Code for reproducing results of Bit-Swap: Recursive Bits-Back Coding for Lossless Compression with Hierarchical Latent Variables, appearing at International Conference on Machine Learning 2019.
The code is written by Friso H. Kingma. The paper is written by Friso H. Kingma, Pieter Abbeel and Jonathan Ho.
We present a lossless compression scheme, called Bit-Swap, that results in compression rates that are empirically superior to existing techniques. Our work builds on BB-ANS that was originally proposed by Townsend et al, 2019. BB-ANS exploits the combination of the ''bits back'' argument (Hinton & Van Camp, 1993), latent variable models and the entropy encoding technique Asymmetric Numeral Systems (ANS) (Duda et al, 2015). We extended BB-ANS to be more efficient for hierarchical latent variable models, that are known to be better density estimators.
In one of the experiments, we compressed 100 unscaled and cropped images of ImageNet with Bit-Swap, BB-ANS and other benchmark compressors. Details of the experimental setup can be found here or in the paper. In this regime, Bit-Swap outperforms the other compression schemes. The experimental setup and results of the other experiments can be found in the paper.
| Compression Scheme | Rate (bits/dim) |
|---|---|
| Uncompressed | 8.00 |
| GNU Gzip | 5.96 |
| bzip2 | 5.07 |
| LZMA | 5.09 |
| PNG | 4.71 |
| WebP | 3.66 |
| BB-ANS | 3.62 |
| Bit-Swap | 3.51 |
For this experiment, we constructed our own train and test set of ImageNet images as described in the instructions here. We trained a model on random 32x23 pixel-patches of the constructed train set. Afterwards we
- independently took 100 images from the constructed test set
- cropped the images to multiples of 32 pixels on each side
- split the images up into grids of 32x32 pixel-blocks
- compressed the resulting sequence of pixel-blocks of every image with Bit-Swap and BB-ANS and finally
- calculated the average bitrate over the pixel-blocks of every image independently.
We also compressed the cropped images resulting from step 2 with other benchmark compressors.
Note: Before executing any of the scripts, look through the requirements section and make sure all the requirements are satisfied.
The repository consists of three main parts:
- A demo to compress and decompress your own image.
- Training of the variational autoencoders.
- Compression with Bit-Swap and BB-ANS using the trained models.
Scripts relating to the demo, including
- Compression of your own image with Bit-Swap to a file (
image_compress.py) - Decompression with Bit-Swap to a reconstruction of your image (
image_decompress.py)
Scripts relating to training of the models on the training sets of
- MNIST (
mnist_train.py) - CIFAR-10 (
cifar_train.py) - ImageNet (32x32) (
imagenet_train.py)
and on random 32x32 pixel-patches of
- ImageNet (original size: unscaled and uncropped) (
imagenetcrop_train.py)
can be found in the subdirectory /model.
Scripts relating to compression with Bit-Swap and BB-ANS of the (partial) test sets of
- MNIST (
mnist_compress.py) - CIFAR-10 (
cifar_compress.py) - ImageNet (32x32) (
imagenet_compress.py)
and on 100 images independently taken from the test set of unscaled ImageNet, cropped to multiples of 32 pixels on each side
- ImageNet (unscaled and cropped) (
imagenetcrop_compress.py)
are in the top directory. The script for compression using the benchmark compressors (benchmark_compress.py) and the script for discretization of the latent space (discretization.py) can also be found in the top directory. The script imagenetcrop_compress.py also directly compresses the images with the benchmark compressors.
The following is required to run the scripts:
- Python (3.7)
- OpenMPI and Horovod (0.16.0)
- Numpy (1.15.4)
- PyTorch (1.0.0)
- Torchvision (0.2.1)
- Tensorflow (1.13.1)
- Tensorboard (1.31.1)
- TensorboardX (1.6)
- tqdm (4.28.1)
- Matplotlib (3.0.2)
- Scipy (1.1.0)
- Scikit-learn (0.20.1)
Add the top directory of the repository to the $PYTHONPATH variable. For example:
export PYTHONPATH=$PYTHONPATH:~/bitswap
Installation instructions for OpenMPI + Horovod are available on the github page of Horovod.
We also highly recommend using GPU's and a machine with a large memory capacity. PyTorch is highly optimized for GPU deployment and some calculations (especially discretization) might take up a large amount of memory.
