Principled Knowledge Extrapolation with GANs
Authors: Ruili Feng, Jie Xiao, Kecheng Zheng, Deli Zhao, Jingren Zhou, Qibin Sun, Zheng-Jun Zha
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present several discoveries from our proposed Principal Knowledge Descent (PKD) method and extrapolation results of state-of-the-art GANs, including Big GAN256-Deep (Brock et al., 2018), Style GAN2 (Karras et al., 2019; 2020b) on FFHQ faces (Karras et al., 2019), and Style GAN2-ADA (Karras et al., 2020a) on Bre Ca HAD (Aksac et al., 2019) which contains breath cancer slices. The details of the experiments are reported in Appendix Sec. B, including choices of the hyper-parameters λ, ϵ, K, m, sources of all pretrained models (i.e., generators and posterior estimation models), and dataset information. In Tab. 1, we report numerical metrics (i.e., Fr echet Inception Distance (FID), Inception Score (IS), and Path Length as in (Karras et al., 2019)) of synthesis quality after knowledge extrapolation, indicating that the influence to synthesis quality is little. |
| Researcher Affiliation | Collaboration | Ruili Feng 1 Jie Xiao 1 Kecheng Zheng 1 Deli Zhao 2 Jingren Zhou 3 Qibin Sun 1 Zheng-Jun Zha 1 1University of Science and Technology of China, Hefei, China 2Ant Research, Hangzhou, China 3Alibaba Group, Hangzhou, China. |
| Pseudocode | Yes | Algorithm 1 Principal Knowledge Descent (PKD) |
| Open Source Code | No | Generative Model Choice For FFHQ data domain, we use the pretrained Style GAN2 generator offered by Awesome Pretrained Style GAN2 2 with config-f and 512 512 resolution. For Bre Ca HAD data domain, we use the official pretrained Style GAN2-ADA generator3. For Image Net data domain, we use the Big GAN256-Deep model in the official TFhub repository 4. Posterior Estimation Model Choice For FFHQ data domain, we use the official pretrained Res Net50 classifiers provided by Style GAN2 authors 5 as the posterior distribution Pl. We use the official code provided by Style Space Analysis authors 7. Footnotes 2-7 provide links to other projects' code/models, not the authors' own implementation of PKD. |
| Open Datasets | Yes | We report knowledge extrapolation results of our PKD method on Image Net (Deng et al., 2009) data domain and FFHQ (Karras et al., 2019) face data domain in Fig. 3 and 5, respectively. In FFHQ facial images domain, we use the Style GAN2 model as the pretrained generator, and Res Net50 classifiers trained on Celeb A-HQ (Karras et al., 2018) annotations for facial attributes like mustache , lipstick , gray hair as the posterior distribution for knowledge of interest. We further conduct Dirac knowledge extrapolation in the Bre Ca HAD data domain. This dataset consists of 162 slice images and each of them has annota-tions for cancer nuclear. |
| Dataset Splits | No | We train a Res Net50 classifier to infer the posterior probability for cancer severity of a tissue slice image on the few-shot annotation images, and use the Style GAN2-ADA model trained on this dataset as the pretrained generator. We halt the training at the error rate of 10% in the training set to avoid overfitting, as the training dataset is small. The whole training data to support our method is composed of merely 162 annotated slice images. While training is mentioned, specific percentages or sample counts for train/validation/test splits are not provided. |
| Hardware Specification | Yes | To train Style GAN2 on 512 512 resolution FFHQ dataset and Inter Face GAN semantic boundaries, we use 8 NVIDIA V100 GPUs. To train the Res Net50 regressive model for Bre Ca HAD data domain, we use 1 NVIDIA GTX1080Ti GPU. For all the remaining experiments of knowledge extrapolation, we use 1 NVIDIA V100 GPU. |
| Software Dependencies | No | For Image Net data domain, we use the official Res Net50 classifier provided by Tensor Flow 6. The paper mentions TensorFlow but does not provide specific version numbers for TensorFlow or any other software libraries used in the experiments. |
| Experiment Setup | Yes | Hyper-parameter Setting For Lipstick extrapolation of FFHQ data domain, we set K = 4; for Gray Hair extrapolation of FFHQ data domain, we set K = 7. For all the other experiments, we set K = 10. For FFHQ domain and Bre Ca HAD domain, we set ϵ = 1e 3; for Image Net domain, we set ϵ = 1e 5. The choice of λ is set according to Fig. 4, where we choose a λ that is slightly smaller than λmax for each experiment. For FFHQ domain, we set λ = 1.8e 4; for Bre Ca HAD domain, we set λ = 4.8e 4; for Image Net domain, we set λ = 1.3e 4. For all Dirac Knowledge Extrapolations, we set ξ = 0.01. |