Image training data format for building models in NeuNetS

The NeuNetS tool in IBM Watson Studio synthesizes a neural network and trains it on your training data without you having to design or build anything by hand. This topic describes how to organize and format image training data. NeuNetS is currently in beta.


To train an image classifier in the NeuNetS tool, you must provide two files:

  • A file named that contains your sample JPEG (.jpg) or PNG (.png) images
  • A file named labels.csv that lists your image file names and corresponding class label


Image .zip file contents

  • Put all of your sample images (the images for all the classes) into a single .zip file, named
  • Supported image file formats: JPEG (.jpg), and PNG (.png)


Labels file format

  • Each row of the file must contain: the image file name (including the path, if the .zip contains folders), a comma, and then a class name
  • When specifying the path (if the .zip contains folders) use a forward slash “/” separator:
    • Correct: my-images/class1/img_01.jpg
    • Incorrect: my-images\class1\img_01.jpg
  • The file must be saved as: UTF-8 encoded
  • Do not include a header in the file like you might in other comma-delimited files


Requirements and limits

Table 1. Requirements for sample images, ``, and `labels.csv`
Requirement Minimum Maximum Notes
Image .zip file size -- 5GB
Image dimensions 32 pixels by 32 pixels Limited only by the 5GB total training set restriction Images larger than 64 pixels by 64 pixels will be scaled down for training
Number of classes 2 Limited only by the 5GB total training set restriction
Number of sample images in each class 100 Limited only by the 5GB total training set restriction A minimum of 250 sample images is recommended for best performance



This example shows a .zip file containing sample search-and-rescue aerial images:

  • A wheelbarrow in a field
  • Large play blocks
  • A grazing horse

Working directory:

Example working directory

Image .zip file:

Example .zip file containing images

Labels file listing the image files and their corresponding class name: