Conditional Generation Using Polynomial Expansions

Authors: Grigorios Chrysos, Markos Georgopoulos, Yannis Panagakis

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

Reproducibility Variable Result LLM Response
Research Type Experimental Co PE is evaluated in five tasks (class-conditional generation, inverse problems, edges-to-image translation, image-to-image translation, attribute-guided generation) involving eight datasets. The thorough evaluation suggests that Co PE can be useful for tackling diverse conditional generation tasks.
Researcher Affiliation Academia Grigorios G Chrysos EPFL, Switzerland grigorios.chrysos@epfl.ch Markos Georgopoulos Imperial College London, UK m.georgopoulos@imperial.ac.uk Yannis Panagakis University of Athens, GR yannisp@di.uoa.gr
Pseudocode No The paper does not contain a clearly labeled pseudocode or algorithm block.
Open Source Code Yes The source code of Co PE is available at https://github.com/grigorisg9gr/ polynomial_nets_for_conditional_generation.
Open Datasets Yes To validate the proposed formulation, the following diverse conditional generation tasks are considered: class-conditional generation trained on CIFAR10, Cars196 and SVHN in sec. 4.1 and sec. H.2. ... inverse problems in imaging, e.g., super-resolution and block-inpainting, trained on Cars196 and Celeb A in sec. 4.3. ... Additionally, the source code of Co PE is available at https://github.com/grigorisg9gr/polynomial_nets_ for_conditional_generation.
Dataset Splits Yes We use standard data splits from well-established datasets (i.e. using the wrappers of Pytorch). The source code is publicly available.
Hardware Specification Yes A single GPU was used in each experiment. The GPU is part of our lab cluster.
Software Dependencies No The paper mentions using "wrappers of Pytorch" for standard data splits, but it does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes To avoid such instability, we take the following three steps: a) Co PE samples the noise vector from the uniform distribution, i.e., from the bounded interval of r 1, 1s, b) a hyperbolic tangent is used in the output of the generator as a normalization, i.e., it constrains the outputs in the bounded interval of r 1, 1s, c) batch normalization [Ioffe and Szegedy, 2015] is used to convert the representations to zero-mean. ... Each experiment is conducted five times and the mean and the standard deviation are reported.