This project predicts missing parts in an image using a simple CNN.
It is based on the sample project provided by Michael Widrich, but heavily modified.
The input dataset was created from thousands of user images
submitted in the first exercise of the semester (not included),
but can create a new dataset from any kind of greyscale images.
Simply call the main function with any configuration json file:
python3 main.py <config>.json
debug_config.json trains less, print more stats and plots more often.
For the final predictions working_config.json was used.
python2_challenge
|- architectures.py
| Classes and functions for network architectures
|- datasets.py
| Dataset classes and dataset helper functions
|- debug_config.json
| A modified copy of the working_config with smaller parameter values for debugging
|- main.py
| Main file
| Also includes training and evaluation routines
|- README.md
| This description
|- utils.py
| Utility functions and classes
| In this case contains a plotting function and a de-normalization function
|- working_config.json
| The configuration file used to train the model and make final predictions
After about 5000 training iterations (with batch-size 16) the plotted output might look something like this:
