This repo contains some of the compressed models from paper AMC: AutoML for Model Compression and Acceleration on Mobile Devices (ECCV18).

If you find the models useful, please kindly cite our paper:

@inproceedings{he2018amc,
  title={AMC: AutoML for Model Compression and Acceleration on Mobile Devices},
  author={He, Yihui and Lin, Ji and Liu, Zhijian and Wang, Hanrui and Li, Li-Jia and Han, Song},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  pages={784--800},
  year={2018}
}

Firstly, download the pretrained models from here and put it in ./checkpoints.

We provide compressed MobileNetV1 by 50% FLOPs and 50% Inference time, and also compressed MobileNetV2 by 70% FLOPs, with PyTorch. The comparison with vanila models as follows:

To test the model, run:

python eval_mobilenet_torch.py --profile={mobilenet_0.5flops, mobilenet_0.5time, mobilenetv2_0.7flops}

We converted the 50% FLOPs and 50% time compressed MobileNetV1 model to TensorFlow. We offer the normal checkpoint format and also the TF-Lite format. We used the TF-Lite format to test the speed on MobileNet.

To replicate the results of PyTorch, we write a new preprocessing function, and also adapt some hyper-parameters from the original TF MobileNetV1. To verify the performance, run following scripts:

python eval_mobilenet_tf.py --profile={0.5flops, 0.5time}

The produced result is:

Here we provide timing logs on Google Pixel 1 using TensorFlow Lite in ./logs directory. We benchmarked the original MobileNetV1 (mobilenet), MobileNetV1 with 0.75 width multiplier (0.75mobilenet), 50% FLOPs pruned MobileNetV1 (0.5flops) and 50% time pruned MobileNetV1 (0.5time). Each model is benchmarked for 200 iterations with extra 100 iterations for warming up, and repeated for 3 runs.

You can also find our PyTorch implementation of AMC Here.

To contact the authors:

Ji Lin, jilin@mit.edu

Song Han, songhan@mit.edu