auroua / Insightface_tf

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

accuracy

model B

accuracy

model C

accuracy

model D

accuracy

Requirements

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

Max Batch Size Test

Environment

Results

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

Test Code

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