FracBNN

This repository serves as the official code release of the paper FracBNN: Accurate and FPGA-Efficient Binary Neural Networks with Fractional Activations (pubilished at FPGA 2021).

FracBNN, as a binary neural network, achieves MobileNetV2-level accuracy by leveraging fractional activations. In the meantime, its input layer is binarized using a novel thermometer encoding with minimal accuracy degradation, which improves the hardware resource efficiceny.

Citation

If FracBNN helps your research, please consider citing:

@article{Zhang2021fracbnn,
    title = "{FracBNN: Accurate and FPGA-Efficient Binary Neural Networks with Fractional Activations}",
    author = {Zhang, Yichi and Pan, Junhao and Liu, Xinheng and Chen, Hongzheng and Chen, Deming and Zhang, Zhiru},
    journal = {The 2021 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays},
    year = {2021}
}

Structure

|   cifar10.py (training script)
|   imagenet.py (training script)
|
└── models/
|   |   fracbnn_cifar10.py
|   |   fracbnn_imagenet.py
|
└───utils/
|   |   quantization.py
|   |   utils.py
|
└───xcel-cifar10/
|   |   High-level synthesis code for FracBNN CIFAR-10 accelerator
|
└───xcel-imagenet/
|   |   High-level synthesis code for FracBNN ImageNet accelerator

Dependency

Python 3.6.8
torch 1.6.0
torchvision 0.7.0
numpy 1.16.4

Run

Pretrained Model Release

• Pretrained CIFAR-10 model
• Pretrained ImageNet model

Test Only

• For CIFAR-10, run python cifar10.py -gpu 0 -t -r /path/to/pretrained-cifar10-model.pt -d /path/to/cifar10-data
• For ImageNet, run python imagenet.py -gpu 0,1,2,3 -t -r /path/to/pretrained-imagenet-model.pt -d /path/to/imagenet-data

Train (Two Step Training):

Please refer to the paper for details such as hyperparameters.

• Step 1: Binary activations, floating-point weights.
  • In utils/quantization.py, use self.binarize = nn.Sequential() in BinaryConv2d(), or modify self.binarize(self.weight) to self.weight in PGBinaryConv2d().
  • Run python cifar10.py -gpu 0 -s
• Step 2: Binary activations, binary weights.
  • In utils/quantization.py, use self.binarize = FastSign() in BinaryConv2d(), or self.binarize(self.weight) in PGBinaryConv2d().
  • Run python cifar10.py -gpu 0 -f -r /path/to/model_checkpoint.pt -s
  • Use -g to set the gating target if training with PGBinaryConv2d

Model Accuracy

Dataset	Precision (W/A)	1-bit Input Layer	Top-1 %
CIFAR-10	1/1.4 (PG)	Yes	89.1
ImageNet	1/1.4 (PG)	Yes	71.7

CIFAR-10 Accelerator

Compile the HLS code

cd ./xcel-cifar10/source/
make hls

Generate the Bitstream

This step should be done after compiling the HLS code. Assume you are in the directory of ./xcel-cifar10/source/.

make vivado

Deploy on Xilinx Ultra96v2

To test the bitstream on the board, the following files (sample images and labels) are needed:

• conv1-input.bin
• conv1-input-uint64.npy
• labels.bin

To deploy the bitstream:

• Step1: Download the files and move them to /xcel-cifar10/deploy/
• Step2: Move the generated bitstream and hardware definition files to /xcel-cifar10/deploy/
• Step3: Upload the entire directory /xcel-cifar10/deploy/ to the board
• Step4: Go to deploy/, run sudo python3 FracNet-CIFAR10.py on the board

ImageNet Accelerator

Compile HLS and Generate Bitstream

Please run the compilation flow using Vivado HLS. The top function is FracNet() in xcel-imagenet/source/bnn.cc.

Deploy on Xilinx Ultra96v2

To test the bitstream on the board, the following files (sample images, weights) are needed:

• sample image
• conv1x1 weights
• conv3x3 weights
• classifier bias
• classifier weights
• batchnorm/BPReLU weights

To deploy the bitstream:

• Step1: Download the files and move them to /xcel-imagenet/deploy/
• Step2: Move the generated bitstream and hardware definition files to /xcel-imagenet/deploy/
• Step3: Upload the entire directory /xcel-imagenet/deploy/ to the board
• Step4: Go to deploy/, run sudo python3 FracNet-ImageNet.py on the board