Low Distortion Block-Resampling with Spatially Stochastic Networks

Authors: Sarah Hong, Martin Arjovsky, Darryl Barnhart, Ian Thompson

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In section 4 we perform both qualitative and qualitative experiments showing the workings and excellent performance of SSNs. We experiment with the FFHQ [18] faces and the LSUN churches [39] datasets at a resolution of 256 256 pixels. We provide both quantitative and qualitative experiments.
Researcher Affiliation Collaboration Sarah Jane Hong Latent Space sarah@latentspace.co Martin Arjovsky École Normale Supérieure martinarjovsky@gmail.com Darryl Barnhart Latent Space darryl@latentspace.co Ian Thompson Latent Space ian@latentspace.co
Pseudocode No The paper describes the architecture and algorithm through text and diagrams (Figure 1, Figure 3), but does not include any formal pseudocode blocks or algorithms.
Open Source Code No The paper cites external GitHub repositories for pretrained StyleGAN2 models ([1], [2]) that were used in the work, but does not provide a link or explicit statement for the release of its own methodology's source code.
Open Datasets Yes We experiment with the FFHQ [18] faces and the LSUN churches [39] datasets at a resolution of 256 256 pixels.
Dataset Splits No The paper mentions using datasets for experiments but does not provide specific details on train/validation/test splits, percentages, or the methodology used for data partitioning.
Hardware Specification Yes training a single Style GAN2 model from scratch on LSUN churches takes 781 GPU hours on V100s
Software Dependencies No The paper mentions using StyleGAN2 and SPADE-like architectures, but does not specify version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or specific dependencies used in their implementation.
Experiment Setup Yes The latent code has dimension z R4 4 512 for SSN and z R1 1 512) as per Style GAN2 s default configuration. We also utilize skip connections, the general architecture, and the R1 and path length regularization (with weights of 1 and 2 respectively) of [18]. We chose the hyperparameter of λD = 100 for our qualitative experiments