The MNIST database of handwritten digits.

To use this dataset:

import tensorflow_datasets as tfds

ds = tfds.load('mnist', split='train')
for ex in ds.take(4):
 print(ex)

See the guide for more informations on tensorflow_datasets.

https://storage.googleapis.com/tfds-data/visualization/fig/mnist-3.0.1.png" alt="Visualization" width="500px">

The MNIST database of handwritten digits, available from this page, has a training set of 60,000 examples, and a test set of 10,000 examples. It is a subset of a larger set available from NIST. The digits have been size-normalized and centered in a fixed-size image. It is a good database for people who want to try learning techniques and pattern recognition methods on real-world data while spending minimal efforts on preprocessing and formatting. The original black and white (bilevel) images from NIST were size normalized to fit in a 20x20 pixel box while preserving their aspect ratio. The resulting images contain grey levels as a result of the anti-aliasing technique used by the normalization algorithm. the images were centered in a 28x28 image by computing the center of mass of the pixels, and translating the image so as to position this point at the center of the 28x28 field. With some classification methods (particuarly template-based methods, such as SVM and K-nearest neighbors),

Author: Yann LeCun, Corinna Cortes, Christopher J.C. Burges
Source: MNIST Website - Date unknown
Please cite:

The MNIST database of handwritten digits with 784 features, raw data available at: http://yann.lecun.com/exdb/mnist/. It can be split in a training set of the first 60,000 examples, and a test set of 10,000 examples

It is a subset of a larger set available from NIST. The digits have been size-normalized and centered in a fixed-size image. It is a good database for people who want to try learning techniques and pattern recognition methods on real-world data while spending minimal efforts on preprocessing and formatting. The original black and white (bilevel) images from NIST were size normalized to fit in a 20x20 pixel box while preserving their aspect ratio. The resulting images contain grey levels as a result of the anti-aliasing technique used by the normalization algorithm. the images were centered in a 28x28 image by computing the center of mass of the pixels, and translating the image so as to position this point at the center of the 28x28 field.

With some classification methods (particularly template-based methods, such as SVM and K-nearest neighbors), the error rate improves when the digits are centered by bounding box rather than center of mass. If you do this kind of pre-processing, you should report it in your publications. The MNIST database was constructed from NIST's NIST originally designated SD-3 as their training set and SD-1 as their test set. However, SD-3 is much cleaner and easier to recognize than SD-1. The reason for this can be found on the fact that SD-3 was collected among Census Bureau employees, while SD-1 was collected among high-school students. Drawing sensible conclusions from learning experiments requires that the result be independent of the choice of training set and test among the complete set of samples. Therefore it was necessary to build a new database by mixing NIST's datasets.

The MNIST training set is composed of 30,000 patterns from SD-3 and 30,000 patterns from SD-1. Our test set was composed of 5,000 patterns from SD-3 and 5,000 patterns from SD-1. The 60,000 pattern training set contained examples from approximately 250 writers. We made sure that the sets of writers of the training set and test set were disjoint. SD-1 contains 58,527 digit images written by 500 different writers. In contrast to SD-3, where blocks of data from each writer appeared in sequence, the data in SD-1 is scrambled. Writer identities for SD-1 is available and we used this information to unscramble the writers. We then split SD-1 in two: characters written by the first 250 writers went into our new training set. The remaining 250 writers were placed in our test set. Thus we had two sets with nearly 30,000 examples each. The new training set was completed with enough examples from SD-3, starting at pattern # 0, to make a full set of 60,000 training patterns. Similarly, the new test set was completed with SD-3 examples starting at pattern # 35,000 to make a full set with 60,000 test patterns. Only a subset of 10,000 test images (5,000 from SD-1 and 5,000 from SD-3) is available on this site. The full 60,000 sample training set is available.

The not-MNIST dataset is a dataset of handwritten digits. It is a challenging dataset that can be used for machine learning and artificial intelligence research. The dataset consists of 100,000 images of handwritten digits. The images are divided into a training set of 60,000 images and a test set of 40,000 images. The images are drawn from a variety of fonts and styles, making them more challenging than the MNIST dataset. The images are 28x28 pixels in size and are grayscale. The dataset is available under the Creative Commons Zero Public Domain Dedication license.

