CausalGAN: Learning Causal Implicit Generative Models with Adversarial Training

Authors: Murat Kocaoglu, Christopher Snyder, Alexandros G. Dimakis, Sriram Vishwanath

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We consider the application of conditional and interventional sampling of face images with binary feature labels, such as mustache, young. We evaluate our Ci GM training framework on the labeled Celeb A data (Liu et al. (2015)). We empirically show that Causal GAN and Causal BEGAN can produce label-consistent images even for label combinations realized under interventions that never occur during training, e.g., "woman with mustache"2.
Researcher Affiliation Academia Department of Electrical and Computer Engineering The University of Texas at Austin Austin, TX, USA mkocaoglu@utexas.edu,22csnyder@gmail.com, dimakis@austin.utexas.edu,sriram@austin.utexas.edu
Pseudocode No The paper describes model architectures and training procedures in text and mathematical formulas, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an unambiguous statement or a direct link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We evaluate our Ci GM training framework on the labeled Celeb A data (Liu et al. (2015)).
Dataset Splits No The paper mentions using the Celeb A dataset but does not explicitly provide details about training, validation, and test dataset splits, such as percentages or sample counts.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper mentions using DCGAN and Wasserstein GAN but does not provide specific software dependency details such as library names with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x).
Experiment Setup Yes For the training, we used 25 Wasserstein discriminator (critic) updates per generator update, with a learning rate of 0.0008. Compared to DCGAN an important distinction is that we make 6 generator updates for each discriminator update on average. We use the same learning rate (0.00008) for both the generator and discriminator and do 1 update of each simultaneously (calculating the for each before applying either). We simply use γ1 = γ2 = γ3 = 0.5. We do use customized margin learning rates λ1 = 0.001, λ2 = 0.00008, λ3 = 0.01