<script type="text/javascript" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js?config=default"></script>

L-DNQ

Codes for AAAI2019 paper: Deep Neural Network Quantization via Layer-Wise Optimization using Limited Training Data

How to use it

Specify your Dataset Root

User need to specify their dataset root in utils/dataset.py. For example, in line 39 for CIFAR10 and line 70 for ImageNet.

Prepare Pre-trained Model

L-DNQ is conducted based on pre-trained model. Therefore you need to:

• CIFAR10 Dataset

run train_base_model.py to train a pre-trained model for CIFAR10 dataset. It will generate a pre-trained model in folder ResNet20:

cd L-DNQ
python train_base_model.py
python train_base_model.py --resume --lr=0.01 # If needed

• ImageNet Dataset

Download the pre-trained model from torchvision model zoo to folder ResNet18-ImageNet. For example, download resnet18 from https://download.pytorch.org/models/resnet18-5c106cde.pth as is instructed in models_ImageNet/resnet.py:

cd L-DNQ
mkdir ResNet18-ImageNet
wget https://download.pytorch.org/models/resnet18-5c106cde.pth

During Quantization

Currently only quantization on ResNet model using CIFAR10/ImageNet dataset is available. To reproduce ResNet20 quantization using CIFAR10:

python main.py

To reproduce ResNet18 quantization using ImageNet:

• Uncomment Line 69, 74 (Comment 68, 73) in main.py to change network.

  python main.py --model_name=ResNet18-ImageNet --dataset_name=ImageNet

To reproduce other experiments, please change the network structure accordingly in the code.

More Specifications

Change portion of dataset used

In our experiment, only 1% of original dataset is used. If you want to change that portion. Change

get_dataloader(dataset_name, 'limited', batch_size = 128, ratio=0.01) in main.py to

get_dataloader(dataset_name, 'limited', batch_size = 128, ratio=Whatever you want, resample=True)

Set resample to False after new selected dataset is generated.

If user want to use the whole dataset, simply use get_dataloader(dataset_name, 'train', batch_size = 128)

Change the quantized bits

Change the argument --kbits to 3,5,7,9,11. For example, 5 means the quantization bits are: $0, \pm \alpha, \pm 2*\alpha, \pm 4*\alpha$, totally 5 bits.

Results

###CIFAR10

Network	bits	Quantized Acc(%)	Original Acc(%)	Acc Improve(%)
ResNet20	3	85.95	91.50	-5.55
ResNet32	3	88.47	92.13	-3.66
ResNet56	3	89.17	92.66	-3.49

###ImageNet

Network	bits	Quantized Acc(%)	Original Acc(%)	Acc Improve(%)
ResNet18	3	45.98/75.73	69.76/89.02	-16.43/-10.67
ResNet34	3	43.99/73.05	73.30/91.42	-29.31/-18.37
ResNet50	3	56.49/81.55	76.15/92.87	-19.66/-11.32

Requirement

PyTorch > 0.4.0

TensorFlow > 1.3.0

Support

Please use github issues for any problem related to the code. Send email to the authors for general questions related to the paper.