The EMNIST dataset is a set of handwritten character digits derived from the NIST Special Database 19 (https://www.nist.gov/srd/nist-special-database-19) and converted to a 28x28 pixel image format and dataset structure that directly matches the MNIST dataset (http://yann.lecun.com/exdb/mnist/). Further information on the dataset contents and conversion process can be found in the paper available at https://arxiv.org/abs/1702.05373v2

The MNIST dataset has become a standard benchmark for learning, classification and computer vision systems. Contributing to its widespread adoption are the understandable and intuitive nature of the task, its relatively small size and storage requirements and the accessibility and ease-of-use of the database itself. The MNIST database was derived from a larger dataset known as the NIST Special Database 19 which contains digits, uppercase and lowercase handwritten letters. This paper introduces a variant of the full NIST dataset, which we have called Extended MNIST (EMNIST), which follows the same conversion paradigm used to create the MNIST dataset. The result is a set of datasets that constitute a more challenging classification tasks involving letters and digits, and that shares the same image structure and parameters as the original MNIST task, allowing for direct compatibility with all existing classifiers and systems. Benchmark results are presented along with a validation of the conversion process through the comparison of the classification results on converted NIST digits and the MNIST digits.

The database is made available in original MNIST format and Matlab format.

Fashion-MNIST is a dataset of Zalando's article images consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes.

To use this dataset:

import tensorflow_datasets as tfds

ds = tfds.load('fashion_mnist', split='train')
for ex in ds.take(4):
 print(ex)

See the guide for more informations on tensorflow_datasets.

https://storage.googleapis.com/tfds-data/visualization/fig/fashion_mnist-3.0.1.png" alt="Visualization" width="500px">

Motivation

The goal of introducing the Rescaled Fashion-MNIST dataset is to provide a dataset that contains scale variations (up to a factor of 4), to evaluate the ability of networks to generalise to scales not present in the training data.

The Rescaled Fashion-MNIST dataset was introduced in the paper:

[1] A. Perzanowski and T. Lindeberg (2025) "Scale generalisation properties of extended scale-covariant and scale-invariant Gaussian derivative networks on image datasets with spatial scaling variations”, Journal of Mathematical Imaging and Vision, 67(29), https://doi.org/10.1007/s10851-025-01245-x.

with a pre-print available at arXiv:

[2] Perzanowski and Lindeberg (2024) "Scale generalisation properties of extended scale-covariant and scale-invariant Gaussian derivative networks on image datasets with spatial scaling variations”, arXiv preprint arXiv:2409.11140.

Importantly, the Rescaled Fashion-MNIST dataset is more challenging than the MNIST Large Scale dataset, introduced in:

[3] Y. Jansson and T. Lindeberg (2022) "Scale-invariant scale-channel networks: Deep networks that generalise to previously unseen scales", Journal of Mathematical Imaging and Vision, 64(5): 506-536, https://doi.org/10.1007/s10851-022-01082-2.

Access and rights

The Rescaled Fashion-MNIST dataset is provided on the condition that you provide proper citation for the original Fashion-MNIST dataset:

[4] Xiao, H., Rasul, K., and Vollgraf, R. (2017) “Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms”, arXiv preprint arXiv:1708.07747

and also for this new rescaled version, using the reference [1] above.

The data set is made available on request. If you would be interested in trying out this data set, please make a request in the system below, and we will grant you access as soon as possible.

The dataset

The Rescaled FashionMNIST dataset is generated by rescaling 28×28 gray-scale images of clothes from the original FashionMNIST dataset [4]. The scale variations are up to a factor of 4, and the images are embedded within black images of size 72x72, with the object in the frame always centred. The imresize() function in Matlab was used for the rescaling, with default anti-aliasing turned on, and bicubic interpolation overshoot removed by clipping to the [0, 255] range. The details of how the dataset was created can be found in [1].

