Reference PyTorch implementation of "End-to-end optimized image compression with competition of prior distributions"

Last update: Jun 16, 2022

Overview

PyTorch reference implementation of "End-to-end optimized image compression with competition of prior distributions" by Benoit Brummer and Christophe De Vleeschouwer ( https://github.com/trougnouf/Manypriors )

Forked from PyTorch implementation of "Variational image compression with a scale hyperprior" by Jiaheng Liu ( https://github.com/liujiaheng/compression )

This code is experimental.

Requirements

TODO torchac should be switched to the standalone release on https://github.com/fab-jul/torchac (which was not yet released at the time of writing this code)

Arch

pacaur -S python-tqdm python-pytorch-torchac python-configargparse python-yaml python-ptflops python-colorspacious python-pypng python-pytorch-piqa-git

Ubuntu / Slurm cluster / misc:

TMPDIR=tmp pip3 install --user torch==1.7.0+cu92 torchvision==0.8.1+cu92 -f https://download.pytorch.org/whl/torch_stable.html
TMPDIR=tmp pip3 install --user tqdm matplotlib tensorboardX scipy scikit-image scikit-video ConfigArgParse pyyaml h5py ptflops colorspacious pypng piqa

torchac must be compiled and installed per https://github.com/trougnouf/L3C-PyTorch/tree/master/src/torchac

torchac $ COMPILE_CUDA=auto python3 setup.py build
torchac $ python3 setup.py install --optimize=1 --skip-build

or (untested)

torchac $ pip install .

Once Ubuntu updates PyTorch then tensorboardX won't be required

Dataset gathering

Copy the kodak dataset into datasets/test/kodak

cd ../common
python tools/wikidownloader.py --category "Category:Featured pictures on Wikimedia Commons"
python tools/wikidownloader.py --category "Category:Formerly featured pictures on Wikimedia Commons"
python tools/wikidownloader.py --category "Category:Photographs taken on Ektachrome and Elite Chrome film"
mv "../../datasets/Category:Featured pictures on Wikimedia Commons" ../../datasets/FeaturedPictures
mv "../../datasets/Category:Formerly featured pictures on Wikimedia Commons" ../../datasets/Formerly_featured_pictures_on_Wikimedia_Commons
mv "../../datasets/Category:Photographs taken on Ektachrome and Elite Chrome film" ../../datasets/Photographs_taken_on_Ektachrome_and_Elite_Chrome_film
python tools/verify_images.py ../../datasets/FeaturedPictures/
python tools/verify_images.py ../../datasets/Formerly_featured_pictures_on_Wikimedia_Commons/
python tools/verify_images.py ../../datasets/Photographs_taken_on_Ektachrome_and_Elite_Chrome_film/

# TODO make a list of train/test img automatically s.t. images don't have to be copied over the network

Crop images to 1024*1024. from src/common: (in python)

import os
from libs import libdsops
for ads in ['Formerly_featured_pictures_on_Wikimedia_Commons', 'Photographs_taken_on_Ektachrome_and_Elite_Chrome_film', 'FeaturedPictures']:
    libdsops.split_traintest(ads)
    libdsops.crop_ds_dpath(ads, 1024, root_ds_dpath=os.path.join(libdsops.ROOT_DS_DPATH, 'train'), num_threads=os.cpu_count()//2)

#verify crops
python3 tools/verify_images.py ../../datasets/train/resized/1024/FeaturedPictures/
python3 tools/verify_images.py ../../datasets/train/resized/1024/Formerly_featured_pictures_on_Wikimedia_Commons/
python3 tools/verify_images.py ../../datasets/train/resized/1024/Photographs_taken_on_Ektachrome_and_Elite_Chrome_film/
# use the --save_img flag at the end of verify_images.py commands if training fails after the simple verification

Move a small subset of the training cropped images to a matching test directory and use it as args.val_dpath

JPEG/BPG compression of the Commons Test Images is done with common/tools/bpg_jpeg_compress_commons.py and comp/tools/bpg_jpeg_test_commons.py

Loading

Loading a model: provide all necessary (non-default) parameters s.a. arch, num_distributions, etc. Saved yaml can be used iff the ConfigArgParse patch from https://github.com/trougnouf/ConfigArgParse is applied, otherwise unset values are overwritten with the "None" string.

