Abstraction and Reasoning Challenge - Kaggle
Shortly: Abstraction and Reasoning Challenge - problem which requires only few, really easy for people, examples as training data-set to solve a task.
Basicly, any algoritms used in Kaggle challenge didn't score any satisfying results. We wanted to check if increasing data set, creates better results.
This program is dataset generator for Kaggle's problem.
To use this this generator you have to:
- Open run.py file
- Comment/uncomment selected problem.
(...) # run_snake() run_labyrinth()
- Set params in
generate_and_save()
function or leave default. - Run run.py file. Now you've got created whole dataset.
To create new generator, you have to:
- Create new ProblemNameGenerator, which overwrites GeneratorInterface, and create functions:
generate_input()
andgenerate_output()
- Create new ProblemNameVariationsGenerator, which overwrites VariationsGeneratorInterface, and create function:
generate_all()
, which return array of ProblemGenerators. You have to iterate over all solutions yourself. - Create new function
run_problemName()
(similarly to previous functions) in run.py file and run this function. - Wait until your new Problem dataset is generated.
If you want to visualise created data, you can open Testing Interface. Press 'Choose file' button, and select file you want to see.
There are example files to visualise:
There is a visualization of mentioned above problems:
There is also CNN Jupiter Notebook Solution - solution based on CNN architecture used during tests.