Counting 2,129 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Author
Contributors
Last Commit
Nov. 26, 2017
Created
Jun. 5, 2017

MNIST Digit Sequence Generator: Images representing sequence of digits using the MNIST handwritten digit image dataset

Problem

The MNIST database is a large database of handwritten digits that is commonly used for training various image processing systems. The database is also widely used for training and testing in the field of machine learning. The goal of this project is to use the above database of handwritten digit images to generate images representing sequences of handwritten digits.

The project also provides a utility to generate and save a set of training/test image dataset of MNIST sequences with labels.

These images could be used to train classifiers and generative deep learning models.

Solution

The solution can be explained using the following steps:

  • Decode the image and label MNIST data given in the special idx format with load.py.
  • Create a label hash map in which each key corresponds to a digit, and the values are the index of digits in the dataset. This will help us randomly choose handwritten image of any particular digit.
  • Calculate allowed spacing according to the given values of minimum spacing, maximum spacing and image width using the formula: allowed_spacing = (total_width - size_sequence * image_width) / ((size_sequence - 1) * 1.0). If the allowed spacing allows a uniform spacing for the sequence, move towards the generation of the image.
  • Randomly choose an image index for each digit in the sequence. Next, horizontally stack each of these sequence digit images separated by uniformly spaced white background as per the allowed spacing. This is done in mnist_sequence.py.
  • Unit tests are available in tests.py.
  • An API endpoint is provided in mnist_sequence_api.py.
  • A CLI script is provided in mnist_sequence_cli.py.A CLI script saves examples of generated images for inspecting the characteristics of the generated images and in inspecting the trained models behaviors.
  • A utility to generate a set of training/test image dataset with labels as an API endpoint is provided in mnist_sequence_api.py.

The above algorithm is implemented in an object-oriented approach which allows great flexibility for future changes. In the API endpoint, the data is loaded only once which will help in scaling the solution. In the future, using parallel processing would improve program's response time.

Dependencies

The library is supported for Python >= 2.7 and Python >= 3.3.

The python modules needed in order to use this library.

Pillow: 3.3.1
numpy: 1.13.3 + mkl
bcolz: 1.1.2

Note: numpy and bcolz can be installed on a Windows machines using binaries provided over here.

Tests

The tests can be run using the below command, currently there are 4 tests.

python tests.py

Examples

The library currently uses the MNIST 10k test data samples, also available in the data folder with the project.

python examples.py

This will run generate three images of sequences and save them in the current directory.

python mnist_sequence_cli.py 01 0 10 66

This will run the CLI script, generate and save the image of the sequence ([0, 1]) of the digits with minimum spacing 0, maximum spacing 10 and width 66.

API:

from mnist_sequence_api import MNIST_Sequence_API

api_object = MNIST_Sequence_API()
api_object.generate_mnist_sequence(digits, spacing_range, image_width)
api_object.generate_data(num_samples, seq_len)

where:

  • digits is a list-like containing the numerical values of the digits from which the sequence will be generated (for example [3, 5, 0])
  • spacing_range is a tuple containing the minimum and maximum spacing (in pixels) between digits.
  • image_width specifies the width of the image in pixels.
  • num_samples is the number of samples of image dataset (MNIST sequence) with labels to be generated.
  • seq_len is the length of the sequence for each image in the image dataset to be generated.

The function returns a (28 x image_width) image.

Images are represented as floating point 32 bits numpy arrays with a scale ranging from 0 (black) to 1 (white), the first dimension corresponding to the height and the second dimension to the width.

Script

The CLI script takes positional arguments:

  • < sequence >: the sequence of digits to be generated
  • <min_spacing>: minimum spacing between consecutive digits
  • <max_spacing>: maximum spacing between consecutive digits
  • <image_width>: width of the generated image

The generated image is saved in the current directory as a < sequence >.png, for example, 1-2-3.png.

License

MIT License

Contribute

Want to work on the project? Any kind of contribution is welcome!

Follow these steps:

  • Fork the project.
  • Create a new branch.
  • Make your changes and write tests when practical.
  • Commit your changes to the new branch.
  • Send a pull request.