HW_face

This project is based on mtcnn and facenet.

We achieve a wonderful result of top1 acc 98.05% and top5 acc 99.45%.

The results can be found in predictions.txt. Or you can follow the instructions below to reproduce training.

We offer trained model on Clound Storage. You can download it and skip sub-steps 1, 2, 3 in Face model multi-train step.

Requirements

Tensorflow>=1.7

Python2.7 or 3.5

Other python packages you can install when running.

Pre-process

set PYTHONPATH environment to this root directory like this on linux:

export PYTHONPATH=path/to/this/root/directory

Prepare data

Move HW_1_Face data to dataset/HW_1_Face/raw

Change the data structure by running

python dataset/HW_1_Face/raw/sort.py

Face detection

Using mtcnn to align face bounding box. Run train set:

python align/align_dataset_mtcnn.py \
dataset/HW_1_Face/raw/train/ \
dataset/HW_1_Face/mtcnn_160/train \
--image_size=160 \
--margin=32 \
--gpu_memory_fraction=0.85

and test set:

python align/align_dataset_mtcnn.py \
dataset/HW_1_Face/raw/test/ \
dataset/HW_1_Face/mtcnn_160/test \
--image_size=160 \
--margin=32 \
--gpu_memory_fraction=0.85

to get face detection of the data

Face model multi-train

If you want to train the model on your own, please follow the multi-train step below.

Download pretrained model on vgg-face2 from FaceNet's github and save it to saved_models/vgg_face2

Fix Inception-V1 vars and only finetune the last fc layer on HW_Face dataset by running:

python train_softmax.py \
--logs_base_dir=logs/softmax_finetune \
--models_base_dir=saved_models/vgg_face2_finetune_softmax/ \
--pretrained_model=saved_models/vgg_face2/modelname \
--data_dir=dataset/HW_1_Face/mtcnn_160/train \
--test_dir=dataset/HW_1_Face/mtcnn_160/test \
--max_nrof_epochs=100 \
--gpu_memory_fraction=0.8 \
--batch_size=100 \
--epoch_size=20 \
--embedding_size=512 \
--learning_rate=0.1 \
--learning_rate_decay_epochs=20 \
--learning_rate_decay_factor=0.8 \
--save_every_n_epochs=10 \
--validate_every_n_epochs=1 \
--keep_probability=0.5 \
--weight_decay=0.01 \
--center_loss_factor=1.0 \
--center_loss_alfa=0.95 \
--is_finetune \
--load_without_fc

Remember to change --pretrained_model to the one you download from github.

This will lead to about top1 acc of 0.62 on test set.

Finetune all vars on HW_Face by running:

python train_softmax.py \
--logs_base_dir=logs/softmax_finetune \
--models_base_dir=saved_models/vgg_face2_finetune_softmax/ \
--pretrained_model=saved_models/vgg_face2_finetune_softmax/modelname \
--data_dir=dataset/HW_1_Face/mtcnn_160/train \
--test_dir=dataset/HW_1_Face/mtcnn_160/test \
--max_nrof_epochs=100 \
--gpu_memory_fraction=0.8 \
--batch_size=300 \
--epoch_size=20 \
--embedding_size=512 \
--learning_rate=0.01 \
--learning_rate_decay_epochs=20 \
--learning_rate_decay_factor=0.8 \
--save_every_n_epochs=10 \
--validate_every_n_epochs=1 \
--keep_probability=0.5 \
--weight_decay=0.001 \
--center_loss_factor=2.0 \
--center_loss_alfa=0.25

Remember to change --pretrained_model to the one last step generates.

This will lead to about top1 acc of 0.82 on test set.

Retrieve via cosine distance Use classifier_cos.py to apply retrieve functions.
```
python classifier_cos.py \
dataset/HW_1_Face/mtcnn_160/train \
dataset/HW_1_Face/mtcnn_160/test \
saved_models/model_path_it_should_be \
--batch_size=256 \
--image_size=160
```
Remember to change the 3rd argument to trained model path of last step or the one you have downloaded.

This will generate predictions_cos.txt and at about top1 acc of 0.98.

Comments

If you have any problems or something went wrong when training, feel free to contact me. Because this instruction and the code is summarized in a hurry and may contain something inaccurate.

Instructions why we use test set for validation:

There is no validation set in the provided data. And we only use validation for intermediate step result generation, not for hyper-parameters modulation. We adjust h-paras according to the loss.

The accuracy achieve 100% on training set very soon because of the small size of training set. So the primary reference during training is the loss.

MeteorsHub / HW_face