shanyou92 / Channel-Bench-Macro

Channel-Bench-Macro

This repository includes the benchmark and code for Channel-Bench-Macro. Channel-Bench-Macro is the corresponding benchmark of NAS-Bench-Marco, which serves for channel number search / channel pruning algorithms.

Channel-Bench-Macro is a channel benchmark on macro search space. The Channel-Bench-Macro consists of 16384 X 2 networks for MobileNet and ResNet search space. We provide the test accuracies, parameters, FLOPs, and training logs on CIFAR-10 dataset.

Each architecture in Channel-Bench-Macro is trained from scratch isolatedly.

Benchmark

All the evaluated architectures are stored in file Resuls_MobileNet.json and Results_ResNet.json with the following format:

{
    arch1: {
        acc: [float, float, float], // the test accuracies of three independent training
        mean: float, // mean accuracy of three test accuracies
        std: float, // the standard deviation of test accuracies
        params: int, // parameters
        flops: int, // FLOPs 
    },
    arch2: ......
}

Search Space

The search space of Channel-Bench-Macro is conducted with 7 independent layers for MobileNet and ResNet, respectivel, the width of skip line is directly derived from the corresponding independent layer. Each layer in contains 4 candidate widths: [0.25, 0.5, 0.75, 1.0], which amounts to 4^7 = 16384 networks for each search space (MobileNet and ResNet).

Network structure

Search space of MobileNet for Channel-Bench-Macro. The ratio for each MobileNet block is set to 6 as default, which influences the maximum channel of the second conv in MB block.

n	input	block	channel	stride
1	32 X 32 X 3	3 X 3 conv	128	2
2	16 X 16 X 128	MB block	128	1
1	16 X 16 X 128	MB block	256	2
1	16 X 16 X 256	MB block	256	1
1	16 X 16 X 256	1 X 1 conv	1024	1
1	16 X 16 X 1024	global avgpool	-	-
1	16 X 16 X 1024	FC	10	-

For the ResNet search space, the ResNet block is introduced as RS block for abbreviate.

n	input	block	channel	stride
1	32 X 32 X 3	3 X 3 conv	256	2
2	16 X 16 X 256	RS block	256	1
1	16 X 16 X 256	RS block	512	2
2	16 X 16 X 512	RS block	512	1
1	16 X 16 X 512	global avgpool	-	-
1	16 X 16 X 512	FC	10	-

Statistics of histograms of networks

Accuracy, FLOPs, and Parameters of the proposed Channel-Bench-Macro w.r.t all architectures for ResNet and MobileNet search space.

Requirements

torch>=1.0.1
torchvision

Reproduce the Results

Training scripts

cd train_folder

python train.py

Training logs

If you need training logs to help your research, i.e., you need the training accuracy or test accuracy of architectures from all training epochs. You can download the training logs of all architectures of MobileNet (Google Drive) and ResNet(Google Drive) for a detailed analysis.

Citation

If you find that Channel-Bench-Macro helps your research, please consider citing it:

@inproceedings{bcnet,
  title={BCNet: Searching for Network Width with Bilaterally Coupled Network},
  author={Su, Xiu and You, Shan and Wang, Fei and Qian, Chen and Zhang, Changshui and Xu, Chang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2175--2184},
  year={2021}
}

@misc{su2021vision,
      title={Vision Transformer Architecture Search}, 
      author={Xiu Su and Shan You and Jiyang Xie and Mingkai Zheng and Fei Wang and Chen Qian and Changshui Zhang and Xiaogang Wang and Chang Xu},
      year={2021},
      eprint={2106.13700},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}