/fashion-mnist

A MNIST-like fashion product database. Benchmark :point_right:

Primary LanguagePythonMIT LicenseMIT

Fashion-MNIST

GitHub stars Gitter Readme-CN Readme-JA

Fashion-MNIST is a dataset of Zalando's article imagesโ€”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 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 git@github.com:zalandoresearch/fashion-mnist.git

Labels

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

Usage

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')

data.train.next_batch(BATCH_SIZE)

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

data = input_data.read_data_sets('data/fashion', source_url='http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/')

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.

Benchmark

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/runner.py. 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 Layers with max pooling (Keras) None 0.876 - Kashif Rasul ๐Ÿ”—
2 Conv Layers with max pooling (Tensorflow) >300 epochs None 0.916 - Tensorflow's doc ๐Ÿ”—
2 Conv Layers with max pooling and ELU activation (PyTorch) None 0.903 - @AbhirajHinge ๐Ÿ”—
2 Conv Layers net Normalization, random horizontal flip, random vertical flip, random translation, random rotation. 0.919 0.971 Kyriakos Efthymiadis ๐Ÿ”—
2 Conv Layers net <100K parameters None 0.925 0.992 @hardmaru ๐Ÿ”—
2 Conv Layers 113K parameters Normalization 0.922 0.993 Abel G. ๐Ÿ”—
2 Conv Layers with 3 FC 1.8M parameters Normalization 0.932 0.994 @Xfan1025 ๐Ÿ”—
3 Conv layers and 2 FC None 0.907 - @Cenk BircanoฤŸlu ๐Ÿ”—
3 Conv+pooling and 2 FC+dropout None 0.926 - @Umberto Griffo ๐Ÿ”—
GRU+SVM None 0.888 0.965 @AFAgarap ๐Ÿ”—
GRU+SVM with dropout None 0.897 0.988 @AFAgarap ๐Ÿ”—
WRN40-4 8.9M params standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.967 - @ajbrock ๐Ÿ”— ๐Ÿ”—
DenseNet-BC 768K params standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.954 - @ajbrock ๐Ÿ”— ๐Ÿ”—
MobileNet augmentation (horizontal flips) 0.950 - @่‹ๅ‰‘ๆž— ๐Ÿ”—
ResNet18 Normalization, random horizontal flip, random vertical flip, random translation, random rotation. 0.949 0.979 Kyriakos Efthymiadis ๐Ÿ”—
GoogleNet with cross-entropy loss None 0.937 - @Cenk BircanoฤŸlu ๐Ÿ”—
AlexNet with Triplet loss None 0.899 - @Cenk BircanoฤŸlu ๐Ÿ”—
SqueezeNet with cyclical learning rate 200 epochs None 0.900 - @snakers4 ๐Ÿ”—
Dual path network with wide resnet 28-10 standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.957 - @Queequeg ๐Ÿ”—
MLP 256-128-64 None 0.900 - @lianghong ๐Ÿ”—
VGG16 26M parameters None 0.935 - @QuantumLiu ๐Ÿ”— ๐Ÿ”—
WRN-28-10 standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.959 - @zhunzhong07 ๐Ÿ”—
WRN-28-10 + Random Erasing standard preprocessing (mean/std subtraction/division) and augmentation (random crops/horizontal flips) 0.963 - @zhunzhong07 ๐Ÿ”—
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 ๐Ÿ”—
HOG+SVM HOG 0.926 - @subalde ๐Ÿ”—
XgBoost scaling the pixel values to mean=0.0 and var=1.0 0.898 0.958 @anktplwl91 ๐Ÿ”—

Other Explorations of Fashion-MNIST

Generative adversarial networks (GANs)

Video Tutorial

Machine Learning Meets Fashion by Yufeng G @ Google Cloud

Machine Learning Meets Fashion

Introduction to Kaggle Kernels by Yufeng G @ Google Cloud

Introduction to Kaggle Kernels

ๅŠจๆ‰‹ๅญฆๆทฑๅบฆๅญฆไน  by Mu Li @ Amazon AI

MXNet/Gluonไธญๆ–‡้ข‘้“

Apache MXNet์œผ๋กœ ๋ฐฐ์›Œ๋ณด๋Š” ๋”ฅ๋Ÿฌ๋‹(Deep Learning) - ๊น€๋ฌดํ˜„ (AWS ์†”๋ฃจ์…˜์ฆˆ์•„ํ‚คํ…ํŠธ)

Apache MXNet์œผ๋กœ ๋ฐฐ์›Œ๋ณด๋Š” ๋”ฅ๋Ÿฌ๋‹(Deep Learning)

Visualization

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

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

Contributing

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.

Contact

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:

@online{xiao2017/online,
  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},
}

Who is citing Fashion-MNIST?

License

The MIT License (MIT) Copyright ยฉ [2017] Zalando SE, https://tech.zalando.com

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the โ€œSoftwareโ€), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED โ€œAS ISโ€, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.