Learning to Discover Cross-Domain Relations with Generative Adversarial Networks

Authors: Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, Jiwon Kim

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 3. Experiments To empirically demonstrate our explanations on the differences between a standard GAN, a GAN with reconstruction loss and our proposed model (Disco GAN), we designed an illustrative experiment based on synthetic data in 2-dimensional domains A and B.
Researcher Affiliation Industry 1SK T-Brain, Seoul, South Korea. Correspondence to: Taeksoo Kim <jazzsaxmafia@sktbrain.com>.
Pseudocode No The paper describes the model architecture and loss functions, but no structured pseudocode or algorithm blocks are provided.
Open Source Code No No statement regarding the release of source code or a link to a code repository was found in the paper.
Open Datasets Yes We used a Car dataset (Fidler et al., 2012)... Next, we use a Face dataset (Paysan et al., 2009)... We also applied the face attribute conversion task on Celeb A and Facescrub dataset (Liu et al., 2015; Ng & Winkler, 2014)... 3D rendered images of chair (Aubry et al., 2014)... generate realistic photos of handbags (Zhu et al., 2016) and shoes (Yu & Grauman, 2014).
Dataset Splits No The paper mentions splitting datasets into 'train set and test set' and further splitting the train set for domain A and B samples, but it does not specify a validation split or provide exact percentages or sample counts for any of the splits.
Hardware Specification Yes All computations were conducted on a single machine with an Nvidia Titan X Pascal GPU and an Intel(R) Xeon(R) E5-1620 CPU.
Software Dependencies No The paper mentions using the Adam optimizer and Batch Normalization, but it does not specify versions for these or any other software libraries or frameworks used.
Experiment Setup Yes In each real domain experiment, all input images and translated images were size of 64 64 3. For training, we employed learning rate of 0.0002 and used the Adam optimizer (Kingma & Ba, 2015) with β1 = 0.5 and β2 = 0.999. We applied Batch Normalization (Ioffe & Szegedy, 2015) to all convolution and deconvolution layers except the first and the last layers, and applied weight decay regularization coefficient of 10 4 and minibatch of size 200.