Implementation for paper "Latent Weights Do Not Exist: Rethinking Binarized Neural Network Optimization".
A poster illustrating the proposed algorithm and its relation to the previous BNN optimization strategy is included at ./poster.pdf.
Note: Bop is now added to Larq, the open source training library for BNNs. We recommend using the Larq implementation of Bop: it is compatible with more versions of TensorFlow and will be more actively maintained.
- Python version
3.6
or3.7
- Tensorflow version
1.14+
or2.0.0
- Larq version
0.2.0
- Zookeeper version
0.1.1
You can also check out one of our prebuilt docker images.
This is a complete Python module. To install it in your local Python environment, cd
into the folder containing setup.py
and run:
pip install -e .
To train a model locally, you can use the cli:
bnno train binarynet --dataset cifar10
To reproduce the runs exploring various hyperparameters, run:
bnno train binarynet \
--dataset cifar10 \
--preprocess-fn resize_and_flip \
--hparams-set bop \
--hparams threshold=1e-6,gamma=1e-3
where you use the appropriate values for threshold and gamma.
To achieve the accuracy in the paper of 91.3%, run:
bnno train binarynet \
--dataset cifar10 \
--preprocess-fn resize_and_flip \
--hparams-set bop_sec52 \
To reproduce the reported results on ImageNet, run:
bnno train alexnet --dataset imagenet2012 --hparams-set bop
bnno train xnornet --dataset imagenet2012 --hparams-set bop
bnno train birealnet --dataset imagenet2012 --hparams-set bop
This should give the results listed below. Click on the tensorboard icons to see training and validation accuracy curves of the reported runs.
Network | Bop - top-1 accuracy | |
---|---|---|
Binary Alexnet | 41.1% | |
XNOR-Net | 45.9% | |
Bi-Real Net | 56.6% |