app.py

# -*- coding: utf-8 -*-
"""image_model.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1SRq7ZIsAwqDYyV4UWM9SZJuRYJvQe59o
"""

import os  # For reading environment variables
import shutil  # For directory cleanup
import zipfile  # For extracting model archives
import pathlib  # For path manipulations
import tempfile  # For creating temporary files/directories

import gradio  # For interactive UI
import pandas  # For tabular data handling
import PIL.Image # For image I/O

import huggingface_hub # For downloading model assets
import autogluon.multimodal  # For loading AutoGluon image classifier

# Hardcoded Hub model (native zip)
MODEL_REPO_ID = "SebastianAndreu/2025-24679-HW1-Part2-image-autogluon-predictor"
ZIP_FILENAME  = "autogluon_image_predictor_dir.zip"
HF_TOKEN = os.getenv("HF_TOKEN", None)

# Local cache/extract dirs
CACHE_DIR   = pathlib.Path("hf_assets")
EXTRACT_DIR = CACHE_DIR / "predictor_native"

# Download & load the native predictor
def _prepare_predictor_dir() -> str:
    CACHE_DIR.mkdir(parents=True, exist_ok=True)
    local_zip = huggingface_hub.hf_hub_download(
        repo_id=MODEL_REPO_ID,
        filename=ZIP_FILENAME,
        repo_type="model",
        token=HF_TOKEN,
        local_dir=str(CACHE_DIR),
        local_dir_use_symlinks=False,
    )
    if EXTRACT_DIR.exists():
        shutil.rmtree(EXTRACT_DIR)
    EXTRACT_DIR.mkdir(parents=True, exist_ok=True)
    with zipfile.ZipFile(local_zip, "r") as zf:
        zf.extractall(str(EXTRACT_DIR))
    contents = list(EXTRACT_DIR.iterdir())
    predictor_root = contents[0] if (len(contents) == 1 and contents[0].is_dir()) else EXTRACT_DIR
    return str(predictor_root)

PREDICTOR_DIR = _prepare_predictor_dir()
PREDICTOR = autogluon.multimodal.MultiModalPredictor.load(PREDICTOR_DIR)

# Explicit class labels (edit copy as desired)
CLASS_LABELS = {0: "No Duo :(", 1: "DUO :-)"}

# Helper to map model class -> human label
def _human_label(c):
    try:
        ci = int(c)
        return CLASS_LABELS.get(ci, str(c))
    except Exception:
        return CLASS_LABELS.get(c, str(c))

# Do the prediction!
def do_predict(pil_img: PIL.Image.Image):
    # Make sure there's actually an image to work with
    if pil_img is None:
        return "No image provided.", {}, pandas.DataFrame(columns=["Predicted label", "Confidence (%)"])

    # IF we have something to work with, save it and prepare the input
    tmpdir = pathlib.Path(tempfile.mkdtemp())
    img_path = tmpdir / "input.png"
    pil_img.save(img_path)

    df = pandas.DataFrame({"image": [str(img_path)]})  # For AutoGluon expected input format

    # For class probabilities
    proba_df = PREDICTOR.predict_proba(df)

    # For user-friendly column names
    proba_df = proba_df.rename(columns={0: "No Duo :(", 1: "DUO :-)"})
    row = proba_df.iloc[0]

    # For pretty ranked dict expected by gr.Label
    pretty_dict = {
        "No Duo :(": float(row.get("No Duo :(", 0.0)),
        "DUO :-)": float(row.get("DUO :-)", 0.0)),
    }

    # Calculate confidence interval (a simple representation for demonstration)
    # This is not a statistically rigorous CI, but rather a representation of the probability spread.
    confidence_info = f"No Duo Probability: {pretty_dict['No Duo :(']:.2f}, DUO Probability: {pretty_dict['DUO :-)']:.2f}"

    return pretty_dict, confidence_info

EXAMPLES = [
    ["duo_1.jpg"],
    ["duo_2.jpg"],
    ["no_duo_1.jpg"],
    ["no_duo_2.jpg"]
]

# Gradio UI
with gradio.Blocks() as demo:

    # Provide an introduction
    gradio.Markdown("# Is Duo Here?")
    gradio.Markdown("""
    This is an app that demonstrates a binary classifier using SebastianAndreu/2025-24679-HW1-Part2-image-autogluon-predictor model
    based on the scottymcgee/duo-image-dataset imageset. This performs binary classification of outdoor images to see
    if the bird from Duolingo is in them. It has an accuracy of 80% based on the model so can have errors. 
    """)

    # Interface for the incoming image
    image_in = gradio.Image(type="pil", label="Input image", sources=["upload", "webcam"])

    # Interface elements to show htte result and probabilities
    proba_pretty = gradio.Label(num_top_classes=2, label="Class probabilities")

    # Whenever a new image is uploaded, update the result
    image_in.change(fn=do_predict, inputs=[image_in], outputs=[proba_pretty])

      # For clickable example images
    gradio.Examples(
        examples=EXAMPLES,
        inputs=[image_in],
        label="Representative examples",
        examples_per_page=8,
        cache_examples=False,
    )

# Gradio UI
with gradio.Blocks() as demo:

    # Provide an introduction
    gradio.Markdown("# Is Duo Here?")
    gradio.Markdown("""
    This is a simple app that demonstrates how to use an autogluon multimodal
    predictor in a gradio space to predict the contents of a picture. To use,
    just upload a photo. The result should be generated automatically.
    """)

    # Interface for the incoming image
    image_in = gradio.Image(type="pil", label="Input image", sources=["upload", "webcam"])

    # Interface elements to show htte result and probabilities
    proba_pretty = gradio.Label(num_top_classes=2, label="Class probabilities")
    confidence_output = gradio.Textbox(label="Confidence Information")

    # Whenever a new image is uploaded, update the result
    image_in.change(fn=do_predict, inputs=[image_in], outputs=[proba_pretty, confidence_output])

      # For clickable example images
    gradio.Examples(
        examples=EXAMPLES,
        inputs=[image_in],
        label="Representative examples",
        examples_per_page=8,
        cache_examples=False,
    )

# Launch here
if __name__ == "__main__":
    demo.launch()