app.py

import gradio as gr
import pandas as pd

CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
CITATION_BUTTON_TEXT = r"""@misc{aienergyscore-leaderboard,
    author = {Sasha Luccioni and Boris Gamazaychikov and Emma Strubell and Sara Hooker and Yacine Jernite and Carole-Jean Wu and Margaret Mitchell},
    title = {AI Energy Score Leaderboard - February 2025},
    year = {2025},
    publisher = {Hugging Face},
    howpublished = "\url{https://huggingface.co/spaces/AIEnergyScore/Leaderboard}",
}"""

# List of tasks (CSV filenames)
tasks = [
    'asr.csv',
    'object_detection.csv',
    'text_classification.csv',
    'image_captioning.csv',
    'question_answering.csv',
    'text_generation.csv',
    'image_classification.csv',
    'sentence_similarity.csv',
    'image_generation.csv',
    'summarization.csv'
]

color_map = {"1": "red", "2": "orange", "3": "yellow", "4": "lightgreen", "5": "green"} # Keep color map

def format_stars(score):
    try:
        score_int = int(score)
    except Exception:
        score_int = 0
    # Render stars in black with a slightly larger font
    return f'<span style="color: black !important; font-size:1.5em !important;">{"★" * score_int}</span>'

def make_link(mname):
    parts = str(mname).split('/')
    display_name = parts[1] if len(parts) > 1 else mname
    return f'[{display_name}](https://huggingface.co/{mname})'

# --- Leaderboard Table Functions (Using gr.HTML Component) ---

def create_minimal_bar_html(energy_value_wh, energy_score, max_energy_value):
    """Generates HTML for the minimal bar chart."""
    if max_energy_value <= 0: # Avoid division by zero if max energy is 0 or negative
        bar_percentage = 0
    else:
        bar_percentage = min(100, (energy_value_wh / max_energy_value) * 100) # Cap at 100%
    bar_color = color_map.get(str(energy_score), "gray") # Default color if score is unexpected

    html = f"""
    <div style="display: flex; align-items: center; gap: 5px;">
        <div style="width: {bar_percentage}%; height: 10px; background-color: {bar_color}; border-radius: 2px;"></div>
        <span>{energy_value_wh:.4f} Wh</span>
    </div>
    """
    return html


def get_model_names(task):
    df = pd.read_csv('data/energy/' + task)
    if df.columns[0].startswith("Unnamed:"):
        df = df.iloc[:, 1:]
    df['total_gpu_energy'] = pd.to_numeric(df['total_gpu_energy'], errors='raise') * 1000 # kWh to Wh conversion
    df['energy_score'] = df['energy_score'].astype(int)

    df = df.sort_values(by='total_gpu_energy') # Sort BEFORE creating HTML column
    max_energy_for_task = df['total_gpu_energy'].max() # Calculate max energy for this task

    # Create HTML bar chart for GPU Energy column, passing dynamic max_energy_for_task
    df['GPU Energy (Wh)'] = df.apply(lambda row: gr.HTML(create_minimal_bar_html(row['total_gpu_energy'], row['energy_score'], max_energy_for_task)), axis=1)

    df['Model'] = df['model'].apply(make_link)
    df['Score'] = df['energy_score'].apply(format_stars)
    df = df[['Model', 'GPU Energy (Wh)', 'Score']] # Keep only these columns
    return df

def get_all_model_names():
    all_df = pd.DataFrame()
    max_energy_overall = 0 # Initialize overall max energy
    for task in tasks:
        df = pd.read_csv('data/energy/' + task)
        df['total_gpu_energy'] = pd.to_numeric(df['total_gpu_energy'], errors='raise') * 1000 # kWh to Wh conversion
        df['energy_score'] = df['energy_score'].astype(int)
        all_df = pd.concat([all_df, df], ignore_index=True)

    all_df = all_df.drop_duplicates(subset=['model'])
    all_df = all_df.sort_values(by='total_gpu_energy') # Sort ALL DATA before calculating max and creating HTML

    max_energy_overall = all_df['total_gpu_energy'].max() # Calculate overall max AFTER sorting

    # Create HTML bar chart for GPU Energy column, passing dynamic max_energy_overall
    all_df['GPU Energy (Wh)'] = all_df.apply(lambda row: gr.HTML(create_minimal_bar_html(row['total_gpu_energy'], row['energy_score'], max_energy_overall)), axis=1)
    all_df['Model'] = all_df['model'].apply(make_link)
    all_df['Score'] = all_df['energy_score'].apply(format_stars)
    all_df = all_df[['Model', 'GPU Energy (Wh)', 'Score']]
    return all_df[['Model', 'GPU Energy (Wh)', 'Score']]


def get_text_generation_model_names(model_class):
    df = pd.read_csv('data/energy/text_generation.csv')
    if df.columns[0].startswith("Unnamed:"):
        df = df.iloc[:, 1:]
    if 'class' in df.columns:
        df = df[df['class'] == model_class]
    df['total_gpu_energy'] = pd.to_numeric(df['total_gpu_energy'], errors='raise') * 1000 # kWh to Wh conversion
    df['energy_score'] = df['energy_score'].astype(int)

    df = df.sort_values(by='total_gpu_energy') # Sort BEFORE creating HTML column
    max_energy_for_class = df['total_gpu_energy'].max() # Calculate max energy for this class

