FaceRecognition/ModelTrainer.py

import tensorflow as tf

import numpy as np
from tensorflow_hub import KerasLayer
import os

from keras.layers import Dense,Dropout,Input,BatchNormalization,Lambda
from keras.models import Model
from keras.optimizers import Adam
from keras import Sequential
from keras.callbacks import EarlyStopping,ModelCheckpoint
from keras.layers.experimental.preprocessing import RandomRotation,RandomFlip,RandomCrop,PreprocessingLayer
from tensorflow.math import l2_normalize
from sys import argv

User=argv[1]

NegativeDatasetPath="./FaceRecognition/ExtactedFaces/Negative/"
TrainUsersDatasetPath="./FaceRecognition/ExtactedFaces/Train/"
TestUsersDatasetPath="./FaceRecognition/ExtactedFaces/Test/"
DatasetPath="./FaceRecognition/ExtactedFaces/Dataset/"
def DatasetPaths(UserDatapath,state,NegativeDatasetPath=None,DatasetPath=None):
    User=os.listdir(UserDatapath)
    UserFiles=[]
    UserLabels=[]
    for folder in User:
        for file in os.listdir(UserDatapath+folder):
            UserFiles.append(UserDatapath+folder+'/'+file)
            UserLabels.append(folder)
    
    if state==True:
        
        Negativefiles=[]
        NegativeLabels=[]
        DatasetPathfiles=[]
        for file in os.listdir(NegativeDatasetPath):
            Negativefiles.append(NegativeDatasetPath+file)
            NegativeLabels.append(file.split(",")[0])
            
            
        for file in os.listdir(DatasetPath):
            DatasetPathfiles.append(DatasetPath+file)
            
            
        return np.array(Negativefiles),np.array(NegativeLabels),np.array(UserFiles),np.array(UserLabels),np.array(DatasetPathfiles)
    return np.array(UserFiles),np.array(UserLabels)


Negativefiles,NegativeLabels,UserFiles,UserLabels,DatasetPathfiles=DatasetPaths(TrainUsersDatasetPath,True,NegativeDatasetPath,DatasetPath)
Negativefiles,NegativeLabels,UserFiles,UserLabels,DatasetPathfiles
TrainClasses=np.unique(UserLabels)
TrainClassesCount=len(TrainClasses)

TestUserFiles,TestUserLabels=DatasetPaths(UserDatapath=TestUsersDatasetPath,state=False)
TestClasses=np.unique(TestUserLabels)
TestClassesCount=len(TestClasses)

mask=np.zeros(shape=(224,224,3))
mask[:,:,0]=200
mask[:,:,1]=100
mask[:,:,2]=200
mask=tf.cast(mask/255,tf.float32)
FliPer=RandomFlip(mode="horizontal",)
Rotater=RandomRotation([-0.135,0.135])
def PreProcessInput(Image,num):
    if num ==0:
        Image=FliPer(Image)
    elif num==1:
        Image= 0.75*Image+0.25*mask
    if num<=2:
        return Rotater(Image)
    
    else:
        return Image
    
@tf.function
def load_image(Anchor,Positive,Nagative,State):

    Anchor=tf.io.read_file(Anchor)
    Anchor=tf.image.decode_jpeg(Anchor)
    Anchor = tf.cast(Anchor, tf.float32)
    Anchor = tf.image.resize(Anchor, [224,224], method = tf.image.ResizeMethod.NEAREST_NEIGHBOR)
    ranA=tf.random.uniform(shape=[1],minval=0,maxval=6,dtype=tf.int32)
    
    Positive=tf.io.read_file(Positive)
    Positive=tf.image.decode_jpeg(Positive)
    Positive = tf.cast(Positive, tf.float32)
    Positive = tf.image.resize(Positive, [224,224], method = tf.image.ResizeMethod.NEAREST_NEIGHBOR)
    ranB=tf.random.uniform(shape=[1],minval=0,maxval=6,dtype=tf.int32)
    
