Facial and gender classification of actors and actresses from the FaceScrub database, with linear regression models written from scratch. Please see faces.pdf for the report.
You must download the dataset before running any parts of the code.
./classification.py --download --log INFO
This will download from the FaceScrub dataset into two directories under dataset/: uncropped (raw images) and cropped (grey-scaled images).
./classification.py --part 2
This process creates non-overlapping, reproducible (seeded) datasets for the actors/actresses Drescher, Ferrera, Chenoweth, Baldwin, Hader, and Carell.
./classification.py --part 3 --log INFO
The implementation details are in the PDF. The optimal parameters for gradient descent were determined by using the --optimal option. A sample report of that is under logs/optimal_params. Those optimal parameters are then used in gradient descent to determine the optimized theta value, which is subsequently saved into part3.pkl.
The classifier is then used to classfy 10 images of hader and carell, and the results are printed to the terminal.
./classification.py --part 4 --log INFO
This visualizes the optimized theta by loading the saved theta value from part 3, for both the full training set and a two-image training set.
./classification.py --part 5 --log INFO
Trains the classifier on images of 6 actors / actresses (100 each) and then tests the performance on images of 6 different actors / actresses (10 each). The size of the training size is varied from 10 to 70 and the performance is plotted.
./classification.py --part 7 --log INFO
Trains the classifier on images of 6 actors / actresses (70 each), each of which are represented by one state of the one-hot-encoder. The performance is tested on 10 images of each respective actor.
./classification.py --part 8 --log INFO
This reproduces images of each dimension of theta from above, representing each state of the one-hot-encoder.