app/Hackathon_setup/face_recognition.py

import numpy as np
import cv2
from matplotlib import pyplot as plt
import torch
# In the below line,remove '.' while working on your local system. However Make sure that '.' is present before face_recognition_model while uploading to the server, Do not remove it.
from .face_recognition_model import *
from PIL import Image
import base64
import io
import os
import joblib
import pickle
# Add more imports if required



###########################################################################################################################################
#         Caution: Don't change any of the filenames, function names and definitions                                                      #
#        Always use the current_path + file_name for refering any files, without it we cannot access files on the server                  # 
###########################################################################################################################################

# Current_path stores absolute path of the file from where it runs. 
current_path = os.path.dirname(os.path.abspath(__file__))
# --- GLOBAL SETUP: Must match your training transforms ---
# Define the transformation pipeline for inference
trnscm = transforms.Compose([
    transforms.Grayscale(num_output_channels=1),
    transforms.Resize((100, 100)),
    transforms.ToTensor()
])
CLASS_NAMES = ['Person0', 'Person1', 'Person2', 'Person3', 'Person4'] # ADJUST THIS!

# --- Model Filenames ---
SIAMESE_MODEL_PATH = current_path + '/siamese_model.t7'
KNN_CLASSIFIER_PATH = current_path + '/decision_tree_model.sav'
SCALER_PATH = current_path + '/face_recognition_scaler.sav'

#1) The below function is used to detect faces in the given image.
#2) It returns only one image which has maximum area out of all the detected faces in the photo.
#3) If no face is detected,then it returns zero(0).

def detected_face(image):
    eye_haar = current_path + '/haarcascade_eye.xml'
    face_haar = current_path + '/haarcascade_frontalface_default.xml'
    face_cascade = cv2.CascadeClassifier(face_haar)
    eye_cascade = cv2.CascadeClassifier(eye_haar)
    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray, 1.3, 5)
    face_areas=[]
    images = []
    required_image=0
    for i, (x,y,w,h) in enumerate(faces):
        face_cropped = gray[y:y+h, x:x+w]
        face_areas.append(w*h)
        images.append(face_cropped)
        required_image = images[np.argmax(face_areas)]
        required_image = Image.fromarray(required_image)
    return required_image


#1) Images captured from mobile is passed as parameter to the below function in the API call. It returns the similarity measure between given images.
#2) The image is passed to the function in base64 encoding, Code for decoding the image is provided within the function.
#3) Define an object to your siamese network here in the function and load the weight from the trained network, set it in evaluation mode.
#4) Get the features for both the faces from the network and return the similarity measure, Euclidean,cosine etc can be it. But choose the Relevant measure.
#5) For loading your model use the current_path+'your model file name', anyhow detailed example is given in comments to the function 
#Caution: Don't change the definition or function name; for loading the model use the current_path for path example is given in comments to the function
def get_similarity(img1, img2):
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
    det_img1 = detected_face(img1)
    det_img2 = detected_face(img2)
    if(det_img1 == 0 or det_img2 == 0):
        det_img1 = Image.fromarray(cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY))
        det_img2 = Image.fromarray(cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY))
    face1 = trnscm(det_img1).unsqueeze(0)
    face2 = trnscm(det_img2).unsqueeze(0)
    ##########################################################################################
    ##Example for loading a model using weight state dictionary:                            ##
    ## feature_net = light_cnn() #Example Network                                           ##
    ## model = torch.load(current_path + '/siamese_model.t7', map_location=device)          ##
    ## feature_net.load_state_dict(model['net_dict'])                                       ##
    ##                                                                                      ##
    ##current_path + '/<network_definition>' is path of the saved model if present in       ##
    ##the same path as this file, we recommend to put in the same directory                 ##
    ##########################################################################################
    ##########################################################################################
    
    # YOUR CODE HERE, load the model
    
    # YOUR CODE HERE, return similarity measure using your model
    # 1. Initialize and Load Siamese Network
    try:
        # Assuming your Siamese Network class is named 'SiameseNetwork'
        siamese_net = SiameseNetwork().to(device)
        siamese_net.load_state_dict(torch.load(SIAMESE_MODEL_PATH, map_location=device))
        siamese_net.eval()
    except Exception as e:
        print(f"Error loading Siamese Model get_similarity: {e}")
        return -1 # Return error code

    # 2. Get Features (Embeddings)
    with torch.no_grad():
        # Get the feature vector from one tower/forward_once method
        # Ensure your SiameseNetwork class has a forward_once or get_embedding method
        embed1 = siamese_net.forward_once(face1).cpu().numpy()
        embed2 = siamese_net.forward_once(face2).cpu().numpy()
        
    # 3. Calculate Similarity Measure
    # The Euclidean distance is the fundamental metric used by the Triplet/Contrastive loss.
    # We return the NEGATIVE Euclidean distance or COSINE similarity, as *higher* value usually means *more* similar.
    
    # Option A: Euclidean Distance (Lower is better) -> return NEGATIVE distance for API expectation
    # distance = euclidean_distances(embed1, embed2)[0][0]
    # similarity = -distance 
    
    # Option B: Cosine Similarity (Higher is better) -> Recommended
    similarity = cosine_similarity(embed1, embed2)[0][0]
    
    return float(similarity)
    
#1) Image captured from mobile is passed as parameter to this function in the API call, It returns the face class in the string form ex: "Person1"
#2) The image is passed to the function in base64 encoding, Code to decode the image provided within the function
#3) Define an object to your network here in the function and load the weight from the trained network, set it in evaluation mode
#4) Perform necessary transformations to the input(detected face using the above function).
#5) Along with the siamese, you need the classifier as well, which is to be finetuned with the faces that you are training
##Caution: Don't change the definition or function name; for loading the model use the current_path for path example is given in comments to the function
def get_face_class(img1):
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
    det_img1 = detected_face(img1)
    if(det_img1 == 0):
        det_img1 = Image.fromarray(cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY))
    ##YOUR CODE HERE, return face class here
    ##Hint: you need a classifier finetuned for your classes, it takes o/p of siamese as i/p to it
    ##Better Hint: Siamese experiment is covered in one of the labs
    face1_tensor = trnscm(det_img1).unsqueeze(0).to(device)

    # 1. Load Siamese Network (Feature Extractor)
    try:
        siamese_net = SiameseNetwork().to(device)
        siamese_net.load_state_dict(torch.load(SIAMESE_MODEL_PATH, map_location=device))
        siamese_net.eval()
    except Exception as e:
        return f"Error loading Siamese Model get_face_class: {e}"

    # 2. Extract Embedding
    with torch.no_grad():
        embedding_np = siamese_net.forward_once(face1_tensor).cpu().numpy()
    
    # 3. Load Sklearn Scaler and Classifier (Joblib)
    try:
        knn_classifier = joblib.load(KNN_CLASSIFIER_PATH)
        scaler = joblib.load(SCALER_PATH)
    except Exception as e:
        return f"Error loading Sklearn models: {e}"
    
    # 4. Preprocess Embedding and Predict
    # The embedding must be reshaped to (1, N_features) for the scaler
    embedding_scaled = scaler.transform(embedding_np.reshape(1, -1))
    
    # Perform prediction (returns a NumPy array with the predicted label index)
    predicted_label_index = knn_classifier.predict(embedding_scaled)[0]
    
    # 5. Map index to Class Name
    if predicted_label_index < len(CLASS_NAMES):
        predicted_class_name = CLASS_NAMES[predicted_label_index]
    else:
        predicted_class_name = "UNKNOWN_CLASS"

    return predicted_class_name