    # Create HTML bar chart for GPU Energy column, passing dynamic max_energy_for_class
    df['GPU Energy (Wh)'] = df.apply(lambda row: gr.HTML(create_minimal_bar_html(row['total_gpu_energy'], row['energy_score'], max_energy_for_class)), axis=1)

    df['Model'] = df['model'].apply(make_link)
    df['Score'] = df['energy_score'].apply(format_stars)
    df = df[['Model', 'GPU Energy (Wh)', 'Score']] # Keep only these columns
    return df

def update_text_generation(model_class):
    table = get_text_generation_model_names(model_class)
    return table

# --- Build the Gradio Interface (Plots Removed, Tables with Dynamic Bars using gr.HTML) ---

demo = gr.Blocks(css="""
.gr-dataframe table {
    table-layout: fixed;
    width: 100%;
}
.gr-dataframe th, .gr-dataframe td {
    max-width: 150px;
    white-space: nowrap;
    overflow: hidden;
    text-overflow: ellipsis;
}
/* CSS for minimal bar chart inside table cell */
.minimal-bar-container {
    display: flex;
    align-items: center;
    gap: 5px; /* space between bar and text */
}
.minimal-bar {
    height: 10px;
    background-color: blue; /* default, will be overridden by dynamic color */
    border-radius: 2px;
}
""")

with demo:
    gr.Markdown(
        """# AI Energy Score Leaderboard
### Welcome to the leaderboard for the [AI Energy Score Project!](https://huggingface.co/AIEnergyScore)
Select different tasks to see scored models. Submit open models for testing and learn about testing proprietary models via the [submission portal](https://huggingface.co/spaces/AIEnergyScore/submission_portal)"""
    )

    with gr.Tabs():
        # --- Text Generation Tab with Dropdown for Model Class ---
        with gr.TabItem("Text Generation 💬"):
            # Dropdown moved above the leaderboard
            model_class_dropdown = gr.Dropdown(choices=["A", "B", "C"],
                                                 label="Select Model Class",
                                                 value="A")
            tg_table = gr.Dataframe(get_text_generation_model_names("A")) # No datatype="markdown" here
            # Update table when the dropdown value changes
            model_class_dropdown.change(fn=update_text_generation,
                                        inputs=model_class_dropdown,
                                        outputs=[tg_table])

        with gr.TabItem("Image Generation 📷"):
            table = gr.Dataframe(get_model_names('image_generation.csv')) # No datatype="markdown" here

        with gr.TabItem("Text Classification 🎭"):
            table = gr.Dataframe(get_model_names('text_classification.csv')) # No datatype="markdown" here

        with gr.TabItem("Image Classification 🖼️"):
            table = gr.Dataframe(get_model_names('image_classification.csv')) # No datatype="markdown" here

        with gr.TabItem("Image Captioning 📝"):
            table = gr.Dataframe(get_model_names('image_captioning.csv')) # No datatype="markdown" here

        with gr.TabItem("Summarization 📃"):
            table = gr.Dataframe(get_model_names('summarization.csv')) # No datatype="markdown" here

        with gr.TabItem("Automatic Speech Recognition 💬"):
            table = gr.Dataframe(get_model_names('asr.csv')) # No datatype="markdown" here

        with gr.TabItem("Object Detection 🚘"):
            table = gr.Dataframe(get_model_names('object_detection.csv')) # No datatype="markdown" here

        with gr.TabItem("Sentence Similarity 📚"):
            table = gr.Dataframe(get_model_names('sentence_similarity.csv')) # No datatype="markdown" here

        with gr.TabItem("Extractive QA ❔"):
            table = gr.Dataframe(get_model_names('question_answering.csv')) # No datatype="markdown" here

        with gr.TabItem("All Tasks 💡"):
            table = gr.Dataframe(get_all_model_names()) # No datatype="markdown" here

    with gr.Accordion("📙 Citation", open=False):
        citation_button = gr.Textbox(
            value=CITATION_BUTTON_TEXT,
            label=CITATION_BUTTON_LABEL,
            elem_id="citation-button",
            lines=10,
            show_copy_button=True,
        )
    gr.Markdown(
        """Last updated: February 2025"""
    )

demo.launch()