app.py

import gradio as gr
import pandas as pd
from green_score import GREEN
import spaces

@spaces.GPU  # Add the GPU decorator for functions that need GPU access
def run_green(ref_text, hyp_text, model_name="StanfordAIMI/GREEN-radllama2-7b"):
    refs = [ref_text.strip()]
    hyps = [hyp_text.strip()]

    green_scorer = GREEN(model_name, output_dir=".")
    mean, std, green_score_list, summary, result_df = green_scorer(refs, hyps)

    final_summary = (
        f"**Model:** {model_name}\n\n"
        f"**Mean GREEN Score:** {mean:.4f}\n"
        f"**Std Deviation:** {std:.4f}\n\n"
        f"**Detailed Summary:**\n{summary}"
    )
    return final_summary, pd.DataFrame(result_df), green_score_list

# Example pairs
examples = {
    "Example 1": {
        "ref": "Interstitial opacities without changes.",
        "hyp": "Interstitial opacities at bases without changes.",
    },
    "Example 2": {
        "ref": "The heart size is normal. Lungs are clear without any infiltrates.",
        "hyp": "The heart size is mildly enlarged. Mild infiltrates in the left upper lobe.",
    },
    "Example 3": {
        "ref": "Lung volumes are low, causing bronchovascular crowding. The cardiomediastinal silhouette is unremarkable.",
        "hyp": "Endotracheal tubes have been removed. Pulmonary aeration is slightly improved with no pleural effusions.",
    }
}

def update_fields(choice):
    if choice == "Custom":
        return gr.update(value="", interactive=True), gr.update(value="", interactive=True)
    else:
        return gr.update(value=examples[choice]["ref"], interactive=False), gr.update(value=examples[choice]["hyp"], interactive=False)

with gr.Blocks(title="GREEN Score Evaluation Demo") as demo:
    gr.Markdown("# GREEN Score Evaluation Demo")
    gr.Markdown(
        "This demo evaluates a single pair of reference and hypothesis reports to compute the GREEN score."
    )

    with gr.Row():
        choice = gr.Radio(
            label="Choose Input Type",
            choices=["Custom"] + list(examples.keys()),
            value="Custom",
            interactive=True
        )

    ref_input = gr.Textbox(
        label="Reference Report",
        lines=3
    )
    hyp_input = gr.Textbox(
        label="Hypothesis Report",
        lines=3
    )

    choice.change(
        update_fields,
        inputs=choice,
        outputs=[ref_input, hyp_input],
    )

    model_name_input = gr.Textbox(
        label="Model Name",
        value="StanfordAIMI/GREEN-radllama2-7b",
        placeholder="Enter the HuggingFace model name"
    )

    run_button = gr.Button("Compute GREEN Score")
    summary_output = gr.Markdown()
    df_output = gr.DataFrame()
    score_list_output = gr.JSON()

    run_button.click(
        run_green,
        inputs=[ref_input, hyp_input, model_name_input],
        outputs=[summary_output, df_output, score_list_output]
    )

demo.launch()