First download the downsized version of ImageNet here. Unpack the train and validation set directories in model/data/train_32x32 and model/data/valid_32x32 respectively. After that, run
python create_imagenet.py
First download the unscaled ImageNet validation set here and the test set here. Unpack the images of both datasets in model/data/imagenetfull/train/class. After that, independently take 5000 images from this folder and move them into model/data/imagenetfull/test/class. In Ubuntu, this can be achieved with the following commands:
cd ~/bitswap/model/data/imagenetfull/train/class
ls | shuf -n 5000 | xargs -i mv {} ~/bitswap/model/data/imagenetcrop/test/class
Pretrained (PyTorch) model checkpoints:
- Download MNIST model checkpoints and unpack in
/model/params/mnist - Download CIFAR-10 model checkpoints and unpack in
/model/params/cifar - Download ImageNet (32x32) model checkpoints and unpack in
/model/params/imagenet - Download ImageNet (unscaled) model checkpoints and unpack in
/model/params/imagenetcrop
- Download MNIST latent space discretization bins and unpack in
/bins - Download CIFAR-10 latent space discretization bins and unpack in
/bins - Download ImageNet (32x32) latent space discretization bins and unpack in
/bins - Download ImageNet (unscaled) and DEMO latent space discretization bins and unpack in
/bins
python mnist_train.py --nz=8 --width=61
python mnist_train.py --nz=4 --width=62
python mnist_train.py --nz=2 --width=63
python mnist_train.py --nz=1 --width=64
mpiexec -np 8 python cifar_train.py --nz=8 --width=252
mpiexec -np 8 python cifar_train.py --nz=4 --width=254
mpiexec -np 8 python cifar_train.py --nz=2 --width=255
mpiexec -np 8 python cifar_train.py --nz=1 --width=256
mpiexec -np 8 python imagenet_train.py --nz=4 --width=254
mpiexec -np 8 python imagenet_train.py --nz=2 --width=255
mpiexec -np 8 python imagenet_train.py --nz=1 --width=256
mpiexec -np 8 python imagenetcrop_train.py --nz=4 --width=256
python mnist_compress.py --nz=8 --bitswap=1
python mnist_compress.py --nz=8 --bitswap=0
python mnist_compress.py --nz=4 --bitswap=1
python mnist_compress.py --nz=4 --bitswap=0
python mnist_compress.py --nz=2 --bitswap=1
python mnist_compress.py --nz=2 --bitswap=0
python cifar_compress.py --nz=8 --bitswap=1
python cifar_compress.py --nz=8 --bitswap=0
python cifar_compress.py --nz=4 --bitswap=1
python cifar_compress.py --nz=4 --bitswap=0
python cifar_compress.py --nz=2 --bitswap=1
python cifar_compress.py --nz=2 --bitswap=0
python imagenet_compress.py --nz=4 --bitswap=1
python imagenet_compress.py --nz=4 --bitswap=0
python imagenet_compress.py --nz=2 --bitswap=1
python imagenet_compress.py --nz=2 --bitswap=0
python imagenetcrop_compress.py
python benchmark_compress.py
python cma.py
python stackplot.py
First, clone the repository to your machine and follow the instructions under requirements. To compress, run
python demo_compress.py
The script will ask for the GPU index, which most likely is 0. Afterwards, it will ask for an image. The image first gets decompressed using it's own file format, which results in raw pixel data, after which the raw pixel data gets compressed by Bit-Swap and other benchmark compressors. The resulting encoded image is saved to a file named after the original filename, appended with '_' and the name of the corresponding compression scheme.
To decompress the Bit-Swap file, run
python demo_decompress.py
It will ask again for the GPU index, after which it will ask for the Bit-Swap encoded image file. The image gets decompressed, reconstructed and saved to a .jpeg format.
If there are any bugs, please contact Friso Kingma by e-mail: fhkingma@gmail.com
If you find our work useful, please cite us in your work.
@inproceedings{kingma2019bitswap,
title={Bit-Swap: Recursive Bits-Back Coding for Lossless Compression with Hierarchical Latent Variables},
author={Kingma, Friso H and Abbeel, Pieter and Ho, Jonathan},
booktitle={International Conference on Machine Learning},
year={2019}
}
Please contact Friso Kingma (fhkingma@gmail.com) if you have any questions.