There are 10 different classes in the dataset: “T-shirt/top”, “trouser”, “pullover”, “dress”, “coat”, “sandal”, “shirt”, “sneaker”, “bag” and “ankle boot”. In the dataset, these are represented by integer labels in the range [0, 9].

The dataset is split into 50 000 training samples, 10 000 validation samples and 10 000 testing samples. The training dataset is generated using the initial 50 000 samples from the original Fashion-MNIST training set. The validation dataset, on the other hand, is formed from the final 10 000 images of that same training set. For testing, all test datasets are built from the 10 000 images contained in the original Fashion-MNIST test set.

The h5 files containing the dataset

The training dataset file (~2.9 GB) for scale 1, which also contains the corresponding validation and test data for the same scale, is:

fashionmnist_with_scale_variations_tr50000_vl10000_te10000_outsize72-72_scte1p000_scte1p000.h5

Additionally, for the Rescaled FashionMNIST dataset, there are 9 datasets (~415 MB each) for testing scale generalisation at scales not present in the training set. Each of these datasets is rescaled using a different image scaling factor, 2^k/4, with k being integers in the range [-4, 4]:

fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p500.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p595.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p707.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte0p841.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p000.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p189.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p414.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte1p682.h5
fashionmnist_with_scale_variations_te10000_outsize72-72_scte2p000.h5

These dataset files were used for the experiments presented in Figures 6, 7, 14, 16, 19 and 23 in [1].

Instructions for loading the data set

The datasets are saved in HDF5 format, with the partitions in the respective h5 files named as
('/x_train', '/x_val', '/x_test', '/y_train', '/y_test', '/y_val'); which ones exist depends on which data split is used.

The training dataset can be loaded in Python as:

with h5py.File(`

x_train = np.array( f["/x_train"], dtype=np.float32)
x_val = np.array( f["/x_val"], dtype=np.float32)
x_test = np.array( f["/x_test"], dtype=np.float32)
y_train = np.array( f["/y_train"], dtype=np.int32)
y_val = np.array( f["/y_val"], dtype=np.int32)
y_test = np.array( f["/y_test"], dtype=np.int32)

We also need to permute the data, since Pytorch uses the format [num_samples, channels, width, height], while the data is saved as [num_samples, width, height, channels]:

x_train = np.transpose(x_train, (0, 3, 1, 2))
x_val = np.transpose(x_val, (0, 3, 1, 2))
x_test = np.transpose(x_test, (0, 3, 1, 2))

The test datasets can be loaded in Python as:

with h5py.File(`

x_test = np.array( f["/x_test"], dtype=np.float32)
y_test = np.array( f["/y_test"], dtype=np.int32)

The test datasets can be loaded in Matlab as:

x_test = h5read(`

The images are stored as [num_samples, x_dim, y_dim, channels] in HDF5 files. The pixel intensity values are not normalised, and are in a [0, 255] range.

There is also a closely related Fashion-MNIST with translations dataset, which in addition to scaling variations also comprises spatial translations of the objects.

The MNIST dataset is a collection of images of handwritten digits, with size n = 70,000 and D = 784.

