Emerging Disentanglement in Auto-Encoder Based Unsupervised Image Content Transfer

Authors: Ori Press, Tomer Galanti, Sagie Benaim, Lior Wolf

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present convincing results in a few visual domains, such as no-glasses to glasses, adding facial hair based on a reference image, etc.
Researcher Affiliation Collaboration Ori Press, Tomer Galanti & Sagie Benaim The School of Computer Science Tel Aviv University [...] Lior Wolf Facebook AI Research & The School of Computer Science Tel Aviv University
Pseudocode No The paper describes the architecture and training losses but does not include any pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement or link indicating that the source code for the described method is publicly available.
Open Datasets Yes Images from the celeb A face image dataset by Yang et al. (2015) were used, since these are conveniently annotated as having the attribute or not.
Dataset Splits No The paper does not explicitly specify validation dataset splits (e.g., percentages or sample counts).
Hardware Specification No The paper does not specify any particular hardware (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No The paper mentions software components like Instance Normalization, Batch Normalization, leaky-ReLUs, and ReLUs, but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Let Ck denote a Convolution Instance Norm-Re LU layer with k filters, where a kernel size of 4 4, with a stride of 2, and a padding of 1 is used. The activations of the encoders e1, e2 are leaky-Re LUs with a slope of 0.2 and the deocder g employs Re LUs. e1 has the following layers C32, C64, C128, C256, C512, C512 d; e2 has a slightly lower capacity C32, C64, C128, C128, C128, Cd, where d = 25. The input images have a size of 128 128, and the encoding is of size 512 2 2 (split between the e1 and e2). g is symmetric to the encoders and employs transposed convolutions for the upsampling. [...] Our method has one weighting hyperparameter, which is fixed throughout the experiments.