Skip to content
First place in the Visual Wake Words challenge (TF-lite track) in LPIRC'19
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
demos Finalized README and formatting Aug 6, 2019 Update Sep 11, 2019 [major] init files Jun 26, 2019 [major] init files Jun 26, 2019
model_fp32.pb Add float point .pb model. Sep 11, 2019
model_quantized.tflite [major] init files Jun 26, 2019 [major] init files Jun 26, 2019

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

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



We have converted our model to tflite format with uint8 quantization. Here we provide a script to evaluate the model with PyTorch data loader in However, to keep consistent with TensorFlow preprocessing, we used the preprocessing function imported from tensorflow. The preprocessing we used is defined in

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.

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





  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)},
You can’t perform that action at this time.