This dataset is derived from the Leaf repository (https://github.com/TalwalkarLab/leaf) pre-processing of the Extended MNIST dataset, grouping examples by writer. Details about Leaf were published in "LEAF: A Benchmark for Federated Settings" https://arxiv.org/abs/1812.01097Note: This dataset does not include some additional preprocessing that MNIST includes, such as size-normalization and centering. In the Federated EMNIST data, the value of 1.0 corresponds to the background, and 0.0 corresponds to the color of the digits themselves; this is the inverse of some MNIST representations, e.g. in tensorflow_datasets, where 0 corresponds to the background color, and 255 represents the color of the digit.Data set sizes:only_digits=True: 3,383 users, 10 label classestrain: 341,873 examplestest: 40,832 examplesonly_digits=False: 3,400 users, 62 label classestrain: 671,585 examplestest: 77,483 examplesRather than holding out specific users, each user's examples are split across train and test so that all users have at least one example in train and one example in test. Writers that had less than 2 examples are excluded from the data set.The tf.data.Datasets returned by tff.simulation.datasets.ClientData.create_tf_dataset_for_client will yield collections.OrderedDict objects at each iteration, with the following keys and values, in lexicographic order by key:'label': a tf.Tensor with dtype=tf.int32 and shape [1], the class label of the corresponding pixels. Labels [0-9] correspond to the digits classes, labels [10-35] correspond to the uppercase classes (e.g., label 11 is 'B'), and labels [36-61] correspond to the lowercase classes (e.g., label 37 is 'b').'pixels': a tf.Tensor with dtype=tf.float32 and shape [28, 28], containing the pixels of the handwritten digit, with values in the range [0.0, 1.0].Argsonly_digits(Optional) whether to only include examples that are from the digits [0-9] classes. If False, includes lower and upper case characters, for a total of 62 class labels.cache_dir(Optional) directory to cache the downloaded file. If None, caches in Keras' default cache directory.ReturnsTuple of (train, test) where the tuple elements are tff.simulation.datasets.ClientData objects.

The EMNIST dataset is a set of handwritten character digits derived from the NIST Special Database 19 and converted to a 28x28 pixel image format and dataset structure that directly matches the MNIST dataset.

Note: Like the original EMNIST data, images provided here are inverted horizontally and rotated 90 anti-clockwise. You can use tf.transpose within ds.map to convert the images to a human-friendlier format.

To use this dataset:

import tensorflow_datasets as tfds

ds = tfds.load('emnist', split='train')
for ex in ds.take(4):
 print(ex)

See the guide for more informations on tensorflow_datasets.

https://storage.googleapis.com/tfds-data/visualization/fig/emnist-byclass-3.1.0.png" alt="Visualization" width="500px">

Database description:

The written and spoken digits database is not a new database but a constructed database from existing ones, in order to provide a ready-to-use database for multimodal fusion [1].

The written digits database is the original MNIST handwritten digits database [2] with no additional processing. It consists of 70000 images (60000 for training and 10000 for test) of 28 x 28 = 784 dimensions.

The spoken digits database was extracted from Google Speech Commands [3], an audio dataset of spoken words that was proposed to train and evaluate keyword spotting systems. It consists of 105829 utterances of 35 words, amongst which 38908 utterances of the ten digits (34801 for training and 4107 for test). A pre-processing was done via the extraction of the Mel Frequency Cepstral Coefficients (MFCC) with a framing window size of 50 ms and frame shift size of 25 ms. Since the speech samples are approximately 1 s long, we end up with 39 time slots. For each one, we extract 12 MFCC coefficients with an additional energy coefficient. Thus, we have a final vector of 39 x 13 = 507 dimensions. Standardization and normalization were applied on the MFCC features.

To construct the multimodal digits dataset, we associated written and spoken digits of the same class respecting the initial partitioning in [2] and [3] for the training and test subsets. Since we have less samples for the spoken digits, we duplicated some random samples to match the number of written digits and have a multimodal digits database of 70000 samples (60000 for training and 10000 for test).

The dataset is provided in six files as described below. Therefore, if a shuffle is performed on the training or test subsets, it must be performed in unison with the same order for the written digits, spoken digits and labels.

Files:

data_wr_train.npy: 60000 samples of 784-dimentional written digits for training;

data_sp_train.npy: 60000 samples of 507-dimentional spoken digits for training;

labels_train.npy: 60000 labels for the training subset;

data_wr_test.npy: 10000 samples of 784-dimentional written digits for test;

data_sp_test.npy: 10000 samples of 507-dimentional spoken digits for test;

labels_test.npy: 10000 labels for the test subset.

References:

Khacef, L. et al. (2020), "Brain-Inspired Self-Organization with Cellular Neuromorphic Computing for Multimodal Unsupervised Learning".

LeCun, Y. & Cortes, C. (1998), “MNIST handwritten digit database”.

Warden, P. (2018), “Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition”.

The MNIST-scale dataset is a dataset of images of handwritten digits, where each image is scaled to a different size.

