main.py

import gradio as gr
from facenet_pytorch import InceptionResnetV1
import torch.nn as nn
import torchvision.transforms as tf
import numpy as np
import torch
import faiss
import h5py
import tqdm
import os
import random
from PIL import Image
import matplotlib.cm as cm
import matplotlib as mpl

img_names = []
with open('list_eval_partition.txt', 'r') as f:
    for line in f:
        img_name, dtype = line.rstrip().split(' ')
        img_names.append(img_name)


# For a model pretrained on VGGFace2
print('Loading model weights ........')

class SiameseModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.backbone = InceptionResnetV1(pretrained='vggface2')
    def forward(self, x):
        x = self.backbone(x)
        x = torch.nn.functional.normalize(x, dim=1)
        return x
    
model = SiameseModel()
model.load_state_dict(torch.load('model.pt', map_location=torch.device('cpu')))
model.eval()


# Make FAISS index
print('Make index .............')
index = faiss.IndexFlatL2(512)

hf = h5py.File('face_vecs_full.h5', 'r')
for key in tqdm.tqdm(hf.keys()):
    vec = np.array(hf.get(key))
    index.add(vec)

hf.close()

print("Finished indexing")

# Function to search image
def image_search(image, k=5):
    
    transform = tf.Compose([
        tf.Resize((160, 160)),
        tf.ToTensor()
    ])

    query_img = transform(image)
    query_img = torch.unsqueeze(query_img, 0)

    model.eval()
    query_vec = model(query_img).detach().numpy()

    D, I = index.search(query_vec, k=k)

    retrieval_imgs = []
    
    FOLDER = 'img_align_celeba'
    for idx in I[0]:
        img_file_name = img_names[idx]
        path = os.path.join(FOLDER, img_file_name)

        image = Image.open(path)
        retrieval_imgs.append((image, ''))
        
    return retrieval_imgs

with gr.Blocks(theme=gr.themes.Monochrome()) as demo:
    gr.Markdown('''
    
    
    # Face Image Retrieval with Content-based Retrieval Image (CBIR) & Saliency Map
    --------
    
    
    ''')
    
    with gr.Row():
        with gr.Column():
            image = gr.Image(type='pil', scale=1)
            slider = gr.Slider(1, 10, value=5, step=1, label='Number of retrieval image')
            with gr.Row():
                btn = gr.Button('Search')
                clear_btn = gr.ClearButton()
        
        gallery = gr.Gallery(label='Retrieval Images', columns=[5], show_label=True, scale=2)
        
    img_dir = './img_align_celeba'    
    examples = random.choices(img_names, k=5)
    examples = [os.path.join(img_dir, ex) for ex in examples]
    examples = [Image.open(img) for img in examples]
                                  
    with gr.Row():
        gr.Examples(
            examples = examples,
            inputs = image
        )
            
            
    btn.click(image_search, 
              inputs= [image, slider], 
              outputs= [gallery])
    
    def clear_image():
        return None
    
    clear_btn.click(
        fn = clear_image,
        inputs = [],
        outputs = [image]
    )
    
if __name__ == "__main__":
    demo.launch(server_name = "0.0.0.0", server_port = 7860)