face-recognition-cnn

Deep Convolutional Network for Face Classification.

Using a dataset of 200 identities in total, this project will present possible solution to build a classifier using CNNs implemented with PyTorch. We tested out three different architectures:

1. The first model presented is the baseline model we were provided. This model is only used as a guideline for what our other models are at least supposed to perform.
2. TronNet, is an extension of what we were provided in the baseline. We made it deeper, in hope of it learning more advanced features.
3. we are using transfer learning to initialize a ResNet18 model. Here we are downloading a pre-trained model while switching out the fully connected layers to fit our problem of 200 different identities.

This project includes:

• Image pre-processing such as: normalization and rotations.
• Xavier weight initialization.
• Our new cnn architecture for solving face classification problem - TronNet.
• Weights Visualization of selected layers.
• Loss Visualizations and comperisons.

Setup

Install the dependencies in the requirements.txt file.

Overview

cnn.py: Main class, this is also where the training and testing happens

models.py: All of our Pytorch models is located here

utils.py: Helper methods

Configuration

Most of our parameters can be configured using command line.

Our default settings is listed below

SETTINGS = {
    'EPOCHS': 50,
    'BATCH_SIZE': 64,
    'LR': 0.001,
    'DECAY': 0,
    'NUM_CLASSES': 201,
    'RANDOM_SEED': 42,
    'WLOSS': True,
    'K-FOLD': False,
    'K-FOLD-NUMBER': 2, # if k-fold is enabled
    'NNET': None,
    'TRANSFORMER': "default",
    'DATA_PATHS': {
        'TRAIN_CSV': 'train.csv',
        'TEST_CSV': 'test.csv',
        'DATASET_PATH': './datasets/cs154-fa19-public/'
    }
}

To run with a small data set, add the parameter --mini True