clcNet Torch Implementation

This is the Torch implementation of clcNet, an efficient convolutional neural network, presented in clcNet: Improving the Efficiency of Convolutional Neural Network using Channel Local Convolutions

The code is modified from fb.resnet.torch.

Training and Validation

1. Install Torch, cuDNN on a machine with CUDA GPU. You can refer here for a step-by-step guide.

2. Download the ImageNet dataset and prepare the data. Please refer here for the instructions of data preparation.

3. Clone this repo: git clone https://github.com/dqzhang17/clcnet.torch.git

4. For training from scratch : Choose clcneta or clcnetb for netType, and change nGPU, batchSize etc. according to your hardware infrastructure:

th main.lua -nGPU 4 -batchSize 256 -nThreads 8 -data [your_imagenet_folder] -netType clcneta -nEpochs 100 -LR 0.1 -shareGradInput true

Or you can build your own custom clcnet by choosing layer counts for different stages. For instance, let a=1, b=1, c=5, d=2:

th main.lua -nGPU 4 -batchSize 256 -nThreads 8 -data [your_imagenet_folder] -a 1 -b 1 -c 5 -d 2 -nEpochs 100 -LR 0.1 -shareGradInput true

5. For validation :

th main.lua -nGPU 4 -batchSize 256 -nThreads 8 -data [your_imagenet_folder] -retrain [your_trained_model.t7] -testOnly true

Results on ImageNet and Pretrained Models

The accuracy rates shown here are 224x224 single-model single-crop test accuracy on the ImageNet-1k validation dataset.

Network	Layer Config (a,b,c,d)	Top-1 Accuracy	Torch Model
clcNet-A	(1, 1, 5, 2)	70.4%	Download (26.2MB)
clcNet-B	(1, 1, 7, 3)	71.6%	Download (32.9MB)

BibText

@inproceedings{zhang2017clcnet,
  title={clcNet: Improving the Efficiency of Convolutional Neural Network using Channel Local Convolutions},
  author={Zhang, Dong-Qing},
  journal={arXiv preprint arXiv:1712.06145},
  year={2017}
}