README.md

Transform quantization for CNN compression (arXiv:2009.01174)

Source code for "transform quantization for CNN compression". Works out of the box for image classification CNNs on ImageNet. Slight modifications are necessary to quantize other types of networks. Commands below are for resnet18py.

1. Construct Karhunen–Loève transforms (KLTs) for convolution/dense layers.
   CUDA_VISIBLE_DEVICES=X python generate_KL_basis.py inter klt resnet18py 10000
   Here, 10000 is the number of ILSVRC2012 validation images for non-KLT transforms. Just keep the 10000 in there.

2. Construct rate–distortion curves for all convolution/dense layers.
   CUDA_VISIBLE_DEVICES=X python generate_RD_curves_base.py inter klt resnet18py 100
CUDA_VISIBLE_DEVICES=X python generate_RD_curves_kern.py inter klt resnet18py 100
   These are separate calls to construct RD curves for transform basis and kernels. Here, 100 is the number of ILSVRC2012 validation images to use.

3. Compute the rate–distortion frontier for your_network_name.
   CUDA_VISIBLE_DEVICES=X python generate_RD_frontier.py inter klt resnet18py 100 1 1 0
   Here, 100 1 1 0 means use 100 ILSVRC2012 validation images, quantize basis 1 and kernels 1, but not activations 0.

4. Modifying the code to work with your_own_cnn instead of resnet18py.
   Essentially, you'll need to have a notion of MSE for your network output. First, make sure dense layers in your_own_cnn are represented as 1x1-conv layers. Second, modify common.py:loadnetwork to load your_own_cnn and common.py:predict to appropriately return the output of your CNN.