Home

Awesome

FNA

Official implementation of Fast Neural Network Adaptation via Parameter Remapping and Architecture Search (ICLR 2020) and FNA++: Fast Network Adaptation via Parameter Remapping and Architecture Search (TPAMI).

Deep neural networks achieve remarkable performance in many computer vision tasks. Most state-of-the-art (SOTA) semantic segmentation and object detection approaches reuse neural network architectures designed for image classification as the backbone, commonly pre-trained on ImageNet. However, performance gains can be achieved by designing network architectures specifically for detection and segmentation, as shown by recent neural architecture search (NAS) research for detection and segmentation. One major challenge though, is that ImageNet pre-training of the search space representation (a.k.a. super network) or the searched networks incurs huge computational cost.

In this paper, we propose a Fast Neural Network Adaptation (FNA) method, which can adapt both the architecture and parameters of a seed network (e.g. a high performing manually designed backbone) to become a network with different depth, width, or kernels via a Parameter Remapping technique, making it possible to utilize NAS for detection/segmentation tasks a lot more efficiently.

framework

Noting

We have released the extension version FNA++: Fast Network Adaptation via Parameter Remapping and Architecture Search and the related code of it will be released in the future.

Results

In our experiments, we conduct FNA on MobileNetV2 to obtain new networks for both segmentation and detection that clearly out-perform existing networks designed both manually and by NAS. The total computation cost of FNA is significantly less than SOTA segmentation/detection NAS approaches: 1737x less than DPC, 6.8x less than Auto-DeepLab and 7.4x less than DetNAS.

<div align="center"> <img src="./imgs/seg_results.png" width = "700"> <img src="./imgs/seg_cost.png" width = "600"> <img src="./imgs/det_results.png" width = "550"> <img src="./imgs/det_cost.png" width = "700"> </div>

FNA on Object Detection

Requirements

Architecture Adaptation

Adapt the architecture of the seed network to the target dataset COCO. The adaptation process is performed on 8 TITAN-XP GPUs. First go to the path of the detection project cd fna_det.

Parameter Adaptation

Adapt the parameters of the target architecture on COCO. The adaptation is performed on 8 TITAN-XP GPUs.

cd fna_det

The seed network MobileNetV2 is trained on ImageNet using the code of DenseNAS. We provide the pre-trained weights and net_config of the seed network in MobileNetV2. The code of MobileNetV2 model is constructed in models/derived_imagenet_net.py.

Evaluation

We provide the adapted parameters and net_config in checkpoint RetinaNet and SSDLite. The complete model zoo is in FNA_modelzoo. You can evaluate the checkpoint with the following script.

FNA on Semantic Segmentation

Requirements

Evaluation

  1. The adapted parameters and net_config are available in DeepLabV3.

  2. Put the adapted parameters epoch-last.pth into fna_seg/model/deeplab/cityscapes.deeplabv3.

  3. You can evaluate the checkpoint with the following script.

    cd fna_seg/model/deeplab/cityscapes.deeplabv3
sh eval.sh.

Citation

If you find this repository/work helpful in your research, welcome to cite it.

@ARTICLE{fna++,
  author={Fang, Jiemin and Sun, Yuzhu and Zhang, Qian and Peng, Kangjian and Li, Yuan and Liu, Wenyu and Wang, Xinggang},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={FNA++: Fast Network Adaptation via Parameter Remapping and Architecture Search}, 
  year={2020},
  doi={10.1109/TPAMI.2020.3044416}}

@inproceedings{
    fang2020fast,
    title={Fast Neural Network Adaptation via Parameter Remapping and Architecture Search},
    author={Jiemin Fang and Yuzhu Sun and Kangjian Peng and Qian Zhang and Yuan Li and Wenyu Liu and Xinggang Wang},
    booktitle={International Conference on Learning Representations},
    year={2020},
}

Acknowledgement

The code of FNA is based on

Thanks for the contribution of the above repositories.