Weikang Shi, 2020011365, shiwk20@mails.tsinghua.edu.cn
- dlib
- imutils
- mtcnn
- numpy
- omegaconf
- opencv-python
- Pillow
- scikit-learn
- tensorflow-gpu
- torch==1.11.0
- torchvision==0.12.0
- tqdm
Firstly create a conda environment by conda create -n FaceRecg python=3.8, then Install the necessary packages using pip install -r requirements.txt.
You can align face image by mtcnn or landmark. For mtcnn, run bash face_detect/run_mtcnn.sh to get mtcnn face detection results by 8 gpus, then run face_detect/process_mtcnn.py to merge mtcnn results and evaluate them. For landmark, run python face_detect/landmark.py to get landmark face detection results and evaluate them.
- ckpt
shape_predictor_68_face_landmarks.datfor landmark face detection can be downloaded here: url. put it undercheckpoints/.
Then, run python face_align -t mtcnn or python face_align -t landmark to align images of both training_set and test_pair. The default crop size for mtcnn is (112, 96), for landmark is (112, 112). you can edit it in face_align.py.
You can download the face detection and face align results here: url. Uncompress data.tar.gz and put it In the root directory.
The file structure of data dir is:
data/
├── test/
├── landmark/
├──align112x112/
├──eval_landmark/
├──landmarks/
├── mtcnn/
├──align112x96/
├──eval_mtcnn/
├──mtcnns/
├── test_pair/
├── train/
├── landmark/
├──align112x112/
├──eval_landmark/
├──indexes/
├──landmarks/
├──error_lmk.json
├── mtcnn/
├──align112x96/
├──eval_mtcnn/
├──indexes/
├──mtcnns/
├── training_set/
├── dataset.zip
{align_type}refer tomtcnnorlandmark.
Run bash train.sh to train the model:
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=23756 train_ddp.py --type {align_type}
Also, you can edit configs/train_{align_type}.yaml to change training params, like batch_size, max_epochs, val_interval and triplets_num.
The pretrained models can be downloaded here: url. After download, put landmark_last_acc_0.8549.pth in log\ckpt\landmark_last_acc_0.8549.pth; put mtcnn_last_acc_0.8693.pth in log\ckpt\mtcnn_last_acc_0.8693.pth.
What's more, My training logs for the pretrained models are in log\logs\train_{align_type}.log.
Run python evaluate.py --type {align_type} --val_ratio 0.85 --seed 0 to evaluate model. My evaluate logs for the pretrained models are in log\logs\evaluate_{align_type}.log.
Run python test.py --type {align_type} to get test results. My test logs are in log\logs\test_{align_type}.log and the test results are in test_{align_type}.txt.