    Negative=tf.io.read_file(Nagative)
    Negative=tf.image.decode_jpeg(Negative)
    Negative = tf.cast(Negative, tf.float32)
    Negative = tf.image.resize(Negative, [224,224], method = tf.image.ResizeMethod.NEAREST_NEIGHBOR)
    ranN=tf.random.uniform(shape=[1],minval=0,maxval=6,dtype=tf.int32)
    if State:
        Anchor=PreProcessInput(Anchor/255,ranA)
        Positive=PreProcessInput(Positive/255,ranB)
        Negative=PreProcessInput(Negative/255,ranN)
    else:
        Anchor=Anchor/255
        Positive=Positive/255
        Negative=Negative/255
    
    return (Anchor,Positive,Negative)


def DatasetTripletsGenerator(State):
    #     Negativefiles=Negativefiles
#     NegativeLabels=NegativeLabels
#     DatasetPathfiles=DatasetPathfiles
    if State:
        UsersImagesPath=UserFiles
        UsersImagesLabel=UserLabels
        ClassesCount=TrainClassesCount
        Classes=TrainClasses
        
    else:
        ImagesName=TestUserFiles
        ImagesLabel=TestUserLabels
        ClassesCount=TestClassesCount
        Classes=TestClasses
        
        
   
    for i in range(ClassesCount):
        class_=Classes[i]
        files=UsersImagesPath[UsersImagesLabel==class_]
        
        files_num=len(files)
        for index in range(files_num-1):
            for j in range(index+1,files_num):
                ancore=files[index]
                positive=files[j]
                random=np.random.randint(0,high=10)
                negative=None
                if random<=3:
                    negative=Negativefiles[NegativeLabels==class_]
                    if type(negative)==list:
                        negative=np.random.choice(negative)
                elif random<=7:
                    negative=UsersImagesPath[UsersImagesLabel != class_]
                    if type(negative)==list:
                        negative=np.random.choice(negative)
                   
                elif random<=10:
                    negative=DatasetPathfiles
                    if type(negative)==list:
                        negative=np.random.choice(negative)
                
                if type(negative)!=str:
                    negative=np.random.choice(DatasetPathfiles)
                
                    
                
                
                yield ancore,positive,negative,State


@tf.function
def EmbeddingImageLoader(Anchor,Label):

    Anchor=tf.io.read_file(Anchor)
    Anchor=tf.image.decode_jpeg(Anchor)
    Anchor = tf.cast(Anchor, tf.float32)
    Anchor = tf.image.resize(Anchor, [224,224], method = tf.image.ResizeMethod.NEAREST_NEIGHBOR)
    
    Anchor=Anchor/255
       
    
    return (Anchor,Label)


TrainData=tf.data.Dataset.from_generator(DatasetTripletsGenerator,args=[True],output_types=(tf.string,tf.string,tf.string,tf.bool),output_shapes=((),(),(),()),name="DataLoaderPipeline")
TrainData=TrainData.map(load_image)
TrainData=TrainData.batch(2).shuffle(buffer_size=10)

TestData=tf.data.Dataset.from_generator(DatasetTripletsGenerator,args=[False],output_types=(tf.string,tf.string,tf.string,tf.bool),output_shapes=((),(),(),()),name="DataLoaderPipeline")
TestData=TestData.map(load_image).batch(2)

EmbeddingData=tf.data.Dataset.from_tensor_slices((list(UserFiles),list(UserLabels))).map(EmbeddingImageLoader).batch(1)


class DistanceLayer(tf.keras.layers.Layer):
    def __init__(self):
        super().__init__()
    def call(self,anchor,positive,negative):
        dis_ap=tf.reduce_sum(tf.square(anchor - positive), 1)  ## distance between anchor and positive 
        dis_an=tf.reduce_sum(tf.square(anchor - negative), 1)   ## distance between anchor and negative
        return  dis_ap , dis_an
 
def GetEncoder():
#     /drive/MyDrive/Model/
    if os.path.isdir("./FaceRecognition/FaceModel/keras/"):
        return tf.keras.models.load_model("./FaceRecognition/FaceModel/")
    else:
        pretrained_model = KerasLayer("./prtrained/archive/",trainable=False) ##pretraind Model

        encode_model = Sequential([
            pretrained_model,
            Dropout(0.2),
            Dense(512, activation='relu'),
            BatchNormalization(),
            Dense(128, activation="relu"),
            Lambda(lambda x:l2_normalize(x))
        ], name="Encoder")
        return encode_model
def SiameseNetwork(inputshape=(224,224,3)):
    An_input=Input(shape=inputshape)
    
