Contains a Convolutional Neural Network (CNN) that is trained to predict missing pixels from images. The missing pixels are stand-alone seperated by an x- and y-spacing which have for the training values between 2 and inlcuding 6 pixels. The missing pixels are starting at an x- and y-offset which have for training values between 0 and 8 pixels. The final accuracy measured as the root mean squared error of the predicted RGB values were around 18.42 pixels.
Usage:
python main.py working_config.json
example_project |- architectures.py | Classes and functions for CNN network architectures | You may select 'SkipCNN' in main file as architecture but this still contains dimension errors |- datasets.py | Dataset creationg classes: provide trainingset in a folder called 'training' | If you remove 'transform_chain_raw' in the train indices when creating the dataset, it will perform data augmentation on the trainingset | uncommenting the code for the full dataset and commenting out the subset code will use all of your image files provided in the training folder | You may choose a random subset size for faster computation or debugging in the subset dataset code | You may provide an inputs.pkl test file for running on unknown data |- main.py | Main file. In this case also includes training and evaluation routines | Uncommenting the rows right below the while training loop will create a new transformed trainingset at the beginning of every epoch (mind the computation time) | Will create a results folder containing test plots, the tensorboard files(e.g. for parsing in web) the best_model.pt file and the best losses in a results.txt file |- Datareader.py | Reads through all files given in the dataset and creates the known_array (containing True/False if pixel is known), input_array (with missing pixels) and target_array (not needed in this build because we compare to full image in evaluation) |- eval_testset.py | If inputs.pkl file for testing reasons is given in this format, run it with this file (best_model.pt is chosen) | Returns predictions.pkl file with all the predicted images |- utils.py | Contains the plotting function for the main file |- working_config.json | Containing the hyperparameters for the main function. Can also be done via command line arguments to main.py. | You may change your device to 'cpu' if you don't have cuda available