OSOA: One-Shot Online Adaptation of Deep Generative Models for Lossless Compression

Authors: Chen Zhang, Shifeng Zhang, Fabio Maria Carlucci, Zhenguo Li

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that vanilla OSOA can save significant time versus training bespoke models and space versus using one model for all targets.
Researcher Affiliation Industry Chen Zhang Shifeng Zhang Fabio M. Carlucci Zhenguo Li Huawei Noah s Ark Lab {chenzhang10, zhangshifeng4, li.zhenguo}@huawei.com
Pseudocode Yes Algorithm 1 One Shot Online Adaptation: Encoding and Decoding and Algorithm 2 The encode_or_cache method in OSOA Encoding
Open Source Code No The paper does not contain an explicit statement or link providing access to open-source code for the described methodology.
Open Datasets Yes The datasets for base model pretraining are the renowned natural image datasets CIFAR10 [28] and Image Net32 [7], including images of size 32 32. We obtain three target datasets randomly sampled from the large image dataset Yahoo Flickr Creative Commons 100 Million (YFCC100m) [46] to test the compression performance.
Dataset Splits Yes The data splitting strategy is the same as Stage 2. For Fine Tune v1, we fine tune the pretrained model for 2 epochs... For Fine Tune v2, we fine tune the pretrained model for 4 epochs for Hi LLo C (and IAF RVAE) and 3 epochs for IDF++... For Fine Tune v3, we fine tune the pretrained model for 20 epochs... We quadruple the batch size as the image size decreases, i.e., batch size 256/64/16 in Hi LLo C and batch size 48/12/3 in IDF++, for SET32/64/128 respectively.
Hardware Specification Yes We use an Nvidia V100 32GB GPU for Hi LLo C (and IAF RVAE) and an Nvidia V100 16 GB GPU for IDF++.
Software Dependencies Yes The time ratio we measured with/without the determinism is 1.98 (Hi LLo C) in Tensor Flow 1.14 [4] with tensorflow-determinism 0.3.0 [3] and 1.34 (IDF++) in Pytorch 1.6 [2].
Experiment Setup Yes For Fine Tune v1, we fine tune the pretrained model for 2 epochs, as the whole OSOA Encoding & Decoding procedures involve 2 epochs of adaptations in total. For Fine Tune v2, we fine tune the pretrained model for 4 epochs for Hi LLo C (and IAF RVAE) and 3 epochs for IDF++... For Fine Tune v3, we fine tune the pretrained model for 20 epochs... We quadruple the batch size as the image size decreases, i.e., batch size 256/64/16 in Hi LLo C and batch size 48/12/3 in IDF++, for SET32/64/128 respectively.