Lightweight Face Recognition via Shrinking Teacher-Student Networks

This repository contains an implementation of the distillation method mentioned in this paper. Using the code from this repository, you can train a lightweight network to recognize faces for embedded devices.

Data preparation

1. Download dataset from here.
2. Extract images using this.
3. Get the labels for each image and create annotations.csv using: create_annotation.ipynb
4. Or get download dataset from here to skip above 2 steps.
5. Create a folder in home directory by the name of CASIA_dataset
6. Save the Images Folder and annotations.csv in it

Training

Create folder for saving model:

mkdir output_logs

Start training using:

python3 train.py

Additional flags

Flag	Default Value	Description
batchsize	8	batch size
epochs	10	number of epochs to train
lr	0.001	learning rate
lmda	0.25	lambda
device	cpu	device using (eg. cuda , mps)
save-dir	output_logs	folder to save model
workers	0	number of workers

Evaluation

Evaluate using:

python3 eval.py

Additional flags

Flag	Default Value	Description
batchsize	8	batch size
device	cpu	device using (eg. cuda , mps)
workers	0	number of workers
test-size	0.1	size of dataset for evaluation
model-path	None (required)	path of model for evaluation

Performance:

CASIA-WebFace

Trained for 30 epochs
Evaluated on 50k random samples

Lambda	0	0.25	0.75
Accuracy	98.87%	98.95%	99.08%

Using HPC

Paramshivay, Supercomputer of IIT BHU (Varanasi) was used for training of this model.

Setting Up Anaconda

Install miniconda locally for more reliable experience and easy package managing.

Download Miniconda

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

Install Miniconda

bash Miniconda3-latest-Linux-x86_64.sh

Activate Miniconda

source miniconda3/bin/activate

Using Interactive Terminal

To get a node with 1 gpu and 20 cores for 1 hour for running the code live.
(Interactive session will over as soon as time is over or your terminal is disconnected from HPC and execution will be interupted.)

srun --partition=gpu --nodes=1 --gres=gpu:1 --time=01:00:00 --ntasks-per-node=20  --pty bash -i

Using Shell Script to Submit the Job

Submit the Job to SLURM for completion.
(No need to be connected to HPC for file to keep executing)

sbatch script.sh

Contributors

• Raunak Pandey
• Rakshit Sawhney