app.py

import PIL.Image
import gradio as gr
import numpy as np
from celeb_indicies import *

import PIL
import os
import platform
import time
from huggingface_hub import hf_hub_download
import huggingface_hub
# import spaces #[uncomment to use ZeroGPU]
from diffusers import DiffusionPipeline
import torch
from torch import nn
import torchvision

# by default, dlib will be compiled locally when installed via pip, which takes so long it 
# causes huggingface to time out during the build process.
# To avoid this, check if we are running on a Linux system, and if so load a binary of dlib compiled for x86_64 Linux
if(platform.system() == 'Linux'):    
    os.system("pip install ./dlib-19.24.99-cp310-cp310-linux_x86_64.whl")
    os.system("pip install face-recognition")

start_time = time.time()
import face_detection
print(f"took {(time.time() - start_time) / 60} minutes to load face_recognition")


# # Function to display the uploaded image
# def process_image(image : PIL.Image.Image):
#     outputs = face_detection.getCroppedImages(image)
#     # do AI stuff here
#     return gr.Image(outputs[0])

model_repo_id = "CSSE416-final-project/faceRecogModel"
weight_file_id = "modelWeights100.bin"


# 1. Load the model from Hugging Face Hub
def load_model(repo_id):
    # Download the model weights from the repo
    weights_path = hf_hub_download(repo_id=model_repo_id, filename=weight_file_id)

    # Initialize the ResNet-18 architecture
    model = torchvision.models.resnet18(pretrained=True) # TODO: does it matter if this is set to true or false?
    num_ftrs = model.fc.in_features
    model.fc = nn.Linear(num_ftrs, 100)  # Adjust for your task (e.g., 128 classes)
    # TODO: check if this number^^ corresponds to the number of classes

    # Load the model weights
    state_dict = torch.load(weights_path, map_location=torch.device("cpu"))
    model.load_state_dict(state_dict)
    model.eval()  # Set the model to evaluation mode
    return model


# 2. Load model
model = load_model(model_repo_id)


# 3. Define how to transform image
transforms = torchvision.transforms.Compose(
[
    torchvision.transforms.ToTensor()
])


# 4. Preprocess and display the image
def process_image_str(groupImageFilePath: str):
    groupImage = PIL.Image.open(groupImageFilePath)
    locations, images = face_detection.getCroppedImages(groupImage)

    outputLabels = []
    for image in images:
        # Process the image
        intputTensor = transforms(image).unsqueeze(0) # unsqueeze? add batch dimension??

        # do AI stuff here
        with torch.no_grad():
            outputs_t = model(intputTensor)
            print(outputs_t)
            temp, pred_t = torch.max(outputs_t, dim=1)
            print(temp)
            outputLabels.append(celeb_list[pred_t.item()])

    #return gr.Image(image)
    print(outputLabels)
    return outputLabels.pop(0)


# 5. Create the Gradio interface
interface = gr.Interface(
    fn=process_image_str,  # Function to process the image
    inputs=gr.Image(type='filepath'),  # Upload input
    # outputs=gr.Image(),  # Display output
    outputs='text',
    allow_flagging='never',
    title="Celebrity Face Detector",
    description="Upload a picture of a celebrity or group of celebrities to identify them"
)

# 6. Launch the app
if __name__ == "__main__":
    interface.launch()