app.py

import gradio as gr
import pandas as pd
import numpy as np
import torch
from transformers import CLIPModel, CLIPProcessor

# -----------------------------
# 1. Load model & processor
# -----------------------------
model_name = "openai/clip-vit-base-patch32"
model = CLIPModel.from_pretrained(model_name)
processor = CLIPProcessor.from_pretrained(model_name)

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
model.eval()

# -----------------------------
# 2. Load your saved embeddings
# -----------------------------
df = pd.read_parquet("animal_embeddings.parquet")

embeddings = df.drop(columns=["label", "index"]).values
labels = df["label"].tolist()
indices = df["index"].tolist()

# Load dataset to retrieve images
from datasets import load_dataset
dataset = load_dataset("mountassir/animals-10")["train"]
sampled_data = dataset.select(indices)

label_names = dataset.features["label"].names

# -----------------------------
# 3. Helper functions
# -----------------------------

def embed_image_query(pil_image):
    with torch.no_grad():
        inputs = processor(images=pil_image, return_tensors="pt").to(device)
        feats = model.get_image_features(**inputs)
        feats = feats / feats.norm(dim=-1, keepdim=True)
    return feats.squeeze().cpu().numpy()

def embed_text_query(text):
    with torch.no_grad():
        inputs = processor(text=[text], return_tensors="pt").to(device)
        feats = model.get_text_features(**inputs)
        feats = feats / feats.norm(dim=-1, keepdim=True)
    return feats.squeeze().cpu().numpy()

from sklearn.metrics.pairwise import cosine_similarity

def get_top_k(query_emb, k=3):
    sims = cosine_similarity(query_emb.reshape(1, -1), embeddings)[0]
    idxs = np.argsort(sims)[::-1][:k]
    return idxs, sims[idxs]

# -----------------------------
# 4. Gradio functions
# -----------------------------

def gradio_image_search(image):
    query_emb = embed_image_query(image)
    idxs, scores = get_top_k(query_emb, 3)
    results = [sampled_data[i]["image"] for i in idxs]
    return results

def gradio_text_search(text):
    query_emb = embed_text_query(text)
    idxs, scores = get_top_k(query_emb, 3)
    results = [sampled_data[i]["image"] for i in idxs]
    return results

# -----------------------------
# 5. Build Gradio App
# -----------------------------
with gr.Blocks() as demo:
    gr.Markdown("""
# 🐾 Animal Similarity Finder

Welcome! This app allows you to find animals that look visually similar using image and text embeddings.

 How it works
- The model uses **CLIP embeddings** to compare your input with a database of animal images.
- It returns the **Top 3 most similar images** from the Animals-10 dataset.

Image Search
Upload a picture of an animal (dog, cat, spider, butterfly, horse, etc.).  
The app will analyze the image and show you the 3 closest matches based on **visual similarity**.

Text Search
Type a description like:
- **"pet"** → finds dogs & cats  
- **"bug"** → finds spiders  
- **"farm animal"** → finds sheep, cows, horses  
- **"bird"** → finds chickens  

The model converts your text into an embedding and returns the 3 images most related to your description.

 Behind the scenes
- Embeddings generated with **CLIP (ViT-B/32)**  
- Similarity is computed using **cosine similarity**  
- All embeddings are precomputed for speed

Enjoy exploring the animal dataset! 🐶🐱🐴🦋🕷️
""")


    with gr.Tab("Image Search"):
        img_in = gr.Image(type="pil")
        img_out = gr.Gallery(label="Top 3 Results", columns=3)
        btn1 = gr.Button("Search")
        btn1.click(fn=gradio_image_search, inputs=img_in, outputs=img_out)

    with gr.Tab("Text Search"):
        txt_in = gr.Textbox(label="e.g. 'pet', 'bug', 'farm animal'")
        txt_out = gr.Gallery(label="Top 3 Results", columns=3)
        btn2 = gr.Button("Search")
        btn2.click(fn=gradio_text_search, inputs=txt_in, outputs=txt_out)

demo.launch()