    Po_input=Input(shape=inputshape)
    
    Ne_input=Input(shape=inputshape)
    
    encoder=GetEncoder()
    
    An_embeding=encoder(An_input)
    Po_embeding=encoder(Po_input)
    Ne_embeding=encoder(Ne_input)
    
    
    distanc=DistanceLayer()(An_embeding,Po_embeding,Ne_embeding) #return distance between (A and B) and (A and N)
    
    return Model(inputs=[An_input,Po_input,Ne_input],outputs=distanc)




siames_net=SiameseNetwork()
class SiamesModel(Model):
    def __init__(self,siames_net,DesiredDistance):
        super(SiamesModel, self).__init__()
        
        self.Model=siames_net
        self.DesiredDistance=DesiredDistance
        self.LossTracker=tf.keras.metrics.Mean(name="Loss")
        
        self.VALTracker=tf.keras.metrics.Mean(name="VAL")
        
        self.PmeanTracker=tf.keras.metrics.Mean(name="P_mean")
        
        self.PmaxTracker=tf.keras.metrics.Mean(name="P_max")
        
        self.PstdTracker=tf.keras.metrics.Mean(name="P_std")
        
        self.FARTracker=tf.keras.metrics.Mean(name="FAR")
        
        self.N_meanTracker=tf.keras.metrics.Mean(name="N_mean")
        
        self.NstdTracker=tf.keras.metrics.Mean(name="N_std")
        self.NminTracker=tf.keras.metrics.Mean(name="N_min")
        
    def call(self,data):
        return self.Model(data)
    
    def train_step(self,data):
        with tf.GradientTape() as Tape:
            AP_distanc,AN_distance=self.Model(data)
            loss=self.TripLoss(AP_distanc,AN_distance)
            gradients=Tape.gradient(loss,self.Model.trainable_weights)
            self.optimizer.apply_gradients(zip(gradients, self.Model.trainable_weights))
        self.DistanceEval(AP_distanc,AN_distance)
        self.LossTracker.update_state(loss)
        return {"VAL":self.VALTracker.result(),
                "P_mean":self.PmeanTracker.result(),
                "P_max":self.PmaxTracker.result(),
                "P_std":self.PstdTracker.result(),
                "FAR":self.FARTracker.result(),
                "N_mean":self.N_meanTracker.result(),
                "N_min":self.NminTracker.result(),
                "N_std":self.NstdTracker.result(),
                "Loss":self.LossTracker.result()}
    
    
    def test_step(self, data):
        AP_distanc,AN_distance=self.Model(data)
        loss=self.TripLoss(AP_distanc,AN_distance)
        self.LossTracker.update_state(loss)
        self.DistanceEval(AP_distanc,AN_distance)
        return {"VAL":self.VALTracker.result(),
                "P_mean":self.PmeanTracker.result(),
                "P_max":self.PmaxTracker.result(),
                "P_std":self.PstdTracker.result(),
                "FAR":self.FARTracker.result(),
                "N_mean":self.N_meanTracker.result(),
                "N_min":self.NminTracker.result(),
                "N_std":self.NstdTracker.result(),
                "Loss":self.LossTracker.result()}
    
    
    
    def TripLoss(self,ap_distance,an_distance):
        return tf.reduce_mean(tf.maximum(ap_distance-0.2*self.DesiredDistance,0)+tf.maximum(self.DesiredDistance-an_distance, 0.0))
    
    
    @property
    def metrics(self):
        return [self.LossTracker,self.VALTracker,self.PmaxTracker,self.PmeanTracker,self.PstdTracker,self.FARTracker,self.N_meanTracker,self.NminTracker,self.NstdTracker]
    
    def DistanceEval(self,P_distance,N_distance):

