Learning Texture Manifolds with the Periodic Spatial GAN

Authors: Urs Bergmann, Nikolay Jetchev, Roland Vollgraf

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We make multiple experiments which show that PSGANs can flexibly handle diverse texture and image data sources, and the method is highly scalable and can generate output images of arbitrary large size. 3. Experiments
Researcher Affiliation Industry 1Zalando Research, Berlin. Correspondence to: Urs Bergmann <urs.bergmann@zalando.de>, Nikolay Jetchev <nikolay.jetchev@zalando.de>, Roland Vollgraf <roland.vollgraf@zalando.de>.
Pseudocode No The paper describes its methods but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our source code is available at https://github.com/ubergmann/psgan
Open Datasets Yes The following image sources were used for the experiments in this paper: the Oxford Describable Textures Dataset (DTD) (Cimpoi et al., 2014), which is composed of various categories, each containing 120 images; the Facades dataset (Radim Tyleˇcek, 2013), which contains 500 facades of different houses in Prague. Both datasets comprise objects of different scales and sizes. We also used satellite images of Sydney from Google Maps.
Dataset Splits No The paper mentions training parameters like 'minibatch size of 25' and 'typical image patch size was 160x160 pixels' but does not specify explicit dataset splits (e.g., percentages or counts) for training, validation, or testing.
Hardware Specification Yes On our hardware (Theano and Nvidia Tesla K80 GPU) we measured 0.006 seconds for the generation of a 256x256 pixels image and 0.26 seconds for a 2048x2048 pixels image.
Software Dependencies No The paper mentions 'Theano' as a software framework used, but it does not specify any version numbers for Theano or other software dependencies required to reproduce the experiments.
Experiment Setup Yes Training was done with ADAM (Kingma & Ba, 2014) with the settings of (Radford et al., 2015) learning rate 0.0002, minibatch size of 25. The typical image patch size was 160x160 pixels. We usually used 5 layers in G and D (see Table 1), kernels of size 5x5 with zero padding, and batch normalization. ... We used dh = 60 for the experiments. All parameters are initialized from an independent random Gaussian distribution N(0, 0.02), except b1 and b2, which have a non-zero mean N(c, 0.02c).