Skip to content

AhmedTElthakeb/dcq.code

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 

Repository files navigation

DCQ (Divide and Conquer for Quantization):

Leveraging Intermediate Feature Representations for Quantized Training of Neural Networks

Image description

Example Usage:

  • Specify the quantization bits of each layer in the network specific yaml file:

dcq.code/dcq/examples/classifier_compression/alexnet_bn_dorefa.yaml

  • Print the model to know the index of each layer of the target network

dcq.py >> print(model)

  • Determine the split point of the network by freezing all layers except those of the particular stage to be trained

dcq.py line 572 >> model.freeze_partial([0, 2])

  • Return the right activation map for intermediate learning by specifiying the right index based on the printed model layers indices

dcq.code/dcq/models/imagenet/alexnet_batchnorm.py (line 84)

  • Run the following command

python3 dcq.py --arch alexnet_bn ../../../data.imagenet_100 --epochs 5 --fpresume alexnet_bn.pth.tar --resume alexnet_bn.pth.tar --lr 0.005 --quantize-eval --compress alexnet_bn_dorefa.yaml --gpus 2,3

--fpresume "path to the full precision reference model"

--resume "path to the model to be quantized and trained"

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published