MaGNET: Uniform Sampling from Deep Generative Network Manifolds Without Retraining

Authors: Ahmed Imtiaz Humayun, Randall Balestriero, Richard Baraniuk

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform a range of experiments on several datasets and DGNs, e.g., for the state-of-the-art Style GAN2 trained on the FFHQ dataset, uniform sampling via Ma GNET increases distribution precision by 4.1% and recall by 3.0% and decreases gender bias by 41.2%, without requiring labels or retraining.
Researcher Affiliation Academia Ahmed Imtiaz Humayun Rice University imtiaz@rice.edu Randall Balestriero Rice University randallbalestriero@gmail.com Richard Baraniuk Rice University richb@rice.edu
Pseudocode Yes Algorithm 1: Ma GNET Sampling as described in Sec. 3.2
Open Source Code Yes Colab and codes at bit.ly/magnet-sampling
Open Datasets Yes For example, the Celeb A dataset contains a large fraction of smiling faces.
Dataset Splits No The paper refers to using existing datasets for training and evaluation but does not explicitly provide details about the specific training, validation, and test dataset splits used for reproducibility.
Hardware Specification Yes All the experiments were run on a Quadro RTX 8000 GPU, which has 48 GB of high-speed GDDR6 memory and 576 Tensor cores.
Software Dependencies Yes In short, we employed TF2 (2.4 at the time of writing), all the usual Python scientific libraries such as Num Py and Py Torch.
Experiment Setup Yes For Style GAN2, we use the official config-e provided in the Git Hub Style GAN2 repo1, unless specified. We use the recommended default of ψ = 0.5 as the interpolating stylespace truncation, to ensure generation quality of faces for the qualitative experiments.