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.

Here's an example of how the data looks (each class takes three-rows):



Why we made Fashion-MNIST

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."

To Serious Machine Learning Researchers

Seriously, we are talking about replacing MNIST. Here are some good reasons:

Get the Data

Many ML libraries already include Fashion-MNIST data/API, give it a try!

You can use direct links to download the dataset. The data is stored in the same format as the original MNIST data.

Name Content Examples Size Link MD5 Checksum
train-images-idx3-ubyte.gz training set images 60,000 26 MBytes Download 8d4fb7e6c68d591d4c3dfef9ec88bf0d
train-labels-idx1-ubyte.gz training set labels 60,000 29 KBytes Download 25c81989df183df01b3e8a0aad5dffbe
t10k-images-idx3-ubyte.gz test set images 10,000 4.3 MBytes Download bef4ecab320f06d8554ea6380940ec79
t10k-labels-idx1-ubyte.gz test set labels 10,000 5.1 KBytes Download bb300cfdad3c16e7a12a480ee83cd310

Alternatively, you can clone this GitHub repository; the dataset appears under data/fashion. This repo also contains some scripts for benchmark and visualization.

git clone [email protected]:zalandoresearch/fashion-mnist.git


Each training and test example is assigned to one of the following labels:

Label Description
0 T-shirt/top
1 Trouser
2 Pullover
3 Dress
4 Coat
5 Sandal
6 Shirt
7 Sneaker
8 Bag
9 Ankle boot


Loading data with Python (requires NumPy)

Use utils/mnist_reader in this repo:

import mnist_reader
X_train, y_train = mnist_reader.load_mnist('data/fashion', kind='train')
X_test, y_test = mnist_reader.load_mnist('data/fashion', kind='t10k')

Loading data with Tensorflow

Make sure you have downloaded the data and placed it in data/fashion. Otherwise, Tensorflow will download and use the original MNIST.

from tensorflow.examples.tutorials.mnist import input_data
data = input_data.read_data_sets('data/fashion')


Note, Tensorflow supports passing in a source url to the read_data_sets. You may use:

data = input_data.read_data_sets('data/fashion', source_url='')

Also, an official Tensorflow tutorial of using tf.keras, a high-level API to train Fashion-MNIST can be found here.

Loading data with other machine learning libraries

To date, the following libraries have included Fashion-MNIST as a built-in dataset. Therefore, you don't need to download Fashion-MNIST by yourself. Just follow their API and you are ready to go.

You are welcome to make pull requests to other open-source machine learning packages, improving their support to Fashion-MNIST dataset.

Loading data with other languages

As one of the Machine Learning community's most popular datasets, MNIST has inspired people to implement loaders in many different languages. You can use these loaders with the Fashion-MNIST dataset as well. (Note: may require decompressing first.) To date, we haven't yet tested all of these loaders with Fashion-MNIST.


We built an automatic benchmarking system based on scikit-learn that covers 129 classifiers (but no deep learning) with different parameters. Find the results here.


You can reproduce the results by running benchmark/ We recommend building and deploying this Dockerfile.

You are welcome to submit your benchmark; simply create a new issue and we'll list your results here. Before doing that, please make sure it does not already appear in this list. Visit our contributor guidelines for additional details.

The table below collects the submitted benchmarks. Note that we haven't yet tested these results. You are welcome to validate the results using the code provided by the submitter. Test accuracy may differ due to the number of epoch, batch size, etc. To correct this table, please create a new issue.

