The 1st place winner of the 4th On-device Visual Intelligence Competition (OVIC) of Low-Power Computer Vision Challenge (LPCVC), both classification track and detection track. The challenge competes for the best accuracy given latency constraint deploying neural networks on mobile phones.

• Tianzhe Wang
• Han Cai
• Shuai Zheng
• Jia Li
• Song Han

The model submitted for the OVIC and implementation code for training and exportation.

• OVIC track: Image Classification, Object Detection

We use Google's Pixel2 to measure the real latency for our exported tflite model.

We provide tflite models for evaluation here. User can use the scripts in the corresponding folder to get checkpoint, frozen graph and tflite.

We address the challenging problem of efficient inference across many devices and resource constraints, especially on edge devices. We propose an Once-for-All Network (OFA, ICLR'2020) that supports diverse architectural settings by decoupling model training and architecture search. We can quickly get a specialized sub-network by selecting from the OFA network without additional training. We also propose a novel progressive shrinking algorithm, a generalized pruning method that reduces the model size across many more dimensions than pruning (depth, width, kernel size, and resolution), which can obtain a surprisingly large number of sub-networks (> 10¹⁹) that can fit different latency constraints. On edge devices, OFA consistently outperforms SOTA NAS methods (up to 4.0% ImageNet top1 accuracy improvement over MobileNetV3, or same accuracy but 1.5x faster than MobileNetV3, 2.6x faster than EfficientNet w.r.t measured latency) while reducing many orders of magnitude GPU hours and CO₂ emission. In particular, OFA achieves a new SOTA 80.0% ImageNet top1 accuracy under 600M MACs. OFA is the winning solution for 4th Low Power Computer Vision Challenge, both classification track and detection track. Code and 50 pre-trained models on CPU/GPU/DSP/mobile CPU/mobile GPU (for different device & different latency constraints) are released at https://github.com/mit-han-lab/once-for-all.

OFA achieves 80.0% top1 accuracy with 595M MACs and 80.1% top1 accuracy with 143ms Pixel1 latency, setting a new SOTA ImageNet Top1 accuracy on the mobile device.

OFA consistently outperforms MobileNetV3 on mobile platforms.

Specialized OFA models consistently achieve significantly higher ImageNet accuracy with similar latency than non-specialized neural networks on CPU, GPU, mGPU, and FPGA. More remarkably, specializing for a new hardware platform does not add training cost using OFA.

• Once for All: Train One Network and Specialize it for Efficient Deployment [Website] [arXiv] [Slides] [Video]

@inproceedings{
  cai2020once,
  title={Once for All: Train One Network and Specialize it for Efficient Deployment},
  author={Han Cai and Chuang Gan and Tianzhe Wang and Zhekai Zhang and Song Han},
  booktitle={International Conference on Learning Representations},
  year={2020},
  url={https://arxiv.org/pdf/1908.09791.pdf}
}

• APQ: Joint Search for Network Architecture, Pruning and Quantization Policy (To be appear in CVPR 2020)

@inproceedings{
  wang2020apq,
  title={APQ: Joint Search for Network Architecture, Pruning and Quantization Policy},  
  author={Wang, Tianzhe and Wang, Kuan and Cai, Han and Lin, Ji and Liu, Zhijian and Wang, Hanrui and Lin, Yujun and Han, Song},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  year={2020}
}

See the corresponding folder for details.

Apache License 2.0