LightCNN

This model is borrowed from A Light CNN for Deep Face Representation with Noisy Labels. The PyTorch version original code can be found here. The official and original Caffe code can be found here.

Installation

• Install PyTorch following the website.
• Clone this repository:

git clone https://github.com/duongminhhieu16/LightCNN

• I currently run it on Python 3.9 with CUDA 11.4.

Datasets

• Download MS-Celeb-1M clean list: Google Drive
• Or download VGG-Face2 dataset Google Drive
• Install gdown to download datasets from Google Drive:

pip install gdown
gdown <link-to-the-dataset>

Training

• Prepare the dataset:

python process_data.py

• After running process_data.py, there will be two files: train_list.txt and test_list.txt.
• Trainng from scratch or from checkpoint:

python train.py

• Checkpoint: Training 500 epochs with VGG-Face2 dataset using SGD optimizer: learning_rate=0.0001, momentum=0.95, weight_decay =1e-5.