dnn-gating

dnn-gating is a collective repository of precision-gating and channel-gating reimplemented in Pytorch.

Precision Gating (PG)

Requirments of PG

python 3.6.8 torch >= 1.3.0 numpy 1.16.4 matplotlib 3.1.0

Usage

With this repo, you can: - Evaluate uniform quantization and PACT. - Evaluate PG on ResNet CIFAR-10. - Apply PG to your own models and datasets.

Example

The following example trains ResNet-20 on CIFAR-10 with activations quantized to 3 bits, 2 MSBs out of which for prediction.

sh $ cd scripts $ source train_pg_pact.sh

Specify the Flags

Make sure to tune the training parameters in to achieve a good model prediction accuracy. -w : bitwidth of weights (floating-point if set to 0) -a : bitwidth of activations (floating-point if set to 0) -pact : use parameterized clipping for activatons -pg : use PG -pb : prediction bitwidth (only valid if -pg is turned on, and the bitwidth of prediction must smaller than that of activations) -gtar : the gating target -sg : the penalty factor on the gating loss -spbp : use sparse back-prop

Channel Gating (CG)

Requirments of CG

python 2.7.12 torch 1.1.0 numpy 1.16.4 matplotlib 2.1.0

Usage

With this repo, you can: - Evaluate CG on ResNet CIFAR-10 (both the original and modified post-activated ResNets). - The post-activated ResNet allows applying channel gating to all convolutional layers in a residual module. - Apply CG to your own models and datasets.

Example

The following examples use one fourth and half of input channels in the base path for the original and post-activated ResNets, respectively.

sh $ cd scripts $ source train_cg.sh $ source train_cg_postact.sh

Specify the Flags

The training parameters can be tuned to achieve different FLOP reduction and model accuracy. -lr : initial learning rate -wd: weigth decaying factor -pt: use 1/pt fraction of channels for prediction -gi: the intital value of gating thresholds -gtar: the target value of gating thresholds -spbp : use sparse back-prop -group: use group conv in the base path -cg : use CG -postact: use post-activated ResNet

Apply PG/CG to Your Own Models & Datasets

The following steps allows you to apply PG/CG to your own models. 1. Copy the model file to model/. 2. Import utils/pg_utils.py /utils/cg_utils.py in the model file, replace convolutional layers followed by activation functions with the PGConv2d/ CGConv2d module. 3. Import model/your_model.py in the generate_model() function in pg-cifar10.py/ cg-cifar10.py.

If you prepare your own training scripts, remember to add the gating loss to the model prediction loss before doing back-prop.

Note

• The way of exporting sparsity in the update phase we are using is only valid while training on a single GPU. This is because Pytorch modifies each model replica on a GPU instead of a global model if DataParallel is activated. For multi-GPU training, we suggest users turn off the sparsity printing during training, save the trained model, and print the sparsity only when testing.

Citation

If you use CG or PG in your research, please cite our NeurIPS'19 and ICLR'20 papers.

Channel Gating Neural Networks ```

@incollection{NIPS2019_8464, title = {Channel Gating Neural Networks}, author = {Hua, Weizhe and Zhou, Yuan and De Sa, Christopher M and Zhang, Zhiru and Suh, G. Edward}, booktitle = {Advances in Neural Information Processing Systems 32}, editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett}, pages = {1886--1896}, year = {2019}, publisher = {Curran Associates, Inc.}, url = {http://papers.nips.cc/paper/8464-channel-gating-neural-networks.pdf} } **Precision Gating: Improving Neural Network Efficiency with Dynamic Dual-Precision Activations** @inproceedings{ Zhang2020Precision, title={Precision Gating: Improving Neural Network Efficiency with Dynamic Dual-Precision Activations}, author={Yichi Zhang and Ritchie Zhao and Weizhe Hua and Nayun Xu and G. Edward Suh and Zhiru Zhang}, booktitle={International Conference on Learning Representations}, year={2020}, url={https://openreview.net/forum?id=SJgVU0EKwS} } ```