app.py

import torch
import torchvision.models as models
import torchvision.transforms as transforms
from PIL import Image
import os
import streamlit as st


@st.cache()
def load_model():
    print("Loading model...")
    # Load the pre-trained ResNet-50 model and set it to eval mode
    model = models.resnet50(pretrained=True)
    model.eval()

    # Define the transform to be applied to each input image
    transform = transforms.Compose(
        [
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ]
    )

    # Define the directory containing the input images
    input_dir = "lfw"

    # Create a dictionary to store the feature vectors
    features_dict = {}

    # Modify the ResNet model to return the output of the penultimate layer
    model = torch.nn.Sequential(*list(model.children())[:-1])

    # Loop over each subdirectory and image file in the input directory
    for root, dirs, files in os.walk(input_dir):
        for file in files:
            # Check if the file is a JPEG image
            if file.endswith(".jpg"):
                # Load the image
                image = Image.open(os.path.join(root, file))

                # Apply the transform to the image
                image = transform(image)

                # Reshape the image to add a batch dimension
                image = image.unsqueeze(0)

                # Extract the features from the model's penultimate layer
                with torch.no_grad():
                    features = model(image).squeeze()

                # Add the feature vector to the dictionary
                features_dict[os.path.join(root, file)] = features.numpy()

    return features_dict


@st.cache()
def create_nearest_neighbors_object(features_dict):
    print("Creating nearest neighbors object...")
    import numpy as np
    from sklearn.neighbors import NearestNeighbors

    # Create a list of the feature vectors
    features_list = list(features_dict.values())

    # Create a NumPy array of the feature vectors
    features_array = np.array(features_list)

    # Create a nearest neighbors object
    nn = NearestNeighbors(n_neighbors=11, metric="euclidean")

    # Fit the nearest neighbors object to the feature vectors
    nn.fit(features_array)

    return nn


# Create a get nearest neighbors function
def get_nearest_neighbors(image_path):
    # Define the query image
    query_image = image_path
    print(query_image)

    # Loop through the dictionary to find the 10 nearest neighbors to the query image
    for key, value in features_dict.items():
        if key == query_image:
            query_features = value
            query_features = query_features.reshape(1, -1)
            distances, indices = nn.kneighbors(query_features)
            indices = indices[0]
            distances = distances[0]
            for i in range(1, 11):
                image = Image.open(list(features_dict.keys())[indices[i]])
                st.image(
                    image,
                    caption="Distance: " + str(distances[i]),
                    use_column_width=True,
                )

# SteamLit App

allow_output_mutation = True
features_dict = load_model()
nn = create_nearest_neighbors_object(features_dict)

# Title
st.title("Similarity Search")

# Subtitle
st.write("This app finds the 10 most similar images to a query image.")

query_image = st.selectbox("Or select an image from the list", os.listdir("lfw"))
# Display the query image from dir ./lfw/query_image/query_image_0001.jpg
print("lfw/" + query_image + "/" + query_image + "_0001.jpg")
st.image(
    "lfw/" + query_image + "/" + query_image + "_0001.jpg",
    caption="Query Image",
    use_column_width=True,
)
# Find the 10 most similar images
if st.button("Find Similar Images"):
    # Call the get nearest neighbors function
    nearest10 = get_nearest_neighbors(
        "lfw/" + query_image + "/" + query_image + "_0001.jpg"
    )