MobileFaceNet with Arc-Face Loss

---

Requirments

numpy 1.15.1
keras 2.2.4
tensorflow-gpu 1.9.0
opencv-python 3.4.3.18

---

Training

python train_MobileFaceNet.py train_directory data_directory pairs_filename  <--optional_arguments>

Arguments

• Data Generator
• Model
• Training
• Eveluattion

Data Generator

• You can define your own data generator in src/data_generators.py

Argument	Description	type	Default
train_directory	Training dataset directory	str
valid_directory	Validation dataset directory	str
batch_size	Batch size of generator.	int	200
aug_freq	Frequency of data augmentation.	float	0.5
image_size	Image size same as model input size.	int	112
shuffle	Shuffle on end of epoch.	bool	True

Model

Argument	Description	type	Default
expansion_ratio	Expansion ratio of res_block.	int	6
embedding_dim	Embedding Dimension.	int	256
loss_scale	Scale parameter of arc-face loss.	int	64
loss_margin	Angular margin of arc-face loss.	float	0.5

Training

Argument	Description	type	Default
pretrained_model	Pre-trained model filename.	str	None
save_model_directory	Directory to save model	str	weights/
checkpoint_epochs	Save checkpoint every n epochs.	int	5
epochs	Max number of training epochs	int	300
valid_split_ratio	Split ratio of validation set	float	0.1
evaluate_epochs	Evaluate model every n epochs	int	5

Evaluation

Argument	Description	type	Default
data_directory	Evaluation dataset directory.	str
pairs_filename	Pairs file name	str
sample_type	Sample type of the task. 0 : balance pos/neg 1 : sample by person and img per person.	0 or 1	0
repeat_times	Repeat times of generation, this argument only be used when sample type is 0.	int	10
num_person	Number of person to sample, this argument only be used when sample type is 1.	int	10
num_sample	Number of sample per person, this argument only be used when sample type is 1.	int	20
far_target	Target FAR(False Accept Rate)	float	1e-2

GPU

Argument	Description	type	Default
gpu	Specify a GPU.	str	'1'

---

infer example

import cv2 
import matplotlib.pyplot as plt
from src.build_model import ArcFaceLossLayer, dummy_loss
from src.feature_extractor import FeatureExtractor

# load model
model_path = 'weights/weights.h5'
fe = FeatureExtractor(model_path, num_classes=24)

# load image and resize to model input size 
path = 'images.jpg'
img = cv2.imread(path)[:, :, ::-1]
img_resize = cv2.resize(img, (112, 112))

# infer 
emb = fe.infer(img_resize)
emb.shape
>> (1, 512)

Reference

ArcFace : https://arxiv.org/abs/1801.07698\ MobileFaceNet : https://arxiv.org/abs/1804.07573