Collapse by Conditioning: Training Class-conditional GANs with Limited Data

Authors: Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, Luc Van Gool

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We analyze the observed mode collapse problem in comprehensive experiments on four datasets. Our approach demonstrates outstanding results compared with stateof-the-art methods and established baselines.
Researcher Affiliation Academia Computer Vision Lab (CVL), ETH Zurich, Switzerland {mshahbazi,martin.danelljan,paudel,vangool}@vision.ee.ethz.ch
Pseudocode No The paper includes architectural diagrams (e.g., Fig. 3) and mathematical equations, but it does not contain any clearly labeled pseudocode blocks or algorithms.
Open Source Code Yes The code is available at https: //github.com/mshahbazi72/transitional-c GAN
Open Datasets Yes We use four datasets to evaluate our method: Image Net Carnivores (Liu et al., 2019), CUB-200-2011 (Wah et al., 2011), Food101 (Bossard et al., 2014), and Animal Face (Si & Zhu, 2011).
Dataset Splits No The paper specifies the number of images used for training (e.g., 'each containing between 1170 and 2410 images for training') and how these limited datasets were constructed (e.g., '20 classes and 100 images per class'). It also mentions evaluating using 'real' images, and in the appendix, 'using all the images of corresponding classes, including the additional images not used for training' for class-wise metrics. However, it does not explicitly provide specific percentages or counts for training, validation, and test splits for reproducibility across all experiments.
Hardware Specification No The paper states 'Training is done with a batch size of 64 using 4 GPUs,' but does not provide specific details on the GPU models (e.g., NVIDIA A100, Tesla V100), CPU, or other hardware components used for the experiments.
Software Dependencies No The paper mentions basing their method on 'the official Py Torch implementation of Style GAN2+ADA,' but it does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes Training is done with a batch size of 64 using 4 GPUs. For the transition function, we use Ts = 2k and Te = 4k in all experiments. In the specific case of the Animal Face dataset, we found that clipping the output of the transition function to the maximum value of 0.2 achieves the best results.