[!NOTE] This dataset card is based on the README file of the authors' GitHub repository: https://github.com/greydanus/mnist1d

  The MNIST-1D Dataset

Most machine learning models get around the same ~99% test accuracy on MNIST. The MNIST-1D dataset is 100x smaller (default sample size: 4000+1000; dimensionality: 40) and does a better job of separating between models with/without nonlinearity and models with/without spatial inductive biases. MNIST-1D is a core teaching dataset in… See the full description on the dataset page: https://huggingface.co/datasets/christopher/mnist1d.

Fashion-MNIST is a dataset of Zalando's article images—consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. We intend Fashion-MNIST to serve as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms. It shares the same image size and structure of training and testing splits.

The original MNIST dataset contains a lot of handwritten digits. Members of the AI/ML/Data Science community love this dataset and use it as a benchmark to validate their algorithms. In fact, MNIST is often the first dataset researchers try. "If it doesn't work on MNIST, it won't work at all", they said. "Well, if it does work on MNIST, it may still fail on others."

Here we present a dataset, MNIST4OD, of large size (number of dimensions and number of instances) suitable for Outliers Detection task.The dataset is based on the famous MNIST dataset (http://yann.lecun.com/exdb/mnist/).We build MNIST4OD in the following way:To distinguish between outliers and inliers, we choose the images belonging to a digit as inliers (e.g. digit 1) and we sample with uniform probability on the remaining images as outliers such as their number is equal to 10% of that of inliers. We repeat this dataset generation process for all digits. For implementation simplicity we then flatten the images (28 X 28) into vectors.Each file MNIST_x.csv.gz contains the corresponding dataset where the inlier class is equal to x.The data contains one instance (vector) in each line where the last column represents the outlier label (yes/no) of the data point. The data contains also a column which indicates the original image class (0-9).See the following numbers for a complete list of the statistics of each datasets ( Name | Instances | Dimensions | Number of Outliers in % ):MNIST_0 | 7594 | 784 | 10MNIST_1 | 8665 | 784 | 10MNIST_2 | 7689 | 784 | 10MNIST_3 | 7856 | 784 | 10MNIST_4 | 7507 | 784 | 10MNIST_5 | 6945 | 784 | 10MNIST_6 | 7564 | 784 | 10MNIST_7 | 8023 | 784 | 10MNIST_8 | 7508 | 784 | 10MNIST_9 | 7654 | 784 | 10

A dataset containing images of handwritten english numerals from 0-9 obtained from National Institute of Standards and Technology. It consists of greyscale images of handwritten digits and consists of 60000 images of size 28*28 for training and 10000 images as test examples.

Kuzushiji-MNIST is a drop-in replacement for the MNIST dataset (28x28 grayscale, 70,000 images), provided in the original MNIST format as well as a NumPy format. Since MNIST restricts us to 10 classes, we chose one character to represent each of the 10 rows of Hiragana when creating Kuzushiji-MNIST.

To use this dataset:

import tensorflow_datasets as tfds

ds = tfds.load('kmnist', split='train')
for ex in ds.take(4):
 print(ex)

See the guide for more informations on tensorflow_datasets.

https://storage.googleapis.com/tfds-data/visualization/fig/kmnist-3.0.1.png" alt="Visualization" width="500px">

Moving variant of MNIST database of handwritten digits. This is the data used by the authors for reporting model performance. See tfds.video.moving_mnist.image_as_moving_sequence for generating training/validation data from the MNIST dataset.

To use this dataset:

import tensorflow_datasets as tfds

ds = tfds.load('moving_mnist', split='train')
for ex in ds.take(4):
 print(ex)

See the guide for more informations on tensorflow_datasets.

