiu]

hi @xiaohe725 , this repository does not contain models for CondenseNet v2.

[xiaohe725]

Yes, but I haven't seen it in the paper and code of CondensenetV2 either