Link Task can be found here. Link to download the .zip file, which contains the rotated pictures, is also here
The problem is to us CNN to identify images orientation and then, rotate them to the upright position. To solve this problem I use a model given by Keras, the same one to solve the cifar10 problem. The main step of the process was to prepare the data to the model. I used a lib called 'imageio' to read the images and transform them into vectors, after some preprocessing they were ready to be used on the model.
Train the model took actually a lot of time, I adjusted the dropout parameter and then I put 20 epochs and the results were really impressive right on the 3 epoch. The final result was ~97%. Really good. The results is on the folder, right here. I also got the rotated images, they can be found on the link above in a .zip file. As the file was really big, I could not push them togethe with the others.
I created a folder name Large Files so I can put all the pictures and stop the git noise because of the large files. But, to run main.py, you will need to create this folder and put under it the 'test.rotfaces' and the 'train.rotfaces'. If that happens, everything should be fine!
It's preatty easy. You just have to run the file 'main.py' and that's it! When you run it, it will build up all the network and train it all over again. The model has already been saved and you can uncomment the piece of code that loads the model and and comment the other piece which train the model, that will speed up things. Comments to guide this were introduce on the code.