Solution to Visual Wakeup Words Challenge'19 (first place).

Participants: Song Han, Ji Lin, Kuan Wang, Tianzhe Wang, Zhanghao Wu (following alphabetical order)

Contact: jilin@mit.edu

Instruction

The demo code on Raspberry Pi and Android is included in this repo under the demos folder.

Youtube: https://youtu.be/7-beBCKVpFE

Article: https://medium.com/tensorflow/visual-wake-words-with-tensorflow-lite-micro-8578e59ea6f9

We have optimized the model with uint8 quantization and converted it to tf-lite format. Here we provide a script to evaluate the model with PyTorch data loader in eval.py. However, to keep consistent with TensorFlow preprocessing, we used the preprocessing function imported from tensorflow. The preprocessing we used is defined in preprocess.py.

Our floating point model (model_fp32.pb) can get 95.40% top-1 accuracy on the minival set of VWW.

Our quantized model (model_quantized.tflite) can get 94.575% top-1 accuracy on the minival set of VWW.

Usage

Run:

python eval.py

Citation

@article{cai2018proxylessnas,
  title={Proxylessnas: Direct neural architecture search on target task and hardware},
  author={Cai, Han and Zhu, Ligeng and Han, Song},
  journal={International Conference on Learning Representations (ICLR)},
  year={2019}
}