seungyonglee0802/face_recognition

Face embedding trainer

This repository contains the framework for training deep embeddings for face recognition. The trainer is intended for the face recognition exercise of the EE488B Deep Learning for Visual Understanding course. This is an adaptation of the speaker recognition model trainer.

20180467 EE488B Experiment Report.pdf

KSTAR Face Map

GOALs

1. Train CNN model that can make appropriate embedding vector(nOut = 512) for korean star's faces. Evaluated by EER(equal error rate).
2. Get the most similar the korean star's face with a face with a random person(not famous)
3. Draw "KSTAR-FaceMap" which put faces nearby when they are similar each other, and vice versa.

Dependencies

pip install -r requirements.txt

Training examples

• Pretrain:

$ python ./trainEmbedNet.py --model ThinResNet50_V2 --train_path data/train/VGGFace2 
--trainfunc amsoftmax --scale 30 --margin 0.1 --save_path exps/T50V2 
--max_epoch 60 --nPerClass 8631 --max_img_per_cls 200 --batch_size 200 --lr 0.001 
--scheduler cosineRestartlr --gpu 0

GPU ID must be specified using --gpu flag.

Use --mixedprec flag to enable mixed precision training. This is recommended for Tesla V100, GeForce RTX 20 series or later models.

• Fine-tuning:

$ python ./trainEmbedNet.py --model ThinResNet50_V2 --initial_model exps/T50V2/model000000050.model 
--trainfunc angproto --nPerClass 2 --save_path exps/transfer_T50V2 
--max_epoch 50 --test_interval 1 --batch_size 250 --lr 0.0005 --scheduler cosineRestartlr --gpu 0

• Evaluation:

$ python ./trainEmbedNet.py --model ThinResNet50_V2 --initial_model exps/transfer_T50V2/model000000020.model
--trainfunc angproto --gpu 0 
--eval --test_path data/test_shuffle --test_list data/test_blind.csv --output output.csv

Implemented loss functions

Softmax (softmax)
Triplet (triplet)

For softmax-based losses, nPerClass should be 1, and nClasses must be specified. For metric-based losses, nPerClass should be 2 or more.

Implemented models

ThinResNet50_V2

Adding new models and loss functions

You can add new models and loss functions to models and loss directories respectively. See the existing definitions for examples.

Data

The test list should contain labels and image pairs, one line per pair, as follows. 1 is a target and 0 is an imposter.

1,id10001/00001.jpg,id10001/00002.jpg
0,id10001/00003.jpg,id10002/00001.jpg

The folders in the training set should contain images for each identity (i.e. identity/image.jpg).

The input transformations can be changed in the code.

Inference

In order to save pairwise similarity scores to file, use --output flag.