The Fashion-MNIST is a dataset of Zalando’s article images. It serves as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms. This dataset includes 60,000 training examples and 10,000 test examples, each being a 28x28 grayscale image associated with a label from 10 classes.

• Description of Fashion-MNIST
• Where to Download the Fashion-MNIST Dataset
• Fashion MNIST in CSV
• Usage and Benchmarks
• Conclusion

Description of Fashion-MNIST

The classes in the Fashion-MNIST dataset represent different types of clothing and fashion items. These include:

Here’s an example of how the data looks

Fashion-MNIST was created with the intention of providing a more challenging benchmark dataset for machine learning and computer vision algorithms, as the original MNIST dataset (comprising hand-written digits) was considered too easy. The dataset aims to reflect real-world scenarios better and provide a more difficult challenge in image classification tasks.

Where to Download the Fashion-MNIST Dataset

• Access the MNIST dataset at its origin https://github.com/zalandoresearch/fashion-mnist/tree/master/data/fashion
• For those who prefer working with tabular data, a CSV version of the Fashion MNIST dataset is available here dataset-MNIST-fashion.zip

Fashion MNIST in CSV

Here’s the zip file dataset-MNIST-fashion.zip

inside it there are two csv files

• mnist_fashion_train.csv (60,000 images)
• mnist_fashion_test.csv (10,000 images)

The format is:

label, 1x1, 1x2, 1x3, … 28x27, 28x28 where ixj is the pixel in the i-th row and j-th column. the label is a number between 0 and 9

Here is the converter

import java.io.FileInputStream;
import java.io.FileWriter;
import java.io.IOException;

public class MNISTConverter {

    public static void convert(String imgFileName, String labelFileName, String outFileName, int n) throws IOException {
        try (FileInputStream f = new FileInputStream(imgFileName); 
             FileWriter o = new FileWriter(outFileName); 
             FileInputStream l = new FileInputStream(labelFileName)) {

            final StringBuilder firstLine = new StringBuilder("label,");
            for (int i = 1; i < 29; i++) {
                for (int j = 1; j < 29; j++) {
                    firstLine.append(i).append("x").append(j);
                    if (i != 28 || j != 28) {
                        firstLine.append(",");
                    }
                }
            }
            o.write(firstLine.toString() + "\n");

            byte[] bytes = new byte[16];
            f.read(bytes, 0, 16); // skip the first 16 bytes
            l.read(bytes, 0, 8);  // skip the first 8 bytes

            for (int i = 0; i < n; i++) {
                StringBuilder image = new StringBuilder();
                image.append(l.read() & 0xFF); // read label

                for (int j = 0; j < 28 * 28; j++) {
                    image.append(",").append(f.read() & 0xFF); // read image bytes
                }

                o.write(image.toString() + "\n");
            }
        }
    }

    public static void main(String[] args) throws IOException {
        convert("train-images-idx3-ubyte", "train-labels-idx1-ubyte", "mnist_fashion_train.csv", 60000);
        convert("t10k-images-idx3-ubyte", "t10k-labels-idx1-ubyte", "mnist_fashion_test.csv", 10000);
    }
}

if you have the uncompressed files (train-images-idx3-ubyte, train-labels-idx1-ubyte, t10k-images-idx3-ubyte, t10k-labels-idx1-ubyte) in the same directory as the script, you could execute it with jbang

jbang MNISTConverter.java

and the mnist_fashion_train.csv and mnist_fashion_test.csv would be generated.

The conversion is inspired by the one from Joseph Redmon https://pjreddie.com/projects/mnist-in-csv/

Usage and Benchmarks

Here is a link to a nice comparison of different implementations agains MNIST and Fashion-MNIST : http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/

a nice article by the Fashion-MNIST creator : https://hanxiao.io/2018/09/28/Fashion-MNIST-Year-In-Review/

Conclusion

In conclusion, the Fashion-MNIST dataset serves as an invaluable resource for both beginners and seasoned practitioners in the field of machine learning. Offering a more complex and realistic challenge than its predecessor, the original MNIST, it pushes the boundaries of image classification techniques. Whether you are conducting academic research, developing commercial applications, or just exploring the fascinating world of machine learning, Fashion-MNIST in its CSV form provides a perfect starting point for your endeavors.

Happy coding!