A Framework for the Quantitative Evaluation of Disentangled Representations

Authors: Cian Eastwood, Christopher K. I. Williams

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To illustrate the appropriateness of the framework, we employ it to compare quantitatively the representations learned by recent state-of-the-art models. We employ the framework to compare quantitatively the codes learned by PCA, the VAE, β-VAE and Info GAN. The results can be reproduced with our open source implementation.
Researcher Affiliation Academia Cian Eastwood School of Informatics University of Edinburgh, UK c.eastwood@ed.ac.uk Christopher K. I. Williams School of Informatics University of Edinburgh, UK and Alan Turing Institute, London, UK ckiw@inf.ed.ac.uk
Pseudocode No No pseudocode or algorithm blocks found.
Open Source Code Yes The results can be reproduced with our open source implementation. Code and dataset available at https://www.github.com/cianeastwood/qedr.
Open Datasets Yes We use the graphics renderer described in (Moreno et al., 2016) to generate 200,000 64 64 colour images of an object (teapot) with varying pose and colour (see Figure 2). ... Code and dataset available at https://www.github.com/cianeastwood/qedr.
Dataset Splits Yes We divide the images into training (160,000), validation (20,000) and test (20,000) sets before removing images which contain particular generative factor combinations to faciliate the evaluation of zeroshot performance (see Appendix B.2). This left 142,927, 17,854 and 17,854 images in the training, validation and test sets respectively.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models or types) are mentioned for running experiments.
Software Dependencies No No specific software dependencies with version numbers (e.g., library names with versions) are explicitly listed.
Experiment Setup Yes For all generative models, we use the Res Net architectures shown in Table 2 for the encoder / discriminatior (D) / auxilary network (Q) and the decoder / generator (G). We optimize using Adam (Kingma & Ba, 2014) with a learning rate of 1e-4 and a batch size of 64.