README.md

metadata

license: cc0-1.0
task_categories:
  - image-classification
  - clustering
tags:
  - mnist
  - digits
  - computer-vision
  - machine-learning
  - matlab
dataset_info:
  features:
    - name: data
      dtype: uint8
      shape:
        - 784
    - name: labels
      dtype: uint8
  splits:
    - name: train
      num_bytes: 3170000000
      num_examples: 8100000

MNIST8M Dataset (.mat format)

Dataset Description

This repository contains the MNIST8M dataset converted to MATLAB .h5 format for convenient use in MATLAB environments. The original data is sourced from the LIBSVM datasets page.

Dataset Summary

• Original Source: LIBSVM Multiclass Datasets - MNIST8M
• Format Conversion: Converted from original LibSVM format to MATLAB .h5 format
• Purpose: Facilitate clustering and machine learning experiments in MATLAB
• Files: MNIST8M_data.h5, labels.mat

Data Specifications

• Samples: 8,100,000 (8.1 million)
• Features: 784 (28×28 pixel images)
• Data Type: uint8
• Value Range: [0, 255]
• Labels: 10 classes (digits 0-9)
• Label Type: uint8
• Label Range: [0, 9]

Storage Format

• MNIST8M_data: uint8 matrix of size 8,100,000 × 784
• labels: uint8 vector of size 8,100,000 × 1

Usage Warning

⚠️ Memory Considerations: Loading the entire dataset directly into memory may cause out-of-memory errors on systems with insufficient RAM. The uncompressed data requires approximately 6GB of memory (8.1M × 784 × 1 byte).

Recommended Usage

For systems with limited memory, consider:

• Loading data in batches
• Using memory-mapped files
• Working with data subsets
• Converting to single precision when possible

Source Attribution

Original dataset courtesy of:

• LIBSVM Datasets
• infinite MNIST

MATLAB Loading Example

dataset_name = 'MNIST8M';

% Data path
data_file = 'MNIST8M_data.h5'; % Your directory
data = h5read(data_file, '/MNIST8M');
data = double(data); % May cause out-of-memory
data_info = h5info(data_file);
data_size = data_info.Datasets.Dataspace.Size;
n_points = data_size(1);   % Total number of points
n_features = data_size(2);       % Number of dimensions

>> h5disp('MNIST8M_data.h5');
HDF5 MNIST8M_data.h5 
Group '/' 
    Dataset 'MNIST8M' 
        Size:  8100000x784
        MaxSize:  8100000x784
        Datatype:   H5T_STD_U8LE (uint8)
        ChunkSize:  100000x784
        Filters:  none
        FillValue:  0