app/Hackathon_setup/exp_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 .exp_recognition_model import *
from PIL import Image
import base64
import io
import os
import torch
import torch.nn as nn
from torchvision import models
import torchvision.transforms as transforms
import torch.nn.functional as F

## 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__))
classes = {0: 'ANGER', 1: 'DISGUST', 2: 'FEAR', 3: 'HAPPINESS', 4: 'NEUTRAL', 5: 'SADNESS', 6: 'SURPRISE'}


#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 Expression detected by your network.
#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 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), this should return the Expression in string form ex: "Anger"
#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_expression(img):
#     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
#
#     # Recreate the same model architecture
#     num_classes = 7  # 👈 change this to match your training setup
#
#     model = models.resnet18(weights=None)
#     model.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
#     # no pretrained weights now
#     model.fc = nn.Sequential(
#         nn.Linear(model.fc.in_features, 256),
#         nn.ReLU(inplace=True),
#         nn.Linear(256, num_classes)
#     )
#
#     model = model.to(device)
#
#     # Create the optimizer (same as training)
#     optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
#
#     # Load the checkpoint
#     BASE_DIR = os.path.dirname(os.path.abspath(__file__))
#     ckpt_path = os.path.join(BASE_DIR, "expression_model.t7")
#     checkpoint = torch.load(ckpt_path, map_location=device)
#
#     # Restore weights and optimizer
#     model.load_state_dict(checkpoint['model_state_dict'])
#     optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
#
#     # Put the model in evaluation mode
#     model.eval()
#
#     ##########################################################################################
#     ##Example for loading a model using weight state dictionary:                            ##
#     ## face_det_net = facExpRec() #Example Network                                          ##
#     ## model = torch.load(current_path + '/exp_recognition_net.t7', map_location=device)    ##
#     ## face_det_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                 ##
#     ##########################################################################################
#     ##########################################################################################
#
#     transform = transforms.Compose([
#         transforms.Grayscale(num_output_channels=1),
#         transforms.Resize(256),
#         transforms.CenterCrop(224),
#         transforms.ToTensor(),
#         transforms.Normalize(mean=[0.5], std=[0.5])
#     ])
#
#     face = detected_face(img)
#     if face==0:
#         face = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY))
#
#     face = transform(face).unsqueeze(0).to(device)
#     # YOUR CODE HERE, return expression using your model
#     with torch.no_grad():
#         outputs = model(face)
#         probs = F.softmax(outputs, dim=1)
#         predicted_class = probs.argmax(dim=1).item()
#     return predicted_class


def get_expression(img):
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    num_classes = 7  # update as per your dataset

    # Recreate exact same architecture as training
    model = models.resnet18(weights=models.ResNet18_Weights.IMAGENET1K_V1)

    # Convert first conv layer to accept 1 channel (grayscale)
    pretrained_conv = model.conv1.weight
    model.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
    with torch.no_grad():
        model.conv1.weight = nn.Parameter(pretrained_conv.mean(dim=1, keepdim=True))

    # Fully connected head (same as training)
    model.fc = nn.Sequential(
        nn.Linear(model.fc.in_features, 256),
        nn.ReLU(),
        nn.Dropout(0.5),
        nn.Linear(256, num_classes)
    )

    model = model.to(device)

    # Load checkpoint
    BASE_DIR = os.path.dirname(os.path.abspath(__file__))
    ckpt_path = os.path.join(BASE_DIR, "expression_model.t7")
    checkpoint = torch.load(ckpt_path, map_location=device)

    # Restore weights (no need for optimizer if inference-only)
    model.load_state_dict(checkpoint['model_state_dict'])
    model.eval()

    # Preprocessing pipeline
    transform = transforms.Compose([
        transforms.Grayscale(num_output_channels=1),
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.5], std=[0.5])
    ])

    face = detected_face(img)
    if face == 0:
        face = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY))

    face = transform(face).unsqueeze(0).to(device)

    with torch.no_grad():
        outputs = model(face)
        probs = F.softmax(outputs, dim=1)
        predicted_class = probs.argmax(dim=1).item()

    return classes[predicted_class]