DLCV-final: Face Recognition

• Kaggle link: Kaggle
• Team Members: R06942120, R06921076, D06921016

---

Task description

1. Beat TA’s Baseline
   • You will be given a dataset of interest (with predetermined data split)
   • You are NOT allowed to apply any external dataset or techniques like transfer learning
2. Squeeze Your Model
   • Design a model (e.g., those with fewer parameters, simpler designs, compact or simplified versions) which would achieve comparable performances but save computation or storage costs.

---

Model descrioption

Baseline model - Inception ResNet

• Network borrowed from Github repo
  • Related paper 1: "FaceNet: A Unified Embedding for Face Recognition and Clustering"
  • Related paper 2: "Deep Face Recognition"
• Center loss is utilized for preventing overfitting
  • Paper: "A Discriminative Feature Learning Approach for Deep Face Recognition"
• Hard instance mining to prevent overfitting those easy-sample
  • Paper: "Focal Loss for Dense Object Detection"
• Triplet loss is also utilized on additional embedding layer as multi-task training tricks
  • Paper: "FaceNet: A Unified Embedding for Face Recognition and Clustering"
  • Code borrowed from Github: tensorflow-triplet-loss
• Training policy - multi-stage training
  1. Train without data augmentation: python train_teacher.py --finetune_level 0
  2. Fine-tune with basic data augmentations: python train_teacher.py --finetune_level 1
     • Including rotation, horizontal flip, scale, crop, hue, contrast, brightness, gray-scale
     • Center loss is included with weighting factor 1e-5
     • pre-logit normalize with weighting 1e-5 also included in stage 2
  3. Fine-tune with seaweed augmentation: python train_teacher.py --finetune_level 2
     • Triplet loss is included
     • Learning rate is decayed to 5e-5
     • Weighting of center loss, PLN loss is increased to 1e-4

---

Compressed model - SqueezeNeXt

• "SqueezeNext: Hardware-Aware Neural Network Design"
  • Implemented SqNxt-23v5 following this repo

---

Teacher Student training

• Github
  • Related paper: Distilling the Knowledge in a Neural Network(2015)

---

Result - Model comparison

	Model size	# of params	P. V.	P. T.	fps	weights
In.-Res.	124MB	26,781,288	88.91%	85.59%	418.96	link
2.0 SqNxt-23v5	13.7MB	3,399,352	71.28%	~	635.68	None
2.0 SqNxt-23v5(T-S)	13.7MB	3,399,352	85.98%	82.48%	635.68	link
1.0 SqNxt-23v5(T-S)	4.5MB	1,106,456	78.42%	73.6%	805.36	link

• T-S refers to Teacher-Student training strategy
• P. V. refers to Performance on Validation set
• the T-S weight is for fine-tuning, thus contains weights of embedding layer

---

Ablation study

Basic A.	Seaweed	Center L.	P. N. L.	H. I. M.	Triplet L.	P. V.	P. T.
✖️	✖️	✖️	✖️	✖️	✖️	~30%
✔️	✖️	✖️	✖️	✖️	✖️	67.4%
✔️	✔️	✖️	✖️	✖️	✖️	72%
✔️	✖️	✔️	✖️	✖️	✖️	75.75%
✔️	✔️	✔️	✖️	✖️	✖️	78.11%	79.11%
✔️	✔️	✔️	✔️	✖️	✖️	81.81%	82.45%
✔️	✔️	✔️	✔️	✔️	✔️	88.91%	85.59%

• Basic A. refers to basic augmentations
• L. refers to loss
• P. N. L. refers to pre-logit norm loss
• H. I. M. refers to hard instance mining
• P. V. refers to Performance on Validation set
• P. T. refers to Performance on Test set (scores on Kaggle)

---

Usage

• Use train_teacher.py to train teacher network with args
• Use train_student.py to train student under normal training policy with args
• Use train_TS.py to train student network under TS leraning policy with args
  • Trained teacher is required