swingxu / torch_quant

This is the Implementation of the uniform/non-uniform quantization of deep learning model.

Implentation of Uniform/Non-Uniform Quantization of Weight Normalization based model

Still Under Construction.

Todo

• kmeans quantization
• finetuning the quantized model for cifar
• freeze the weights and training the scale factor after quantization
• finetuning the quantized model for imagenet
• expansion and compression of the weights
• depth-wise convolution quantization

Usage

Test resnet

python main_imgnet_training.py ~/workspace/dataset/torch_imagenet/CLS-LOC/ -a resnet --pretrained  --resume ./R-50-GN-WS.pth.tar -e --dist-url 'tcp://127.0.0.1:8888' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0、

Train Activation quantized model for cifar10

python train_cifar.py -a alexnet --epochs 164 --schedule 81 122 --gamma 0.1 --lr 0.01 --a_quant --a_bit 8 --checkpoint checkpoints/cifar10/alexnet_act8   

Test with cifar10

 python train_cifar.py -a alexnet --epochs 164 --schedule 81 122 --gamma 0.1 --lr 0.01 --checkpoint checkpoints/cifar10/alexnet_dorefa --resume checkpoints/cifar10/alexnet/model_best.pth.tar

Save the Standard Weight based model as normal model

python save_WN_as_norm.py

Quantize the weight

kmeans quantization：

python weightquant.py --input_ckpt ./R-50-GN-WS-flod.pth.tar --quant_method kmeans --output_ckpt ./R-50-GN-WS-q --weight_bits 4 --n_sample 10000

linear quantization：

python weightquant.py --input_ckpt ./R-50-GN-WS-flod.pth.tar --quant_method linear --output_ckpt ./R-50-GN-WS-q --weight_bits 4

Citaiton:

* Weight Standardlization

* AlexNet

* VGG (Imported from pytorch-cifar)

* ResNet

* Pre-act-ResNet

* ResNeXt (Imported from ResNeXt.pytorch)

* Wide Residual Networks (Imported from WideResNet-pytorch)

* DenseNet