Finite-State Autoregressive Entropy Coding for Efficient Learned Lossless Compression

Authors: Yufeng Zhang, Hang Yu, Jianguo Li, Weiyao Lin

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that the proposed lossless compression method could improve the compression ratio by up to 6% compared to the baseline, with negligible extra computational time. and 6 EXPERIMENTS 6.1 DATASETS AND METRICS In our experiments, we focused on compressing and decompressing image datasets, specifically CIFAR10 (CF10) (Krizhevsky, 2009) and Image Net32/64 (IN32, IN64) (Deng et al., 2009).
Researcher Affiliation Collaboration Yufeng Zhang1,2 , Hang Yu2 , Jianguo Li2 , Weiyao Lin1 1Shanghai Jiao Tong University, 2Ant Group
Pseudocode Yes Algorithm 1 Finite-State Autoregressive t ANS Coding Algorithm. The steps that are different from t ANS are highlighted in red.
Open Source Code Yes Code is available at https://github.com/alipay/Finite_ State_Autoregressive_Entropy_Coding.
Open Datasets Yes In our experiments, we focused on compressing and decompressing image datasets, specifically CIFAR10 (CF10) (Krizhevsky, 2009) and Image Net32/64 (IN32, IN64) (Deng et al., 2009). and We calculate the BPD based on the length of the compression bitstream to evaluate practical compression performance. For a qualitative understanding of the results, we present the BPD versus Speed comparison for different methods on CIFAR10 in Figure 1. For complete numerical results, please refer to Appendix E.2.
Dataset Splits No The paper mentions the datasets used (CIFAR10, ImageNet32/64) and training epochs/batch sizes, but it does not specify the explicit training, validation, and test dataset splits (e.g., percentages or sample counts) used for their experiments.
Hardware Specification Yes For CPU-based experiments, we utilize a desktop machine equipped with an Intel i7-6800K CPU. For GPU-based model training, we employ a virtual machine on a Kubermaker cluster featuring 8 NVIDIA V100 GPUs.
Software Dependencies No The paper mentions various libraries and tools used (e.g., zstd, craystack, Pillow, imageio-flif) in footnotes but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Regarding FSAR, each variable relies on its previous neighboring variables in the spatial domain. Specifically, the latent prior utilizing the Order-1 Markov model is defined as p(yi,j,k|yi,j,k 1), while Order-2 is p(yi,j,k|yi,j 1,k, yi,j,k 1), and Order-3 is p(yi,j,k|yi,j 1,k 1, yi,j 1,k, yi,j,k 1), where i, j, k represent the indices of the channel, height, and width dimensions, respectively. During training, the Markov models are implemented using a 3-layer network consisting of 3 Linear layers and 2 Re LU layers. For the learnable state number, the initial number of states is set to 256, and α = 1.5 is used in the α-entmax method. For STHQ, the temperature for Gumbel-softmax GSτ is set as a constant 0.5. and The training process utilized the Adam optimizer with a learning rate of 10 3. For the CIFAR10 dataset, the training was performed for 1000 epochs, with a batch size of 64 per GPU. Regarding the Image Net32 dataset, the training was conducted for 50 epochs, with a batch size of 64 per GPU. As for Image Net64, the training setup was identical to Image Net32, except that the batch size was set to 16 per GPU.