duyngudao / arcface_pytorch

Distributed Arcface Training in Pytorch

Requirements

RAM > 10G
GPU

How to Training

To train a model, run train.py with the path to the configs.
The example commands below show how to run distributed training.

python -m torch.distributed.launch --nproc_per_node=1 --nnodes=1 --node_rank=0 train.py configs/wf_mbf

config.num_classes = 10572 config.num_image = 501196

Download Datasets or Prepare Datasets

• WebFace0.5M (10572 IDs, 501196 images)

Guide training on Kaggle

1. install network

!conda install -y gdown
!pip install -U --no-cache-dir gdown --pre

2. Clone backbone train

!git clone https://github.com/DuyNguDao/ArcFace_Pytorch.git

3. Download dataset and unzip

Note: change id gdrive of dataset

!gdown --id 1yaLoTdjybeLtXLsODjsJCqPRGvlk2a9k
!unzip /kaggle/working/faces_webface_112x112.zip

4. Change config

%load ./configs/wf4m_mbf.py
Change:
+ config.rec = "path of dataset" 
+ config.num_classes = xxx
+ config.image = xxx
+ config.epoch = xxx
+ config.batch_size = xxx
+ config.val_targets = ['lfw', "agedb_30"]: datatest (lfw, fcp-cp, agedb_30)
writefile:
%%writefile ./configs/wf4m_mbf.py

5. Install environments

!pip install -r requirement.txt

6. Training on GPU

!python -m torch.distributed.launch --nproc_per_node=1 --nnodes=1 --node_rank=0 train.py configs/wf4m_mbf

7. Zip and download

!zip -r result.zip ./work_dirs