Insight Face in TensorFlow

Tasks

• mxnet dataset to tfrecords
• backbone network architectures [vgg16, vgg19, resnet]
• backbone network architectures [resnet-se, resnext]
• LResNet50E-IR
• LResNet100E-IR
• Additive Angular Margin Loss
• CosineFace Loss
• train network code
• add validate during training
• multi-gpu training
• combine losses contributed by RogerLo.
• evaluate code

Training Tips(Continual updates)

• If you can't use large batch size(>128), you should use small learning rate
• If you can't use large batch size(>128), you can try batch renormalization(file L_Resnet_E_IR_RBN.py)
• If use multiple gpus, you should keep at least 16 images each gpu.
• Try Group Normalization, you can use the code L_Resnet_E_IR_GBN.py
• Using the current model, and the lr schedule in train_nets.py, you can get the results as model c
• The bug about model size is 1.6G have fixed based on issues #9. If you want to get a small model, you should use L_Resnet_E_IR_fix_issues9.py
• multi-gpu training code's bug have fixed. If you want to use the correct version, you should use train_nets_mgpu_new.py

Training models (Continual updates)

model A

model name	depth	normalization layer	batch size	total_steps	download	password
model A	50	group normalization	16	1060k	model a	2q72

accuracy

dbname	accuracy
lfw	0.9897
cfp_ff	0.9876
cfp_fp	0.84357
age_db30	0.914

model B

model name	depth	normalization layer	batch size	total_steps	download	password
model B	50	batch normalization	16	1100k	model_b	h6ai

accuracy

dbname	accuracy
lfw	0.9933
cfp_ff	0.99357
cfp_fp	0.8766
age_db30	0.9342

model C

model name	depth	normalization layer	batch size	total_steps	download	password
model C	50	batch normalization	16	1950k	model_c	8mdi

accuracy

dbname	accuracy
lfw	0.9963
cfp_ff	0.99586
cfp_fp	0.9087
age_db30	0.96367

model D

model name	depth	normalization layer	batch size	total_steps	model_size	download	password
model D	50	batch normalization	136	710k	348.9MB	model_d	amdt

accuracy

dbname	accuracy
lfw	0.9968
cfp_ff	0.9973
cfp_fp	0.9271
age_db30	0.9725

Requirements

1. TensorFlow 1.4 1.6
2. TensorLayer 1.7
3. cuda8&cudnn6 or cuda9&cudnn7
4. Python3

Max Batch Size Test

Environment

GPU	cuda	cudnn	TensorFlow	TensorLayer	Maxnet	Gluon
Titan xp	9.0	7.0	1.6	1.7	1.1.0	1.1.0

Results

DL Tools	Max BatchSize(without bn and prelu)	Max BatchSize(with bn only)	Max BatchSize(with prelu only)	Max BatchSize(with bn and prelu)
TensorLayer	(8000, 9000)	(5000, 6000)	(3000, 4000)	(2000, 3000)
Mxnet	(40000, 50000)	(20000, 30000)	(20000, 30000)	(10000, 20000)
Gluon	(7000, 8000)	(3000, 4000)	no official method	None

(8000, 9000) : 8000 without OOM, 9000 OOM Error

Test Code

TensorLayer	Maxnet	Gluon
tensorlayer_batchsize_test.py	mxnet_batchsize_test.py	gluon_batchsize_test.py

pretrained model download link

• resnet_v1_50
• resnet_v1_101
• resnet_v1_152
• vgg16
• vgg19

References

1. InsightFace mxnet
2. InsightFace : Additive Angular Margin Loss for Deep Face Recognition
3. Group Normalization
4. tensorlayer_vgg16
5. tensorlayer_vgg19
6. tf_slim
7. Deep Residual Learning for Image Recognition
8. Very Deep Convolutional Networks For Large-Scale Image Recognition
9. Squeeze-and-Excitation Networks