Differentiable Augmentation for Data-Efficient GAN Training
Authors: Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, Song Han
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate consistent gains of our method over a variety of GAN architectures and loss functions for both unconditional and class-conditional generation. With Diff Augment, we achieve a state-of-the-art FID of 6.80 with an IS of 100.8 on Image Net 128 128 and 2-4 reductions of FID given 1,000 images on FFHQ and LSUN. Furthermore, with only 20% training data, we can match the top performance on CIFAR-10 and CIFAR-100. |
| Researcher Affiliation | Collaboration | Shengyu Zhao IIIS, Tsinghua University and MIT Zhijian Liu MIT Ji Lin MIT Jun-Yan Zhu Adobe and CMU Song Han MIT |
| Pseudocode | No | The paper includes a diagram (Figure 4) illustrating the data flow for Diff Augment but does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/mit-han-lab/data-efficient-gans. |
| Open Datasets | Yes | We conduct extensive experiments on Image Net [6], CIFAR-10 [19], CIFAR-100, FFHQ [17], and LSUN-Cat [46] based on the leading class-conditional Big GAN [2] and unconditional Style GAN2 [18]. |
| Dataset Splits | Yes | IS and FID are measured using 10k samples; the validation set is used as the reference distribution for FID calculation. [...] IS and FID are measured using 50k samples; the validation set is used as the reference distribution for FID. [...] FID is measured using 50k generated samples; the full training set is used as the reference distribution. |
| Hardware Specification | Yes | Research supported with Cloud TPUs from Google s Tensor Flow Research Cloud (TFRC). |
| Software Dependencies | No | The paper mentions 'Google s Tensor Flow Research Cloud (TFRC)' but does not specify software versions for TensorFlow or any other libraries. |
| Experiment Setup | Yes | For Diff Augment, we adopt Translation + Cutout for the Big GAN models, Color + Cutout for Style GAN2 with 100% data, and Color + Translation + Cutout for Style GAN2 with 10% or 20% data. [...] As Style GAN2 adopts the R1 regularization [27] to stabilize training, we increase its strength from γ = 0.1 to up to 104 and plot the FID curves in Figure 8b. While we initially find that γ = 0.1 works best under the 100% data setting, the choice of γ = 103 boosts its performance from 34.05 to 26.87 under the 10% data setting. |