facenet

Table of Content

• FaceNet
• W&B
• AT&T Faces
• LFW

FaceNet

FaceNet is a face recognition system developed in 2015 by researchers at Google that achieved then state-of-the-art results. It maps the face images to euclidean space and learns on the L2 distance between the embeddings. Our paper notes on FaceNet can be found here.

This is a pytorch implementation of FaceNet paper with ResNet as the backbone architechture. At first the implementation was done on AT&T Dataset of Faces, then on LFW Dataset. We used online triplet mining method for selecting triplets.

W&B

Wandb was used throughout this part of the project for metric tracking, hyperparameter tuning, sweeps, visualization, etc.

(a) Metrics of ResNet18 on LFW (b) Sweeps of ResNet18 on ATT

AT&T Faces

The dataset was split in 35 training classes and 5 test classes

Training

Parameter	Value
Architechture	ResNet18
Embeddings Dimension	64
No. of Learnable Parameters	11,209,344
Epochs	200
Learning Rate	0.0002
Optimizer	Adam
Batch Size	100
Margin	1

Results

Results	Train Set	Test Set
Accuracy	1.0	0.984
Recall	1.0	0.978
Precision	1.0	0.936
ROC area under curve	1.0	0.981
Euclidean Distance Threshold	0.91	0.89

Plots

(a) Epoch Loss. (b) EER Curve. (c) t-SNE Emdeddings.
(d) ROC Curve on train set. (e) ROC Curve on test set

LFW

Deep Funneled set of LFW images was used for training and evaluation purpose.

The faces were extracted by center crop and then resized to match input shape. Further they were normalized overall data's mean and standard deviation.

MEAN = torch.Tensor([0.5929, 0.4496, 0.3654])
STD = torch.Tensor([0.2287, 0.1959, 0.1876])
transform = transforms.Compose([
    transforms.CenterCrop((128,98)),
    transforms.Resize((224,224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=MEAN, std=STD),
])

LFWDataset.py contains the custom dataset classes for loading LFW data in all configurations. This dataset class was later contributed to Torchvision library.

Training Configuration

	Architechture	Embeddings Dimension	No. of Learnable Parameters	Epochs	Learning Rate	Batch Size
Training	ResNet-18	128	11,242,176	200	0.002 (Reduced by factor of 2 every 50 epochs)	256

To train run

train.py --config configs/resnet18lfw.yml --data_dir ../datasets/lfw  --wandb true

To resume training

train.py --config configs/resnet18lfw.yml --data_dir ../datasets/lfw  --wandb true --resume "checkpoints/model_resnet18_triplet_epoch_120_08-Dec 15:57.pt" 

Model State Dict

state = {
    'epoch': epoch+1,
    'embedding_dimension': p.fc_layer_size,
    'batch_size_training': p.batch_size,
    'model_state_dict': model.state_dict(),
    'model_architecture': p.backbone,
    'optimizer_state_dict': optimizer.state_dict(),
    'scheduler_state_dict': scheduler.state_dict(),
    'best_distance_threshold': best_threshold,
    'accuracy':accuracy
}

Results

Accuracy	Precision	Recall	ROC Area Under Curve	Euclidean Distance	TAR @ FAR=1e-2
88.35%	88.46%	88.23%	0.9508	1.104	61.07%

Plots

(a) Epoch Loss (b) ROC Curve (c) Accuracy (d) Learning Rate