CNN Project to recognise famous people.

Covering from data loading to training/testing and computing predictions on unseen images.

Input Image

RGB Channel Decomposition

First Convolution Feature Maps

Second Convolution Feature Maps

Files explained:

• data_loading.py: This file contains the code for loading the data for testing/training and predicting. It includes transformations of the data, loading into test and train splits and batching.

• download_data.py: Script to download the data (old one, changed from casia web-face to imdb_wiki).

• main.py: contains all the code to run test/train or pred. Ran through args. It also contains the optimizer and the loss function.

• model.py: Contains the simple model built for this project. A CNN that doubles the amount of channels every convolution while it reduces the features of the image in half every convolution. It has 347140 parameters with the current image loading code.

• plot.py: Contains the code to create the plots after training.

• prediction.py: Contains the code to run predictions on a single image found in /prediction_img

• test.py: Code for the test steps

• train.py: Code for the train steps

• utils.py: Contains accuracy_fn

Training plot over +65 epochs

Disclaimer: This is a very simple project, I haven't trained nor improved the model as much as possible, due to time constraints, and to GPU constraints too. Accuracy might not be the best.