joonb14/jhface

Face Recognition training and testing framework with tensorflow 2.0 based on the well implemented arcface-tf2. Changes are added to provide tensorflow lite conversion, and provide additional backbones, loss functions.

JHFace

This project is based on the well implemented arcface-tf2. The things that rise error when converting the model to tflite was changed.

Training the Model

For preparing data, follow the instructions provided in data-preparing Then checkout the TensorFlow ArcFace.ipynb.

Backbones w/ ImageNet pretrained weights:

NasNet case, because of this issue, we manually download the weight file and explicitly load it in models.py file. We tried to provide pretrained weights for the MobileNet and EfficientNet models. However the official Keras implementations of the MobileNetV3 has built in preprocessing layers inside the model. Also this accounts to EfficientNet as well. So for the EfficientNet and EfficientNetLite, we used Weights Transfer.ipynb for extracting pretrained weights. To use the pretrained weights for NasNet, EfficientNet, EfficientNetLite please download the weights using this link. Then unzip it inside the path/to/JHFace/weights/ directory. All of the pretrained weights are provided by(or extracted from) tf.keras.applications

• MobileNet
• MobileNetV2
• InceptionResNetV2
• InceptionV3
• ResNet50
• ResNet50V2
• ResNet101V2
• NASNetLarge
• NASNetMobile
• Xception
• MobileNetV3Large
• MobileNetV3Small
• EfficientNetLite0 ~ Lite6
• EfficientNetB0 ~ B7

Backbones w/o ImageNet pretrained weights:

We implemented MnasNet models looking at the official code. If there's a bug, please tell us through the github issue page!

• MnasNetA1
• MnasNetB1
• MnasNetSmall

Loss Function

• ArcFace
• CosFace
• SphereFace

Configuration

in the TensorFlow ArcFace.ipynb, we provided simple configuration values. To change the model backbone, just change the backbone_type parameter. To change the loss function, just change the head_type parameter.

### MS1M dataset

batch_size = 128 # Initially 128
input_size = 112
embd_shape = 512
head_type = 'ArcHead' # 'ArcHead', 'CosHead', 'SphereHead'
# Backbones w/ pretrained weights:
#     MobileNet, MobileNetV2, InceptionResNetV2, InceptionV3, ResNet50, ResNet50V2, ResNet101V2, NASNetLarge, NASNetMobile, Xception, MobileNetV3Large, MobileNetV3Small, EfficientNetLite0~6, EfficientNetB0~7
# Backbones w/o pretrained weights:
#      MnasNetA1, MnasNetB1, MnasNetSmall 
backbone_type = 'EfficientNetLite0' 
w_decay=5e-4
num_classes = 85742 
dataset_len = 5822653 
if head_type == 'SphereHead':
    base_lr = 0.01 
    margin = 1.35
    logist_scale = 30.0 
elif head_type == 'CosHead':
    base_lr = 0.01 
    margin=0.35
    logist_scale=64
elif head_type == 'ArcHead':
    base_lr = 0.01 
    margin=0.5
    logist_scale=64
else:
    base_lr = 0.01 # initially 0.01
epochs = 20
save_steps = 1000
train_size = int(0.8 * dataset_len)
print(train_size)
steps_per_epoch = train_size // batch_size
print(steps_per_epoch)
val_size = dataset_len - train_size
print(val_size)
validation_steps = val_size // batch_size
print(validation_steps)
steps = 1
is_ccrop=False
binary_img=True
is_Adam = False

Converting the Model to TensorFlow Lite

checkout the TFLite conversion.ipynb. Int8 quantization is supported, we checked with MobileNetV2 and EfficientNet-lite0.

For Face Verification

We downloaded the testing dataset from here. With the data use the verification.ipynb for verification test.

For Face Identificaiton

checkout the Face Identification with Centroid Vector.ipynb.

jhface's Issues

Dataset format

Hello,
Thanks for this good work.
Does one have to convert the dataset to tfrecord.
Or will it work by just pointing to directories.
Thanks

Request: TFlite model

Can you provide us with the pre-trained tflite model ?