Results

Hardware : GTX 1060 took roughly 4.5 hours to complete 28k iterations/7 epochs) p.s : Disk stored dataset was horrible the first epoch due to what I believe to be TLB misses. After the first epoch, disk reads and file i/o buffers helped training go way faster.
CASIA (training) : ~0.98 training accuracy, ~1.92 Training Loss, 10575(CASIA) + 1(me) classes
LFW (evaluation) : 0.9902 accuracy

File Tree

├── live_inference.py '(Live inferencing of the model using laptop camera)'
├── cosFace
│   ├── casia_train.py '(MAIN TRAINING SCRIPT)'
│   ├── casia_train_summary_CosFace_20200625-115034 '(Tensorboard metrics)'
│   │   └── events.out.tfevents.1593111034.joshuakang-Alienware
│   ├── checkpoint 
│   │   ├── accuracy.npy
│   │   ├── casia_train.py
│   │   ├── dataLoader.py
│   │   ├── faceNet.py
│   │   ├── images.png
│   │   ├── inference.py
│   │   ├── lfw_eval.py
│   │   ├── loss.npy
│   │   ├── matlab_cp2tform.py
│   │   ├── net_1.pth '(model after nth epoch)'
│   │   ├── net_2.pth
│   │   ├── net_3.pth
│   │   ├── net_4.pth
│   │   ├── net_5.pth
│   │   ├── net_6.pth
│   │   ├── net_7.pth
│   │   ├── netFinal_8.pth
│   │   ├── trainingLog_0.txt '(models training loss during nth epoch)'
│   │   ├── trainingLog_1.txt
│   │   ├── trainingLog_2.txt
│   │   ├── trainingLog_3.txt
│   │   ├── trainingLog_4.txt
│   │   ├── trainingLog_5.txt
│   │   ├── trainingLog_6.txt
│   │   └── trainingLog_7.txt 
│   ├── data
│   │   ├── casia_landmarkMTCNN.txt '(CASIA face landmarks generated from /mtcnn/get_landmarks.py + joshs faces)'
│   │   ├── casia_landmark.txt '(CASIA face landmarks as generated from other source)'
│   │   ├── josh_landmarksMTCNN.txt '(Facial landmarks for my face)'
│   │   ├── lfw_landmark.txt '(LFW dataset facial landmarks generated by /mtcnn/get_landmarks.py)'
│   │   └── pairs.txt '(LFW given pairs for Face cosin similarity in evaluation)'
│   ├── dataLoader.py '(dataloader for model. Landmarks file with labels required)'
│   ├── faceNet.py '(Models)'
│   ├── random_10_inference.py '(creates embeddings for random faces. For Tsne visualization mostly)'
│   ├── lfw_eval.py '(evaluation script on lfw dataset)'
│   ├── matlab_cp2tform.py
│   ├── plot_graphs.ipynb
│   └── __pycache__
│       ├── dataLoader.cpython-37.pyc
│       ├── faceNet.cpython-37.pyc
│       └── matlab_cp2tform.cpython-37.pyc
├── mtcnn
│   ├── example.jpg
│   ├── get_landmarks.py '(get landmarks for images in dataset and label)'
│   ├── LICENSE
│   └── src
│       ├── box_utils.py
│       ├── detector.py 
│       ├── first_stage.py
│       ├── get_nets.py
│       ├── __init__.py
│       ├── __pycache__
│       │   ├── box_utils.cpython-37.pyc
│       │   ├── detector.cpython-37.pyc
│       │   ├── first_stage.cpython-37.pyc
│       │   ├── get_nets.cpython-37.pyc
│       │   ├── __init__.cpython-37.pyc
│       │   └── visualization_utils.cpython-37.pyc
│       ├── visualization_utils.py
│       └── weights
│           ├── onet.npy
│           ├── pnet.npy
│           └── rnet.npy
└── README.md

No usage of out of the box implementations (some changes to mtcnn) or usage of opencv (cuz I'm a boss). Other than cv2 usage for camera accessibility.

MTCNN for facial landmark detection

Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks.

20 residual Layer'ed (batch-norm) CNN using cosFace loss function

[CosFace: Large Margin Cosine Loss for Deep Face Recognition] (https://arxiv.org/pdf/1801.09414.pdf) Fully trained model on CASIA + 215 images of Me (could have done transfer learning, but meh this is way cooler). Can then finetune with more images of me.
Model : (https://drive.google.com/file/d/1UJW8chHcD8KEl28yGSy3vwD2KzOMb1em/view?usp=sharing)