app.py

import torch
import torch.nn.functional as F
import numpy as np
import pandas as pd
import gradio as gr
from transformers import CLIPModel, CLIPTokenizer
from sklearn.metrics.pairwise import cosine_similarity
from datasets import load_dataset
from PIL import Image
import traceback


MODEL_ID  = "openai/clip-vit-base-patch32"
tokenizer = CLIPTokenizer.from_pretrained(MODEL_ID)
model     = CLIPModel.from_pretrained(MODEL_ID)
model.eval()


embeddings = np.load("car_embeddings.npz")["embeddings"]
metadata   = pd.read_csv("car_metadata.csv")


ds = load_dataset("DamianBoborzi/car_images", split="train").select(range(10000))


top3_default_indices = (
    metadata
    .nlargest(3, "aesthetic_score")["dataset_index"]
    .astype(int)
    .tolist()
)

def get_default_images():
    imgs = []
    for ds_idx in top3_default_indices:
        img = ds[ds_idx]["image"]
        if not isinstance(img, Image.Image):
            img = Image.fromarray(img)
        imgs.append(img.convert("RGB"))
    return imgs

def get_text_vector(user_query: str) -> np.ndarray:
    inputs = tokenizer(user_query, return_tensors="pt",
                       padding=True, truncation=True, max_length=77)
    with torch.no_grad():
        text_out = model.text_model(**inputs)
        feat     = model.text_projection(text_out.pooler_output)
    return F.normalize(feat, p=2, dim=-1).cpu().numpy()


def recommend(user_query: str):
    if not user_query or not user_query.strip():
        defaults = get_default_images()
        rows     = [metadata[metadata["dataset_index"] == i].iloc[0]
                    for i in top3_default_indices]
        results  = []
        for rank, (img, row) in enumerate(zip(defaults, rows)):
            score    = float(row["aesthetic_score"])
            car_name = str(row["text"]).split(",")[0].split("(")[0].strip()[:55]
            label    = f"#{rank+1}  Aesthetic Score: {score:.2f}\n{car_name}"
            results += [img, label]
        return results[0], results[1], results[2], results[3], results[4], results[5]

    try:
        query_vec = get_text_vector(user_query)
        scores    = cosine_similarity(query_vec, embeddings).flatten()
        top3      = scores.argsort()[-3:][::-1]

        results = []
        for rank, df_idx in enumerate(top3):
            row    = metadata.iloc[int(df_idx)]
            score  = float(scores[int(df_idx)])
            ds_idx = int(row["dataset_index"])

            img = ds[ds_idx]["image"]
            if not isinstance(img, Image.Image):
                img = Image.fromarray(img)
            img = img.convert("RGB")

            car_name = str(row["text"]).split(",")[0].split("(")[0].strip()[:55]
            label    = f"#{rank+1}  Score: {score:.4f}  |  Cluster {int(row['cluster'])}\n{car_name}"
            results += [img, label]

        return results[0], results[1], results[2], results[3], results[4], results[5]

    except Exception as e:
        traceback.print_exc()
        msg = f"Error: {e}"
        return None, msg, None, msg, None, msg


default_imgs   = get_default_images()
default_rows   = [metadata[metadata["dataset_index"] == i].iloc[0]
                  for i in top3_default_indices]
default_labels = [
    f"#{r+1}  Aesthetic Score: {float(row['aesthetic_score']):.2f}\n"
    f"{str(row['text']).split(',')[0].split('(')[0].strip()[:55]}"
    for r, row in enumerate(default_rows)
]


with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🏎️ Car Recommendation Engine")
    gr.Markdown(
        "Describe the car you're looking for and CLIP will find the 3 most similar cars "
        "from a dataset of **10,000 real car images**.\n\n"
        "*Default view shows the top 3 highest-rated cars by aesthetic score.*"
    )

    with gr.Row():
        query_box  = gr.Textbox(
            label="Describe your ideal car",
            placeholder="e.g. A red aggressive supercar with wide body kit",
            scale=5
        )
        search_btn = gr.Button("🔍 Find Matches", variant="primary", scale=1)

    gr.Markdown("### 🏆 Top 3 Matches")
    with gr.Row():
        with gr.Column():
            img1 = gr.Image(label="Match 1", type="pil", value=default_imgs[0])
            lbl1 = gr.Textbox(show_label=False, interactive=False, value=default_labels[0])
        with gr.Column():
            img2 = gr.Image(label="Match 2", type="pil", value=default_imgs[1])
            lbl2 = gr.Textbox(show_label=False, interactive=False, value=default_labels[1])
        with gr.Column():
            img3 = gr.Image(label="Match 3", type="pil", value=default_imgs[2])
            lbl3 = gr.Textbox(show_label=False, interactive=False, value=default_labels[2])

    search_btn.click(fn=recommend, inputs=query_box,
                     outputs=[img1, lbl1, img2, lbl2, img3, lbl3])

    
    gr.Markdown("---")
    gr.Markdown("## 📽️ Project Presentation Video")
    with gr.Row():
        gr.HTML("""
        <div style="display: flex; justify-content: center; align-items: center; width: 100%;">
            <iframe width="760" height="415"
            src="https://www.youtube.com/embed/q1pvsyz1A9g"
            title="YouTube video player"
            frameborder="0"
            allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture"
            allowfullscreen>
            </iframe>
        </div>
        """)

    
    gr.Examples(
        examples=[
            ["A red aggressive supercar with wide body kit"],
            ["Vintage blue convertible classic car"],
            ["White family SUV with clean modern design"],
            ["Black 3D rendered sports coupe low angle"],
        ],
        inputs=query_box,
    )

demo.launch()