Code [GitHub] | Publication [Nature Scientific Data'23 / ISBI'21] | Preprint [arXiv]

Abstract

We introduce MedMNIST, a large-scale MNIST-like collection of standardized biomedical images, including 12 datasets for 2D and 6 datasets for 3D. All images are pre-processed into 28x28 (2D) or 28x28x28 (3D) with the corresponding classification labels, so that no background knowledge is required for users. Covering primary data modalities in biomedical images, MedMNIST is designed to perform classification on lightweight 2D and 3D images with various data scales (from 100 to 100,000) and diverse tasks (binary/multi-class, ordinal regression and multi-label). The resulting dataset, consisting of approximately 708K 2D images and 10K 3D images in total, could support numerous research and educational purposes in biomedical image analysis, computer vision and machine learning. We benchmark several baseline methods on MedMNIST, including 2D / 3D neural networks and open-source / commercial AutoML tools. The data and code are publicly available at https://medmnist.com/.

Disclaimer: The only official distribution link for the MedMNIST dataset is Zenodo. We kindly request users to refer to this original dataset link for accurate and up-to-date data.

Update: We are thrilled to release MedMNIST+ with larger sizes: 64x64, 128x128, and 224x224 for 2D, and 64x64x64 for 3D. As a complement to the previous 28-size MedMNIST, the large-size version could serve as a standardized benchmark for medical foundation models. Install the latest API to try it out!

Python Usage

We recommend our official code to download, parse and use the MedMNIST dataset:

% pip install medmnist% python

To use the standard 28-size (MNIST-like) version utilizing the downloaded files:

from medmnist import PathMNIST

train_dataset = PathMNIST(split="train")

To enable automatic downloading by setting download=True:

from medmnist import NoduleMNIST3D

val_dataset = NoduleMNIST3D(split="val", download=True)

Alternatively, you can access MedMNIST+ with larger image sizes by specifying the size parameter:

from medmnist import ChestMNIST

test_dataset = ChestMNIST(split="test", download=True, size=224)

Citation

If you find this project useful, please cite both v1 and v2 paper as:

Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, Bingbing Ni. Yang, Jiancheng, et al. "MedMNIST v2-A large-scale lightweight benchmark for 2D and 3D biomedical image classification." Scientific Data, 2023.

Jiancheng Yang, Rui Shi, Bingbing Ni. "MedMNIST Classification Decathlon: A Lightweight AutoML Benchmark for Medical Image Analysis". IEEE 18th International Symposium on Biomedical Imaging (ISBI), 2021.

or using bibtex:

@article{medmnistv2, title={MedMNIST v2-A large-scale lightweight benchmark for 2D and 3D biomedical image classification}, author={Yang, Jiancheng and Shi, Rui and Wei, Donglai and Liu, Zequan and Zhao, Lin and Ke, Bilian and Pfister, Hanspeter and Ni, Bingbing}, journal={Scientific Data}, volume={10}, number={1}, pages={41}, year={2023}, publisher={Nature Publishing Group UK London} }

@inproceedings{medmnistv1, title={MedMNIST Classification Decathlon: A Lightweight AutoML Benchmark for Medical Image Analysis}, author={Yang, Jiancheng and Shi, Rui and Ni, Bingbing}, booktitle={IEEE 18th International Symposium on Biomedical Imaging (ISBI)}, pages={191--195}, year={2021} }

Please also cite the corresponding paper(s) of source data if you use any subset of MedMNIST as per the description on the project website.

License

The MedMNIST dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0), except DermaMNIST under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

The code is under Apache-2.0 License.

Changelog

v3.0 (this repository): Released MedMNIST+ featuring larger sizes: 64x64, 128x128, and 224x224 for 2D, and 64x64x64 for 3D.

v2.2: Removed a small number of mistakenly included blank samples in OrganAMNIST, OrganCMNIST, OrganSMNIST, OrganMNIST3D, and VesselMNIST3D.

v2.1: Addressed an issue in the NoduleMNIST3D file (i.e., nodulemnist3d.npz). Further details can be found in this issue.

v2.0: Launched the initial repository of MedMNIST v2, adding 6 datasets for 3D and 2 for 2D.

v1.0: Established the initial repository (in a separate repository) of MedMNIST v1, featuring 10 datasets for 2D.

Note: This dataset is NOT intended for clinical use.

This dataset contains a simplified version of the famous MNIST handwritten digits dataset. This version involves distinguishing between digits 3 and 5 rather than the full range 0-9.