Optimally Controllable Perceptual Lossy Compression
Authors: Zeyu Yan, Fei Wen, Peilin Liu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we validate our theoretical finding and demonstrate the superiority of our frameworks via experiments. Experiments on MNIST, depth images and RGB images, which validate our theoretical finding and demonstrate the performance of our frameworks. |
| Researcher Affiliation | Academia | 1Brain-inspired Application Technology Center (BATC), School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China. Correspondence to: Fei Wen <wenfei@sjtu.edu.cn>. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at: https://github.com/Zeyu Yan/Controllable Perceptual-Compression |
| Open Datasets | Yes | Experiments on MNIST, depth images and RGB images. We provide experiment results on MNIST (Lecun et al., 1998), depth and RGB images. We further provide results on the SUNCG dataset (Song et al., 2017). In the experiment on RGB images, ... We train Gd, Gp, Gh on COCO2014 (Lin et al., 2014) in two different bit-rates... |
| Dataset Splits | No | The paper trains models on datasets like MNIST, SUNCG, and COCO2014 and evaluates on KODAK, but it does not specify explicit training/validation/test splits (e.g., percentages or sample counts) for these datasets, nor does it refer to predefined standard splits for reproduction beyond the datasets themselves. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like WGAN-gp and Hi Fi C, but it does not specify their version numbers or versions for other key software dependencies like programming languages or deep learning frameworks. |
| Experiment Setup | Yes | We first pretrain an encoder-decoder pair (Ed, Gd) for 100 epochs by MMSE with regime set as low. Then our generator Gp is trained by (12) with λ = 0.005. For comparison, we train the Hi Fi C model for 100 epochs using the same hyper-parameters as in (Mentzer et al., 2020), denoted by Gh. In the experiment on RGB images, the networks are set the same as the depth image experiment, except the input channel number is 3 and λ = 0.01. We train Gd, Gp, Gh on COCO2014 (Lin et al., 2014) in two different bit-rates, with regime set as low and high. |