Counterfactual Generative Networks

Authors: Axel Sauer, Andreas Geiger

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the ability of our model to generate such images on MNIST and Image Net. Further, we show that the counterfactual images can improve out-of-distribution robustness with a marginal drop in performance on the original classification task, despite being synthetic. Lastly, our generative model can be trained efficiently on a single GPU, exploiting common pre-trained models as inductive biases.
Researcher Affiliation Academia Axel Sauer1,2 & Andreas Geiger1,2 Autonomous Vision Group 1Max Planck Institute for Intelligent Systems, T ubingen 2University of T ubingen {firstname.lastname}@tue.mpg.de
Pseudocode No The paper describes its methods verbally and through architectural diagrams (e.g., Figure 2) but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes We release our code at https://github.com/autonomousvision/counterfactual generative networks
Open Datasets Yes We demonstrate the ability of our model to generate such images on MNIST and Image Net.
Dataset Splits No The paper mentions training and testing, and uses metrics like Inception Score during training, but does not explicitly detail a separate 'validation' dataset split with percentages or counts for reproducibility.
Hardware Specification Yes whole CGN on a single NVIDIA GTX 1080Ti within 12 hours
Software Dependencies No We use a Res Net-50 from Py Torch torchvision. We use the pre-trained Big GAN models from https://github.com/huggingface/ pytorch-pretrained-Big GAN. We use Adam (Kingma & Ba, 2014).
Experiment Setup Yes We use the following lambdas: λ1 = 100, λ2 = 5, λ3 = 300, λ4 = 500, λ5 = 5, λ6 = 2000. For the optimization we use Adam (Kingma & Ba, 2014), and set the learning rate of fshape to 8e-6, and for both ftext and fbg to 1e-5. We train for 70 episodes with Stochastic Gradient Descent using a batch size of 512. Of the 512 images, 256 are real images, 256 are counterfactual images. We use a momentum of 0.9, weight decay (1e-4), and a learning rate of 0.1, multiplied by a factor of 0.001 after 30 and 60 epochs.