Awesome
Learning to Quantize
Code for the ICML submission 6132.
Dependencies
We recommend using Anaconda to install the following packages,
- Python 3.7.1
- PyTorch (>1.1.0)
- TensorboardX
- Pyyaml
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
Cuda kernel installation
cd nuq/cuda/;
python setup.py install
cd ../../
Running experiments in the paper
The commands used to run the experiments can be found in pjobs/ directory.
These commands are generated using the grid_run.py script. Each experiment is
described using a function in the grid/nuq.py file.
Each of the experiments can be run using bash pjobs/supp_{i}.sh where i is
the experiment number being used. The small configuration changes in each
experiment corresponds to variations in the quantization method used for
generating the plots.
After running the experiments, the results will be stored in a directory
specified in the configuration of the job that can be served using tensorboard
to view the results.
Quantization
Quantization methods are implemented in NumPy (nuq/quantize.py) as well is in
Cuda (nuq/cuda/src/ops_gpu.cu).