We decided to tackle image colorization. We used jupyter notebook at first, before moving to standard .py files. Keras, skimage, flask, and numpy were Python module we wanted to work with. Flask was used for API and GUI.

Before doing anything, a dataset needs to be in place. It was too large to put on GitHub so it is linked below.

Our mission docs are located in /missions/NAME where NAME is team member's name.

Documentation is located in /documentation/ folder.

Alpha version of our baseline can be seen in the alpha_version_notebook.ipynb or alpha version snapshot link

Beta version of our baseline can be seen in baseline.ipynb or beta version snapshot link

Final version of the baseline (current implementation) is split across main.py, model.py, colorize.py and utils.py

Alpha version works with one photo at a time. Epochs value works best around 1000. Its performance is bad with cold color pallet or when a different image is used for testing and for predicting.

API needs static and templates folder to be in place to function correctly, as well as all python files.

To run the API, simply run "python api.py". Afterwards, a localserver should be created.

127.0.0.1:5000/ is the default link it creates. After accessing it, a list of possible options should appear. It can take a while to train a model so if a page is loading in for a while, check out output logs.

To train the model, test its performance, and run some images through it: run main.py (it is esentially the training loop). To run some of the functions independently, utils.py can be a great resource to check out. Colorize.py is essentially what the user would see. It loads in the model and colorizes the selected photos. To see the gui, run gui.py (requires template folder to work).

To ensure everyting works with the GUI, Result/, Test/, Train/ and Example Results/ folders were moved to "static/" folder. This was done to ensure that flask can access those files easily, allowing for easier integration with the API and GUI. Template folder was also created to ensure html code is ready to go when flask needs it.

Required Libraries:

• keras
• skimage
• numpy
• flask

To install them:

pip install keras skimage numpy flask

Os and json, which are also required, are usually included with python.

Other setup:

There are some default variables set around the program and its functions. Below is a list of those functions and their values. Users can change them freely in the code. Parsing arguments was part of the team missions, which we sadly had to drop but I feel like this is an ok alternative.

1. load_images loads images from static/Train/ folder by default.
2. when training, tensorboard is only used when last parameter is False. Across 4 different machines I had varring results (based on installation, OS and so on) with getting callbacks to work, so it is disabled by default.
3. when saving the model, the default name is "model". This means the model is saved as model.json and its weights are saved as model.hdf5.
4. When preparing accuracy visualisation images a.k.a testing the model, the default folder to load the images from is "static/Test/".
5. When images are being saved, they are saved under "static/Result/".

Dataset used: https://www.floydhub.com/emilwallner/datasets/colornet This dataset contains almost 10 thousand photos, all 256x256, that I used for the training session.

Alpha version was based on: https://www.freecodecamp.org/news/colorize-b-w-photos-with-a-100-line-neural-network-53d9b4449f8d/