Digit Sequence Generator

generate_numbers_sequence.py

A script and API for generating and augmenting a sequence of digits based on a specified input list of digits. The purpose of this code is to aid in training classifiers and generative deep-learning models.

generate_numbers_sequence.py is a semi-vectorized implementation:

• Time is saved by vectorizing the method to generate an image.
• Space is saved by recomputing the method to generate an image for every digit in an input list.
  
  

Set-up

1. Clone the repo in a local directory:

git clone https://github.com/arikanev/digit_sequence_generator.git

2. cd into digit_sequence_generator, and de-compress data files:

cd digit_sequence_generator

unzip -a data.zip

Running

Running generate_numbers_sequence.py with augmentation saves a pair of image sequences with the following filenames:

• 'sequenceX.png'
• 'aug_sequenceX.png'

(Where X is an integer denoting number of existing sequence files + 1)

These files contain the exact same digit images in their sequences, and differ only by RGB and Greyscale value.

To run as a script:

python generate_numbers_sequence.py -d DIGITS (space-separated ints) -r SPACING_RANGE (two space-separated ints) -w IMAGE_WIDTH (int)

optional arguments:

-a AUGMENTATION (str)

Example: python generate_numbers_sequence.py -d 4 5 2 1 3 4 4 3 2 1 2 4 -r 5 40 -w 300 -a mnistm

Currently the augmentation supported is 'mnistm', which consists of mnist masks super-imposed on imagenet image backgrounds. A full mnistm dataset can be found here.

(As of now you can only access augmentation options when running generate_numbers_sequence as a script.)

To call as an API in python code:

import generate_numbers_sequence

sequence, height = generate_numbers_sequence.generate_numbers_sequence([digit_list], (range_tuple), width_int)

Example: sequence, height = generate_numbers_sequence.generate_numbers_sequence([4, 5, 6, 2, 3], (1,30), 100)

generate_numbers_sequence.generate_numbers_sequence(d, r, w) returns 2 values:

• A numpy array of size (height, width_int) and dtype float32.
• An int value representing sequence height.

(The height value is translated unchanged from the sampled digit image height)

Testing

To test the above API and script, run python run_tests.py.

In the future, tests should be added to:

• Assert the shape match between sequenceX.png and aug_sequenceX.png.
• Ensure lack of runtime errors when generating images from other datasets. (For generalizability/extensability)
  
  

Future plans

Could/Should focus on:

• Modifying mnistm augmentation to clearly visualize the full digit sequence.
• Adding parse code to enable reading a sequence of digits without spaces, or with commas.
• Adding an option for sequence margins to be extended, as opposed to stretching the entire image.
• Expanding on the number of augmentation methods.