app.py

# app.py
"""
Streamlit BLIP-2 Image Captioning demo
- Uses HuggingFace transformers' Blip2Processor + Blip2ForConditionalGeneration
- Caches the model & processor with st.cache_resource so they load once per Space/session.
- Designed for deployment on Hugging Face Spaces (use Docker SDK / Streamlit template).
"""

import streamlit as st
from PIL import Image
import io
import torch
from transformers import Blip2Processor, Blip2ForConditionalGeneration

st.set_page_config(
    page_title="BLIP-2 Image Captioning",
    layout="wide",
    initial_sidebar_state="expanded",
)

# --- Sidebar / Info ---
st.sidebar.title("BLIP-2 Caption Demo")
st.sidebar.markdown(
    """
    Upload an image and BLIP-2 will generate a caption.
    - Model choices: choose a BLIP-2 model (large models may need GPU / won’t fit on CPU).
    - For Spaces deployment, prefer smaller/flan-xl variants or use inference API.
    """
)

# Recommended default model (change if you want)
DEFAULT_MODEL = "Salesforce/blip2-opt-2.7b"

@st.cache_resource(show_spinner=False)
def load_model_and_processor(model_name: str):
    """Load and cache the BLIP-2 processor and model."""
    # Note: large models will require a GPU; smaller variants or hosted inference endpoints recommended for CPU-only Spaces.
    processor = Blip2Processor.from_pretrained(model_name)
    model = Blip2ForConditionalGeneration.from_pretrained(model_name)
    # move to GPU if available
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)
    return processor, model, device

def generate_caption(processor, model, device, pil_image: Image.Image, max_new_tokens=50, num_beams=4):
    """Generate caption text for a PIL image using BLIP-2."""
    if pil_image.mode != "RGB":
        pil_image = pil_image.convert("RGB")

    inputs = processor(images=pil_image, return_tensors="pt").to(device)

    # Generate - tune generation args as needed
    generated_ids = model.generate(**inputs, max_new_tokens=max_new_tokens, num_beams=num_beams)
    # decode using the tokenizer in the processor
    caption = processor.tokenizer.decode(generated_ids[0], skip_special_tokens=True)
    return caption

# --- UI layout ---
col1, col2 = st.columns([1, 1.2])

with col1:
    st.header("Upload image")
    uploaded = st.file_uploader("Choose an image", type=["png", "jpg", "jpeg"], accept_multiple_files=False)

    st.markdown("**Model selection**")
    model_name = st.selectbox(
        "Pick BLIP-2 model (large models may not run on CPU)",
        options=[
            "Salesforce/blip2-flan-t5-xl",
            "Salesforce/blip2-opt-2.7b",
            "Salesforce/blip2-flan-t5-xxl",
        ],
        index=1 if DEFAULT_MODEL.endswith("2.7b") else 0,
        help="Large models require GPU or HF Inference API; choose smaller if you have no GPU.",
    )

    max_tokens = st.slider("Max caption length (tokens)", min_value=10, max_value=200, value=50)
    num_beams = st.slider("Beam search width (num_beams)", min_value=1, max_value=8, value=4)

    st.write("---")
    st.markdown("Tips:")
    st.markdown(
        "- If deploying on CPU-only Spaces, use a smaller/flan model or use the Hugging Face Inference API.\n"
        "- Model loading is cached to speed up subsequent requests."
    )

with col2:
    st.header("Preview & Caption")
    if uploaded is None:
        st.info("Upload an image on the left to generate a caption.")
        st.empty()
    else:
        # display image
        image_bytes = uploaded.read()
        pil_image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
        st.image(pil_image, use_column_width=True)

        # Load model & processor (cached)
        with st.spinner("Loading model (cached after first load)..."):
            processor, model, device = load_model_and_processor(model_name)

        # Generate caption
        if st.button("Generate caption"):
            with st.spinner("Generating caption..."):
                try:
                    caption = generate_caption(processor, model, device, pil_image, max_new_tokens=max_tokens, num_beams=num_beams)
                    st.success("Caption generated")
                    st.markdown(f"**Caption:** {caption}")
                    # Provide a copy button and simple download
                    st.download_button("Download caption (.txt)", caption, file_name="caption.txt")
                except Exception as e:
                    st.error(f"Error during generation: {e}")
                    st.info("If model is too large or out-of-memory, try a smaller model or use GPU.")

# --- Footer / Resources ---
st.markdown("---")
st.markdown(
    "Built with BLIP-2 + Transformers. For production or public Spaces hosting, consider using Hugging Face Inference API or a smaller model variant to avoid OOM on CPU-only hosts."
)
st.caption("Docs: BLIP-2 (Transformers), Hugging Face Spaces (Streamlit), Streamlit caching & uploader.")