deepface/deepface/models/facial_recognition/ArcFace.py

import os
import gdown
from deepface.commons import package_utils, folder_utils
from deepface.models.FacialRecognition import FacialRecognition

from deepface.commons.logger import Logger

logger = Logger()

# pylint: disable=unsubscriptable-object

# --------------------------------
# dependency configuration

tf_version = package_utils.get_tf_major_version()

if tf_version == 1:
    from keras.models import Model
    from keras.engine import training
    from keras.layers import (
        ZeroPadding2D,
        Input,
        Conv2D,
        BatchNormalization,
        PReLU,
        Add,
        Dropout,
        Flatten,
        Dense,
    )
else:
    from tensorflow.keras.models import Model
    from tensorflow.python.keras.engine import training
    from tensorflow.keras.layers import (
        ZeroPadding2D,
        Input,
        Conv2D,
        BatchNormalization,
        PReLU,
        Add,
        Dropout,
        Flatten,
        Dense,
    )

# pylint: disable=too-few-public-methods
class ArcFaceClient(FacialRecognition):
    """
    ArcFace model class
    """

    def __init__(self):
        self.model = load_model()
        self.model_name = "ArcFace"
        self.input_shape = (112, 112)
        self.output_shape = 512


def load_model(
    url="https://github.com/serengil/deepface_models/releases/download/v1.0/arcface_weights.h5",
) -> Model:
    """
    Construct ArcFace model, download its weights and load
    Returns:
        model (Model)
    """
    base_model = ResNet34()
    inputs = base_model.inputs[0]
    arcface_model = base_model.outputs[0]
    arcface_model = BatchNormalization(momentum=0.9, epsilon=2e-5)(arcface_model)
    arcface_model = Dropout(0.4)(arcface_model)
    arcface_model = Flatten()(arcface_model)
    arcface_model = Dense(512, activation=None, use_bias=True, kernel_initializer="glorot_normal")(
        arcface_model
    )
    embedding = BatchNormalization(momentum=0.9, epsilon=2e-5, name="embedding", scale=True)(
        arcface_model
    )
    model = Model(inputs, embedding, name=base_model.name)

    # ---------------------------------------
    # check the availability of pre-trained weights

    home = folder_utils.get_deepface_home()

    file_name = "arcface_weights.h5"
    output = os.path.join(home, ".deepface/weights", file_name)

    if not os.path.isfile(output):
        logger.info(f"{file_name} will be downloaded to {output}")
        gdown.download(url, output, quiet=False)
    # ---------------------------------------

    model.load_weights(output)

    return model


def ResNet34() -> Model:
    """
    ResNet34 model
    Returns:
        model (Model)
    """
    img_input = Input(shape=(112, 112, 3))

    x = ZeroPadding2D(padding=1, name="conv1_pad")(img_input)
    x = Conv2D(
        64, 3, strides=1, use_bias=False, kernel_initializer="glorot_normal", name="conv1_conv"
    )(x)
    x = BatchNormalization(axis=3, epsilon=2e-5, momentum=0.9, name="conv1_bn")(x)
    x = PReLU(shared_axes=[1, 2], name="conv1_prelu")(x)
    x = stack_fn(x)

    model = training.Model(img_input, x, name="ResNet34")

    return model


def block1(x, filters, kernel_size=3, stride=1, conv_shortcut=True, name=None):
    bn_axis = 3

    if conv_shortcut:
        shortcut = Conv2D(
            filters,
            1,
            strides=stride,
            use_bias=False,
            kernel_initializer="glorot_normal",
            name=name + "_0_conv",
        )(x)
        shortcut = BatchNormalization(
            axis=bn_axis, epsilon=2e-5, momentum=0.9, name=name + "_0_bn"
        )(shortcut)
    else:
        shortcut = x

    x = BatchNormalization(axis=bn_axis, epsilon=2e-5, momentum=0.9, name=name + "_1_bn")(x)
    x = ZeroPadding2D(padding=1, name=name + "_1_pad")(x)
    x = Conv2D(
        filters,
        3,
        strides=1,
        kernel_initializer="glorot_normal",
        use_bias=False,
        name=name + "_1_conv",
    )(x)
    x = BatchNormalization(axis=bn_axis, epsilon=2e-5, momentum=0.9, name=name + "_2_bn")(x)
    x = PReLU(shared_axes=[1, 2], name=name + "_1_prelu")(x)

    x = ZeroPadding2D(padding=1, name=name + "_2_pad")(x)
    x = Conv2D(
        filters,
        kernel_size,
        strides=stride,
        kernel_initializer="glorot_normal",
        use_bias=False,
        name=name + "_2_conv",
    )(x)
    x = BatchNormalization(axis=bn_axis, epsilon=2e-5, momentum=0.9, name=name + "_3_bn")(x)

    x = Add(name=name + "_add")([shortcut, x])
    return x


def stack1(x, filters, blocks, stride1=2, name=None):
    x = block1(x, filters, stride=stride1, name=name + "_block1")
    for i in range(2, blocks + 1):
        x = block1(x, filters, conv_shortcut=False, name=name + "_block" + str(i))
    return x


def stack_fn(x):
    x = stack1(x, 64, 3, name="conv2")
    x = stack1(x, 128, 4, name="conv3")
    x = stack1(x, 256, 6, name="conv4")
    return stack1(x, 512, 3, name="conv5")