app.py

import torch
import numpy as np
import pandas as pd
from PIL import Image
import gradio as gr
from transformers import CLIPModel, CLIPProcessor
from datasets import load_dataset

# --- Device & model ---
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
MODEL_NAME = "openai/clip-vit-base-patch32"

model = CLIPModel.from_pretrained(MODEL_NAME).to(DEVICE)
processor = CLIPProcessor.from_pretrained(MODEL_NAME)
model.eval()

# --- Load dataset & embeddings ---
dataset = load_dataset("JotDe/birds", split="train")

embedding_df = pd.read_parquet("bird_image_embeddings.parquet")
embedding_vectors = embedding_df.filter(like="emb_").values


# --- Scene detection ---
def detect_scene(image: Image.Image):
    image = image.convert("RGB")
    arr = np.array(image) / 255.0

    mean_r = arr[:, :, 0].mean()
    mean_g = arr[:, :, 1].mean()
    mean_b = arr[:, :, 2].mean()

    if mean_b > 0.45 and mean_b > mean_g and mean_b > mean_r:
        return "sky"
    elif mean_g > 0.40 and mean_g > mean_r:
        return "forest"
    elif mean_b > 0.35 and mean_g < 0.35:
        return "water"
    else:
        return "mixed"


# --- Embed user image ---
def embed_user_image(image: Image.Image):
    image = image.convert("RGB")
    inputs = processor(images=image, return_tensors="pt").to(DEVICE)

    with torch.no_grad():
        features = model.get_image_features(**inputs)
        features = features / features.norm(dim=-1, keepdim=True)

    return features.cpu().numpy().flatten()


# --- Recommendation function ---
def recommend_images(image: Image.Image):
    user_emb = embed_user_image(image)
    scene = detect_scene(image)

    # Add scene column if missing
    if "scene" not in embedding_df.columns:
        scenes = []
        for i in range(len(embedding_df)):
            img = dataset[int(embedding_df.iloc[i]["row_in_subset"])]["image"]
            scenes.append(detect_scene(img))
        embedding_df["scene"] = scenes

    # Filter by scene
    filtered_df = embedding_df[embedding_df["scene"] == scene]

    # Fallback if too few images
    if len(filtered_df) < 10:
        filtered_df = embedding_df

    filtered_vectors = filtered_df.filter(like="emb_").values

    similarities = filtered_vectors @ user_emb
    top_indices = np.argsort(similarities)[-3:][::-1]

    results = []
    for idx in top_indices:
        row = filtered_df.iloc[int(idx)]
        img = dataset[int(row["row_in_subset"])]["image"]
        results.append(img)

    return results


# --- Gradio interface ---
interface = gr.Interface(
    fn=recommend_images,
    inputs=gr.Image(type="pil", label="Upload a bird image"),
    outputs=[
        gr.Image(label="Recommendation 1"),
        gr.Image(label="Recommendation 2"),
        gr.Image(label="Recommendation 3"),
    ],
    title="Bird Image Recommendation System",
    description="Scene-aware bird image recommendations using CLIP embeddings."
)

interface.launch()


import gradio as gr

with gr.Blocks() as demo:
    gr.HTML("""
        <iframe
            width="1280"
            height="720"
            src="https://www.youtube.com/embed/bWRN4ZAKyQ4"
            title="YouTube video player"
            frameborder="0"
            allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture"
            allowfullscreen>
        </iframe>
    """)

demo.launch()