Each file is in the pickle format and was generated with Python 3.8.5. Each file contains a Python dictionary with the following fields: 'train_images': 640,000 float32 images used for model training. 20px images contain 400 pixel intensities, 40px images contain 1600 pixel intensities each. 'train_labels': float32 labels for each image in 'train_images'. All natural images are labeled with 0.0. Texture images are labeled with 0.0, 1,0, 2.0, 3.0, or 4.0, according to their texture family. 'test_images': 64,000 float32 images used for model testing. 20px images contain 400 pixel intensities, 40px images contain 1600 pixel intensities each. 'test_labels': float32 labels for each image in 'test_images'. All natural images are labeled with 0.0. Texture images are labeled with 0.0, 1,0, 2.0, 3.0, or 4.0, according to their texture family. For more details, see the paper "Top-down inference in an early visual cortex inspired hierarchical Variational Autoencoder" submitted to NeurIPS 2022 (to be published soon on arXiv.org). Details on saved models coming soon.

{"references": ["J Hans Van Hateren and Arjen van der Schaaf. Independent component filters of natural images compared445 with simple cells in primary visual cortex. Proceedings of the Royal Society of London. Series B: Biological446 Sciences, 265(1394):359\u2013366, 1998.", "Javier Portilla and Eero P Simoncelli. A parametric texture model based on joint statistics of complex430 wavelet coefficients. International journal of computer vision, 40(1):49\u201370, 2000.", "https://www.textures.com/", "Phil Brodatz. Textures: a photographic album for artists and designers. Dover publications, 1966.", "Gabriel Barello, Adam S Charles, and Jonathan W Pillow. Sparse-coding variational auto-encoders.355 bioRxiv, page 399246, 2018."]}