Training

Train a base model (given arch and num_distributions) for 6M steps at train_lambda=4096, fine-tune for 4M steps with lower train_lambda and/or msssim lossf Set arch to Manypriors for this work, use num_distributions 1 for Balle2017, or set arch to Balle2018PTTFExp for Balle2018 (hyperprior) egrun:

python train.py --num_distributions 64 --arch ManyPriors --train_lambda 4096 --expname mse_4096_manypriors_64_CLI
# and/or
python train.py --config configs/mse_4096_manypriors_64pr.yaml
# and/or
python train.py --config configs/mse_2048_manypriors_64pr.yaml --pretrain mse_4096_manypriors_64pr --reset_lr --reset_global_step # --reset_optimizer
# and/or
python train.py --config configs/mse_4096_hyperprior.yaml

--passthrough_ae is now activated by default. It was not used in the paper, but should result in better rate-distortion. To turn it off, change config/defaults.yaml or use --no_passthrough_ae

Tests

egruns: Test complexity:

python tests.py --complexity --pretrain mse_4096_manypriors_64pr --arch ManyPriors --num_distributions 64

Test timing:

python tests.py --timing "../../datasets/test/Commons_Test_Photographs" --pretrain mse_4096_manypriors_64pr --arch ManyPriors --num_distributions 64

Segment the images in commons_test_dpath by distribution index:

python tests.py --segmentation --commons_test_dpath "../../datasets/test/Commons_Test_Photographs" --pretrain mse_4096_manypriors_64pr --arch ManyPriors --num_distributions 64

Visualize cumulative distribution functions:

python tests.py --plot --pretrain mse_4096_manypriors_64pr --arch ManyPriors --num_distributions 64

Test on kodak images:

python tests.py --encdec_kodak --test_dpath "../../datasets/test/kodak/" --pretrain mse_4096_manypriors_64pr --arch ManyPriors --num_distributions 64

Test on commons images (larger, uses CPU):

python tests.py --encdec_commons --test_commons_dpath "../../datasets/test/Commons_Test_Photographs/" --pretrain checkpoints/mse_4096_manypriors_64pr/saved_models/checkpoint.pth --arch ManyPriors --num_distributions 64

Encode an image:

python tests.py --encode "../../datasets/test/Commons_Test_Photographs/Garden_snail_moving_down_the_Vennbahn_in_disputed_territory_(DSCF5879).png" --pretrain mse_4096_manypriors_64pr --arch ManyPriors --num_distributions 64 --device -1

Decode that image:

python tests.py --decode "checkpoints/mse_4096_manypriors_64pr/encoded/Garden_snail_moving_down_the_Vennbahn_in_disputed_territory_(DSCF5879).png" --pretrain mse_4096_manypriors_64pr --arch ManyPriors --num_distributions 64 --device -1

Reference PyTorch implementation of "End-to-end optimized image compression with competition of prior distributions"

Related tags

Overview

Requirements

Arch

Ubuntu / Slurm cluster / misc:

Dataset gathering

Loading

Training

Tests

Owner

Benoit Brummer

Code for the paper "On the Power of Edge Independent Graph Models"

NeurIPS workshop paper 'Counter-Strike Deathmatch with Large-Scale Behavioural Cloning'

Binary Stochastic Neurons in PyTorch

Implementation of paper "Self-supervised Learning on Graphs:Deep Insights and New Directions"

Official implementation of NeurIPS 2021 paper "Contextual Similarity Aggregation with Self-attention for Visual Re-ranking"

Awesome Remote Sensing Toolkit based on PaddlePaddle.

TensorFlow Ranking is a library for Learning-to-Rank (LTR) techniques on the TensorFlow platform

UPSNet: A Unified Panoptic Segmentation Network

DivNoising is an unsupervised denoising method to generate diverse denoised samples for any noisy input image. This repository contains the code to reproduce the results reported in the paper https://openreview.net/pdf?id=agHLCOBM5jP

For IBM Quantum Challenge Africa 2021, 9 September (07:00 UTC) - 20 September (23:00 UTC).

4D Human Body Capture from Egocentric Video via 3D Scene Grounding

FSL-Mate: A collection of resources for few-shot learning (FSL).

PyTorch implementation of the Quasi-Recurrent Neural Network - up to 16 times faster than NVIDIA's cuDNN LSTM

Numenta published papers code and data

This repository contains the code for the CVPR 2021 paper "GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields"

Auto-Lama combines object detection and image inpainting to automate object removals

An pytorch implementation of Masked Autoencoders Are Scalable Vision Learners

Official Implementation for "StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery" (ICCV 2021 Oral)

Pseudo-mask Matters in Weakly-supervised Semantic Segmentation