        P_pred,N_pred=self.TDEvaluation(P_distance,N_distance)
        PCDCount=tf.size(tf.where(P_pred))

        VAL=PCDCount/tf.size(P_pred)
        self.VALTracker.update_state(VAL)
        
        NCDcount=tf.size(tf.where(N_pred))
        FAR=1-(NCDcount/tf.size(P_pred))
        self.FARTracker.update_state(FAR)
        P_mean=tf.reduce_mean(P_distance)
        self.PmeanTracker.update_state(P_mean)
        N_mean=tf.reduce_mean(N_distance)
        self.N_meanTracker.update_state(N_mean)
        P_std=tf.math.reduce_std(P_distance)
        self.PstdTracker.update_state(P_std)
        N_std=tf.math.reduce_std(N_distance)
        self.NstdTracker.update_state(N_std)
        P_max=tf.reduce_max(P_distance)
        self.PmaxTracker.update_state(P_max)
        N_min=tf.reduce_min(N_distance)
        self.NminTracker.update_state(N_min)
    
    def TDEvaluation(self,P_distance,N_distance):
        return tf.cast(P_distance<=self.DesiredDistance,dtype=tf.int8),tf.cast(N_distance>self.DesiredDistance,dtype=tf.int8)
    
    
DesiredDistance=1
Optimizer= Adam(learning_rate=1e-4)
Siamesmodel=SiamesModel(siames_net,DesiredDistance)
Siamesmodel.compile(optimizer=Adam(1e-4),weighted_metrics=[])
Siamesmodel.fit(TrainData,validation_data=TestData,epochs=1,callbacks=[EarlyStopping(patience=3),ModelCheckpoint(f"./FaceRecognition/FaceModel/{User}/kerasModel")])



def EmbeddingMaker(DataPipline,Model):
    Embedding={}
    NamesTimer={}
    for  Image,Name in DataPipline:
        Name=str(Name[0].numpy())[2:-1]
        
        if Name[0] not in Embedding.keys():
            NamesTimer[Name]=1
            Embedding[Name]=tf.squeeze(Model(Image)).numpy()
            
        else:
            Embedding[Name]=Embedding[Name]+tf.squeeze(Model(Image)).numpy()
            NamesTimer[Name]=NamesTimer[Name]+1
    for Name in Embedding:
        Embedding[Name]=Embedding[Name]/NamesTimer[Name]
    return Embedding


Embedding=EmbeddingMaker(EmbeddingData,siames_net.layers[3])
EmbeddingLabel,EmbeddingNames=[[Embedding[Name] for Name in Embedding] , {Name:Index+1 for Index,Name in enumerate(Embedding) } ]



class LiteModel(tf.Module):
    def __init__(self,FaceModel,FacesEmbedding,name="FaceLiteModel"):
        self.FaceModel=FaceModel
        self.FacesEmdedding=FacesEmbedding
       
    @tf.function(input_signature=[tf.TensorSpec(shape=[None,224,224,3],dtype=tf.float32),tf.TensorSpec(shape=[],dtype=tf.float32)])
    def __call__(self,Image,Threshold):
        Embedding=self.FaceModel(Image)
        Distance=tf.cast(Threshold,tf.float32)
        Name=0
        for Index,StoredEmbedding in enumerate(self.FacesEmdedding):
            distance=tf.reduce_sum(tf.math.pow(Embedding-StoredEmbedding,2))
            if distance<Distance:
                Name=Index+1
                Distance=distance
        return Name,Distance
    
litemodel=LiteModel(siames_net.layers[3],FacesEmbedding=EmbeddingLabel)


converter=tf.lite.TFLiteConverter.from_concrete_functions([litemodel.__call__.get_concrete_function()],litemodel)
converter.optimizations=[tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types=[tf.float16]
tflitemodel=converter.convert()

with open(f"./FaceRecognition/FaceModel/{User}/FaceXModel.tflite","wb") as file:
    file.write(tflitemodel)