face-recognition-supervised-by-center-loss

Deep Face Recognition ; pytorch ; center_loss ; triplet_loss;

Introduction

In this repository,we provide code to train deep face neural network using pytorch. the network was training supervised by center loss. we also provide triplet loss Method to train the network,but my experients indicate the result is not good comparing using center loss. so I recommend you use center loss.

the architecture I used is standard vgg16. you can replace it with other architecture like resnet50,101... these will have higher accurancy. As for why I use vgg16. maybe.....too young too sample.... if you use resnet50, the accurancy could improve two or three points I guess.

the training data I used is vggface2 which have 8631 identities and 3 million pictures. MS1M have larger pictures. Using MS1M the accurancy would have a higher accurancy comparing vggface2,maybe one point or two.

the test data I used is LFW.

My Result:

vgg16+vggface2: test accurancy: 0.967 auc: 0.99

(the code will drawing roc curve in checkpoints when test)

Environment：

python3, pytorch

the code can be run under both GPU and CPU

Data prepare

All face images should be aligned by MTCNN and cropped to 112x112

if you don't alighed,the accurancy would be very low

LFW download path(BaiduNetdisk):

https://pan.baidu.com/s/1eilswvW-qXy3XsHoTeHRzg

passward：iyd7

the LFW data above is alighed. vggface2 is too large.You can get it in https://github.com/deepinsight/insightface. this repo not only provide alighed datasets like vggface2 and mslm, but also alighed code in src/align. You can use the code in src/align to align your own datasets.

you must put the data in the directory "./datasets". And the format is as follows:

--datasets
   
   --vggface2
   
        --people name/id
        
         --111.jpg

that is say the images must be grouped by id or name.

Train

Using center loss (recommand)

paper: A Discriminative Feature Learning Approach for Deep Face Recognition

1.modify config file -------> training_certer.json

I write all the config parameters in this json file. the explain is as follows:

parameter	default value	explain
dataroot	"datasets/vggface2"	traing data dir
lfw_dir	"datasets/lfw_alighed"	test data dir
lfw_pairs_path	"datasets/lfw_pairs.txt"	triplet pairs of LFW
log_dir	"./checkpoints"	is loding pretrained checkpoint
resume	false	is loding pretrained checkpoint
start_epoch	0	start epoch index
epochs	50	epoch num
no_cuda	false	is not using gpu.false is use gpu.true is not use gpu
gpu_id	"0,1"	gpu index
num_workers	10	threads num of loding data
seed	0	random seed
val_interval	200	every $(val_interval) batchs test on test dataset
log_interval	10	every $(log_interval) batchs print training message contain loss
save_interval	500	every $(save_interval) batchs save checkpoint
batch_size	256	traing batch_size
test_batch_size	128	test batch_size
optimizer	"sgd"	optimizer in sgd/adam/adagrad
lr	0.1	learning rate
center_loss_weight	0.5	center_loss weight
alpha	0.5	center_loss learning rate
beta1	0.9	adam param1
lr_decay	1e-4	adam param2
wd	0.0	adam param3

python train_center.py

Using triplet loss

paper：FaceNet: A Unified Embedding for Face Recognition and Clustering

1.modify config file -------> training_triplet.json

there are almost the same with training_certer.json, but two more parameters:

parameter	default value	explain
n_triplets	1000000	triplet pairs num for training
margin	0.5	margin in paper FaceNet

2.run

python train_triplet.py

you can check training process using tensorboard. the specific way is writen in "checkpoints/readme.txt"

pretrained checkpoints

There is a checkpoint I trained using vgg16. You can use it as pretrained model.

download path(BaiduNetdisk): https://pan.baidu.com/s/1IPdmFy0bFfPt1xqV8S7eDg

passward：89sf

Contact

Email: [email protected] vx: 18811423186

llllllllllllllily / face-recognition-supervised-by-center-loss Goto Github PK