app.py

import io
import tarfile
import numpy as np
import random as RAND
import torch
import torchvision.transforms as TRNSFM
import torchvision.models as MDLS
from PIL import Image as IMG
from scipy.spatial.distance import cosine
import streamlit as st

def similar(image): # Function for Streamlit App
    pict = form(IMG.open(image).convert('RGB'))
    pictFeats = mod(pict.unsqueeze(0)).detach().numpy().squeeze()

    for na, feat in feats.items():
        s = 1 - cosine(pictFeats, feat)
        simScores.append((na, s))

    simScores.sort(key=lambda x: x[1], reverse=True)

    st.write("### Selected Image")
    test = IMG.open(image)
    test.show()
    print('\n')

    # 10 Most Similar Images from Dictionary
    st.write("### 10 Most Similar Images")
    for na in simScores[:10]:
        for x in range(10):
            st.write(f"### {x}")

        new_na = na[:3] + "images/" + na[3:]
        new_path = "http://vis-www.cs.umass.edu/" + new_na
        simImages = IMG.open(new_path)
        simImages.show()

mod = MDLS.resnet50(pretrained=True)
mod.eval()
mod = torch.nn.Sequential(*list(mod.children())[:-1])
inFile = tarfile.open('./datasets/lfw.tar', 'r')

feats = {}
simScores = [] # Similarity Scores for Later

form = TRNSFM.Compose([
    TRNSFM.Resize(256),
    TRNSFM.CenterCrop(224),
    TRNSFM.ToTensor(),
    TRNSFM.Normalize(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225])
]) # Image Pre-processing

stuffs = inFile.getmembers()

for stuff in stuffs: # Going through the TAR file
    f = inFile.extractfile(stuff)

    if stuff.isdir():
        continue

    if stuff.name.lower().endswith(('.jpg', '.jpeg', '.png')): 
        n = stuff.name
        pic = form(IMG.open(io.BytesIO(f.read())).convert('RGB')) # Pre-processes the image before feeding it into the model

        feats[n] = mod(pic.unsqueeze(0)).detach().numpy().squeeze()

# Stuff for App
st.title("Similar Image Finder")

upload = st.file_uploader("Upload an Image...", type=['.jpg', '.jpeg', '.png'])

if upload is not None:
    similar(upload)

st.write("## OR")

# Random Image Selector from 5 Pictures
randImages = [
    './datasets/random-images/img1.jpg',
    './datasets/random-images/img2.jpg',
    './datasets/random-images/img3.jpg',
    './datasets/random-images/img4.jpg',
    './datasets/random-images/img5.jpg'
]

if st.button("Surprise Me!"): # Button
    imageOptOne = RAND.choice(randImages)
    similar(imageOptOne)
        
inFile.close()