Classifier Preprocessing Fashion test accuracy MNIST test accuracy Submitter Code
2 Conv+pooling None 0.876 - Kashif Rasul :link:
2 Conv+pooling None 0.916 - Tensorflow's doc :link:
2 Conv+pooling+ELU activation (PyTorch) None 0.903 - @AbhirajHinge :link:
2 Conv Normalization, random horizontal flip, random vertical flip, random translation, random rotation. 0.919 0.971 Kyriakos Efthymiadis :link:
2 Conv <100K parameters None 0.925 0.992 @hardmaru :link:
2 Conv ~113K parameters Normalization 0.922 0.993 Abel G. :link:
2 Conv+3 FC ~1.8M parameters Normalization 0.932 0.994 @Xfan1025 :link:
2 Conv+3 FC ~500K parameters Augmentation, batch normalization 0.934 0.994 @cmasch :link:
2 Conv+pooling+BN None 0.934 - @khanguyen1207 :link:
2 Conv+2 FC Random Horizontal Flips 0.939 - @ashmeet13 :link:
3 Conv+2 FC None 0.907 - @Cenk Bircanoğlu :link:
3 Conv+pooling+BN None 0.903 0.994 @meghanabhange :link:
3 Conv+pooling+2 FC+dropout None 0.926 - @Umberto Griffo :link:
3 Conv+BN+pooling None 0.921 0.992 @gchhablani :link:
5 Conv+BN+pooling None 0.931 - @Noumanmufc1 :link:
CNN with optional shortcuts, dense-like connectivity standardization+augmentation+random erasing 0.947 - @kennivich :link:
GRU+SVM None 0.888 0.965 @AFAgarap :link:
GRU+SVM with dropout None 0.897 0.988 @AFAgarap :link:
WRN40-4 8.9M params standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.967 - @ajbrock :link: :link:
DenseNet-BC 768K params standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.954 - @ajbrock :link: :link:
MobileNet augmentation (horizontal flips) 0.950 - @苏剑林 :link:
ResNet18 Normalization, random horizontal flip, random vertical flip, random translation, random rotation. 0.949 0.979 Kyriakos Efthymiadis :link:
GoogleNet with cross-entropy loss None 0.937 - @Cenk Bircanoğlu :link:
AlexNet with Triplet loss None 0.899 - @Cenk Bircanoğlu :link:
SqueezeNet with cyclical learning rate 200 epochs None 0.900 - @snakers4 :link:
Dual path network with wide resnet 28-10 standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.957 - @Queequeg :link:
MLP 256-128-100 None 0.8833 - @heitorrapela :link:
VGG16 26M parameters None 0.935 - @QuantumLiu :link: :link:
WRN-28-10 standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.959 - @zhunzhong07 :link:
WRN-28-10 + Random Erasing standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.963 - @zhunzhong07 :link:
Human Performance Crowd-sourced evaluation of human (with no fashion expertise) performance. 1000 randomly sampled test images, 3 labels per image, majority labelling. 0.835 - Leo -
Capsule Network 8M parameters Normalization and shift at most 2 pixel and horizontal flip 0.936 - @XifengGuo :link:
HOG+SVM HOG 0.926 - @subalde :link:
XgBoost scaling the pixel values to mean=0.0 and var=1.0 0.898 0.958 @anktplwl91 :link:
DENSER - 0.953 0.997 @fillassuncao :link: :link:
Dyra-Net Rescale to unit interval 0.906 - @Dirk Schäfer :link: :link:
Google AutoML 24 compute hours (higher quality) 0.939 - @Sebastian Heinz :link:
Fastai Resnet50+Fine-tuning+Softmax on last layer's activations 0.9312 - @Sayak :link:

Other Explorations of Fashion-MNIST

Fashion-MNIST: Year in Review

Fashion-MNIST on Google Scholar

Generative adversarial networks (GANs)


Video Tutorial

Machine Learning Meets Fashion by Yufeng G @ Google Cloud

Introduction to Kaggle Kernels by Yufeng G @ Google Cloud

动手学深度学习 by Mu Li @ Amazon AI

Apache MXNet으로 배워보는 딥러닝(Deep Learning) - 김무현 (AWS 솔루션즈아키텍트)


t-SNE on Fashion-MNIST (left) and original MNIST (right)


PCA on Fashion-MNIST (left) and original MNIST (right)


UMAP on Fashion-MNIST (left) and original MNIST (right)


PyMDE on Fashion-MNIST (left) and original MNIST (right)



Thanks for your interest in contributing! There are many ways to get involved; start with our contributor guidelines and then check these open issues for specific tasks.


To discuss the dataset, please use Gitter.

Citing Fashion-MNIST

If you use Fashion-MNIST in a scientific publication, we would appreciate references to the following paper:

Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. Han Xiao, Kashif Rasul, Roland Vollgraf. arXiv:1708.07747

Biblatex entry:

  author       = {Han Xiao and Kashif Rasul and Roland Vollgraf},
  title        = {Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms},
  date         = {2017-08-28},
  year         = {2017},
  eprintclass  = {cs.LG},
  eprinttype   = {arXiv},
  eprint       = {cs.LG/1708.07747},