Towards Efficient Image Compression Without Autoregressive Models

Authors: Muhammad Salman Ali, Yeongwoong Kim, Maryam Qamar, Sung-Chang Lim, Donghyun Kim, Chaoning Zhang, Sung-Ho Bae, Hui Yong Kim

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We train all our models on the Vimeo-90k dataset Xue et al. [2019]... We tested our model on a commonly used Kodak lossless images dataset Kodak [1993]... Figure 1: Performance-complexity tradeoff using various entropy models... Figure 7: RD rate comparison... Table 1: Average encoding and decoding time... Ablation Studies: Comprehensive ablation studies regarding various mask types, mask sizes, and α values are presented in the supplementary material.
Researcher Affiliation Academia Muhammad Salman Ali 1, Yeongwoong Kim1, Maryam Qamar 1, Sung-Chang Lim2, Donghyun Kim2, Chaoning Zhang1, Sung-Ho Bae 1, Hui Yong Kim 1 1 Kyung Hee University, Republic of Korea 2 Electronics and Telecommunications Research Institute (ETRI), Republic of Korea
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper states, 'We perform all our experiments on the Pytorch framework Paszke et al. [2017] and use the Compress AI library Bégaint et al. [2020].' but does not provide an explicit statement or link for the authors' own source code for the methodology described.
Open Datasets Yes We train all our models on the Vimeo-90k dataset Xue et al. [2019]... We tested our model on a commonly used Kodak lossless images dataset Kodak [1993]...
Dataset Splits No The paper states it trains on the Vimeo-90k dataset and tests on the Kodak dataset, but it does not specify a separate validation split for either dataset.
Hardware Specification Yes Minnen s and Cheng s models were trained using an NVIDIA 2080Ti, whereas the Swin T model was trained on an NVIDIA 3070Ti due to the transformers high memory requirement.
Software Dependencies No The paper mentions 'Pytorch framework Paszke et al. [2017] and use the Compress AI library Bégaint et al. [2020]' but does not provide specific version numbers for these software components.
Experiment Setup Yes The models were optimized using the Adam optimizer Kingma and Ba [2015] with a batch size of 16 and trained for 1.5 million iterations with a learning rate of 1 10 4 for the first million iterations and then halved every 50,000 iterations till 1.25 million iterations. ... The rate-distortion tradeoff is guided by λ, whose value is contained in the set [0.0009, 0.0018, 0.0035, 0.0067, 0.0130, 0.0250].