Deceive D: Adaptive Pseudo Augmentation for GAN Training with Limited Data
Authors: Liming Jiang, Bo Dai, Wayne Wu, Chen Change Loy
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the effectiveness of APA in improving synthesis quality in the low-data regime. We provide a theoretical analysis to examine the convergence and rationality of our new training strategy. APA is simple and effective. It can be added seamlessly to powerful contemporary GANs, such as Style GAN2, with negligible computational cost. Code: https://github.com/Endless Sora/Deceive D. 4 Experiments Datasets. We use four datasets in our main experiments: Flickr-Faces-HQ (FFHQ) [19] with 70,000 human face images, AFHQ-Cat [8] with 5,153 cat faces, Caltech-UCSD Birds-200-2011 (CUB) [43] with 11,788 images of birds, and Danbooru2019 Portraits (Anime) [4] with 302,652 anime portraits. Evaluation metrics. We follow the standard evaluation protocol [5, 42] for the quantitative evaluation. Specifically, we use the Fr echet Inception Distance (FID, lower is better) [11], which quantifies the distance between distributions for the real and generated images. We also apply the Inception Score (IS, higher is better) [38]. 4.1 The Effectiveness of APA |
| Researcher Affiliation | Collaboration | Liming Jiang1 Bo Dai1 Wayne Wu2 Chen Change Loy1 1S-Lab, Nanyang Technological University 2Sense Time Research {liming002, bo.dai, ccloy}@ntu.edu.sg wuwenyan@sensetime.com |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code: https://github.com/Endless Sora/Deceive D. |
| Open Datasets | Yes | Datasets. We use four datasets in our main experiments: Flickr-Faces-HQ (FFHQ) [19] with 70,000 human face images, AFHQ-Cat [8] with 5,153 cat faces, Caltech-UCSD Birds-200-2011 (CUB) [43] with 11,788 images of birds, and Danbooru2019 Portraits (Anime) [4] with 302,652 anime portraits. |
| Dataset Splits | No | The paper mentions using subsets of datasets and calculating FID for models using 50k generated images and all real images, but it does not specify explicit train/validation/test splits of the data used for training itself. |
| Hardware Specification | Yes | All the models are trained on 8 NVIDIA Tesla V100 GPUs. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies (e.g., Python version, PyTorch/TensorFlow version). |
| Experiment Setup | Yes | Implementation details. We choose the state-of-the-art Style GAN2 [20] as the backbone to verify the effectiveness of APA on limited data. We use the default setups of APA provided in Section 3.1 unless specified otherwise. For a fair and controllable comparison, we reimplement all baselines and run the experiments from scratch using official code. ... We set a threshold t (in most cases of our experiments, t = 0.6) and initialize p to be zero. If λ signifies too much/little overfitting regarding t (i.e., larger/smaller than t), the probability p will be increased/decreased by one fixed step. Using this step size, p can increase from zero to one in 500k images shown to D. We adjust p once every four iterations and clamp p from below to zero after each adjustment. |