o Ni(@sdZddlmZddlmZddlmZddlZddlmm Z ddl m Z dZdZdZd Zd d gZGd d d e jjZGddde jjZdS)Cifar-10.1 dataset.)absolute_import)division)print_functionNae@article{recht2018cifar10.1, author = {Benjamin Recht and Rebecca Roelofs and Ludwig Schmidt and Vaishaal Shankar}, title = {Do CIFAR-10 Classifiers Generalize to CIFAR-10?}, year = {2018}, note = {\url{https://arxiv.org/abs/1806.00451}}, } @article{torralba2008tinyimages, author = {Antonio Torralba and Rob Fergus and William T. Freeman}, journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence}, title = {80 Million Tiny Images: A Large Data Set for Nonparametric Object and Scene Recognition}, year = {2008}, volume = {30}, number = {11}, pages = {1958-1970} } a#The CIFAR-10.1 dataset is a new test set for CIFAR-10. CIFAR-10.1 contains roughly 2,000 new test images that were sampled after multiple years of research on the original CIFAR-10 dataset. The data collection for CIFAR-10.1 was designed to minimize distribution shift relative to the original dataset. We describe the creation of CIFAR-10.1 in the paper "Do CIFAR-10 Classifiers Generalize to CIFAR-10?". The images in CIFAR-10.1 are a subset of the TinyImages dataset. There are currently two versions of the CIFAR-10.1 dataset: v4 and v6. z^https://github.com/modestyachts/CIFAR-10.1/blob/master/datasets/cifar10.1_{}_data.npy?raw=truez`https://github.com/modestyachts/CIFAR-10.1/blob/master/datasets/cifar10.1_{}_labels.npy?raw=truev4v6cs*eZdZdZejjdfdd ZZS)Cifar10_1ConfigzBuilderConfig for Cifar-10.1.Nc s@|tvr tdt|d|tt|jdi|||_dS)zConstructs a Cifar10_1Config. Args: data: `str`, one of `_DATA_OPTIONS`. **kwargs: keyword arguments forwarded to super. zdata must be one of %snameN) _DATA_OPTIONS ValueError setdefaultsuperr__init__data)selfrkwargs __class__r f/home/ubuntu/.local/lib/python3.10/site-packages/tensorflow_datasets/image_classification/cifar10_1.pyrBs    zCifar10_1Config.__init__)N) __name__ __module__ __qualname____doc__tfdscoredisallow_positional_argsr __classcell__r r rrr?src@sTeZdZdZedejdddedejdddgZdd Z d d Z d d Z dS) Cifar10_1raIt is the first version of our dataset on which we tested any classifier. As mentioned above, this makes the v4 dataset independent of the classifiers we evaluate. The numbers reported in the main sections of our paper use this version of the dataset. It was built from the top 25 TinyImages keywords for each class, which led to a slight class imbalance. The largest difference is that ships make up only 8% of the test set instead of 10%. v4 contains 2,021 images.z1.1.0r) descriptionversionrzIt is derived from a slightly improved keyword allocation that is exactly class balanced. This version of the dataset corresponds to the results in Appendix D of our paper. v6 contains 2,000 images.rc Cs:tjj|ttjtjjddtjjdddddtdS)N) r!)shape ) num_classesimagelabelz*https://github.com/modestyachts/CIFAR-10.1)builderrfeaturessupervised_keyshomepagecitation) rr DatasetInfo _DESCRIPTIONr* FeaturesDictImage ClassLabel _CITATION)rr r r_infojs  zCifar10_1._infocCsJt|jj}t|jj}|||g\}}tjj tj j ||ddgS)zReturns SplitGenerators.) image_path label_path)r gen_kwargs) _DL_URL_IMAGESformatbuilder_configr_DL_URL_LABELSr downloadrrSplitGeneratorSplitTEST)r dl_manager image_url label_urlr5r6r r r_split_generatorsws zCifar10_1._split_generatorsc cstjj|d }t|}Wdn1swYtjj|d }t|}Wdn1s6wYtt||D]\}\}}||d} || fVqBdS)Nrbr&)tfiogfileGFilenpload enumeratezip) rr5r6fimageslabelsir'r(recordr r r_generate_exampless   zCifar10_1._generate_examplesN) rrrrrrrVersionBUILDER_CONFIGSr4rCrRr r r rrQs     r)r __future__rrrnumpyrItensorflow.compat.v2compatv2rEtensorflow_datasets.public_api public_apirr3r/r8r;r r BuilderConfigrGeneratorBasedBuilderrr r r rs