app.py

import os
import streamlit as st
from PIL import Image
import torch
from transformers import PaliGemmaProcessor, PaliGemmaForConditionalGeneration

# Get Hugging Face API key from environment variables
HF_TOKEN = os.getenv("HF_KEY")

# Ensure API key is available
if not HF_TOKEN:
    st.error("❌ Hugging Face API key not found! Set it as 'HF_KEY' in Spaces secrets.")
    st.stop()

# Load the model and processor with authentication
@st.cache_resource
def load_model():
    model_id = "google/paligemma2-3b-mix-224"
    
    try:
        model = PaliGemmaForConditionalGeneration.from_pretrained(
            model_id, token=HF_TOKEN, torch_dtype=torch.bfloat16, device_map="auto"
        ).eval()

        processor = PaliGemmaProcessor.from_pretrained(
            model_id, token=HF_TOKEN
        )

        return processor, model

    except Exception as e:
        st.error(f"❌ Error loading model: {str(e)}")
        st.stop()

processor, model = load_model()

# Streamlit UI
st.title("🖼️ Image Understanding with PaliGemma")

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

if uploaded_file:
    image = Image.open(uploaded_file).convert("RGB")
    st.image(image, caption="Uploaded Image", use_container_width=True)

    # User input for task selection
    task = st.selectbox(
        "Select a task:",
        ["Generate a caption", "Answer a question", "Detect objects", "Generate segmentation"]
    )

    # User prompt
    prompt = st.text_area("Enter a prompt (e.g., 'Describe the image' or 'What objects are present?')")

    if st.button("Run"):
        if prompt:
            inputs = processor(text=prompt, images=image, return_tensors="pt").to(torch.bfloat16).to(model.device)
            input_len = inputs["input_ids"].shape[-1]  # Get input length

            with torch.inference_mode():
                generation = model.generate(**inputs, max_new_tokens=100, do_sample=False)
                generation = generation[0][input_len:]  # Remove input tokens from output
                answer = processor.decode(generation, skip_special_tokens=True)

            st.success(f"✅ Result: {answer}")