Fine-tuning AdaFace on images from news papers

This is a fork of the AdaFace (https://github.com/mk-minchul/AdaFace) repository.

This repository contains a modified version of the AdaFace repository. The main modifications were done to the following files:

evaluate_utils.py
requirements.txt
train_val.py
utils.py
run_ir50_ms1mv2.sh

The modifications were done to support a different dataset than what was originally used to train the model and especially to make the model runnable on the metacentrum clusters.

Usage guide

The model was fine-tuned and tested on metacentrum. We have had issues when running the scripts on various different nodes. We have had success with the zia and galdor clusters. Other ones didn't work.

The first step is to create a singularity (if used on metacentrum). To do this, run the following command:

singularity shell --nv /cvmfs/singularity.metacentrum.cz/NGC/PyTorch\:21.09-py3.SIF

It is important to use this singularity, because the scripts only support python3.8 and this singularity comes with this particular python distribution. (we tried our best to make it run on newer python versions but we failed)

Installation

The original requirements.txt file didn't work on metacentrum at all. When we tried fixing it, we were unsuccessful. The only solution we found was to manually install some packages one by one using pip. Pasting in the sequence of commands below should install all required packages without issues.

conda create --name adaface pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=10.2 python=3.8.10 -c pytorch
source /opt/conda/bin/activate adaface
conda install matplotlib 
pip install nvidia_cublas_cu11==11.10.3.66 --no-cache-dir
pip install nvidia_cudnn_cu11==8.5.0.96 --no-cache-dir
pip install opencv_python_headless==4.9.0.80 --no-cache-dir
pip install scikit_learn==1.3.2 --no-cache-dir
pip install PyWavelets==1.4.1 --no-cache-dir
pip install aiohttp==3.9.5 --no-cache-dir
pip install menpo --no-cache-dir
pip install -r requirements.txt --no-cache-dir

Testing

To test the model, make sure your pre-trained model adaface_ir50_ms1mv2.ckpt is stored in the /pretrained/ directory. Then you need to prepare your testing dataset in the .bin format. Name it lfw.bin and place it in the /data/faces_webface_112x112/ directory. Afterward, run the following command to prepare the testing set to be used:

python3.8 convert.py --rec_path ./data/faces_webface_112x112 --make_validation_memfiles

After that, the data set will be ready, and you can run the testing via the following command:

python3.8 main.py --data_root ./data/ --val_data_path faces_webface_112x112 --prefix adaface --use_mxrecord --start_from_model_statedict ./pretrained/adaface_ir50_ms1mv2.ckpt --use_16bit --gpus 1 --evaluate --arch ir_50

The results of the test will be printed to stdout and saved to a new file in the /experiments/ directory.

Fine-tuning

To fine-tune the pre-trained model on style-transferred data, prepare your training dataset in the .rec format, alongside the .lst and .idx files and place it into /data/faces_webface_112x112/train/. The files must be named train.rec, train.idx and train.lst. Then place your validation data set into the /data/faces_webface_112x112/val/ directory in the .bin format. It has to be named lfw.bin. After that, it must be preprocessed using the following command:

python3.8 convert.py --rec_path ./data/faces_webface_112x112/val --make_validation_memfiles

Once all that is done, place your fine-tuned model adaface_ir50_ms1mv2.ckpt into the /pretrained/ directory and run the command

./scripts/run_ir50_ms1mv2.sh