app.py

import gradio as gr
import pandas as pd
import plotly.express as px

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'
]

def format_stars(score):
    """
    Convert the energy_score (assumed to be an integer from 1 to 5)
    into that many star characters wrapped in a span styled with color #3fa45bff
    and with a font size increased to 2em.
    """
    try:
        score_int = int(score)
    except Exception:
        score_int = 0
    return f'<span style="color: #3fa45bff; font-size:2em;">{"★" * score_int}</span>'

def make_link(mname):
    """
    Create a markdown link for the model.
    For example, if mname is "org/model", display "model" and link to its HF page.
    """
    parts = str(mname).split('/')
    display_name = parts[1] if len(parts) > 1 else mname
    return f'[{display_name}](https://huggingface.co/{mname})'

def get_plots(task):
    """
    Read the energy CSV for a given task and return a Plotly scatter plot.
    Now the x-axis is the numeric energy (GPU Energy (Wh)) and
    the y-axis displays the model name.
    """
    df = pd.read_csv('data/energy/' + task)
    df['energy_score'] = df['energy_score'].astype(int)
    # Do not multiply by 1000; simply round to 4 decimals
    df['GPU Energy (Wh)'] = df['total_gpu_energy'].round(4)
    
    # Define a 5-level color mapping: 1 = red, 5 = green.
    color_map = {
        1: "red",
        2: "orange",
        3: "yellow",
        4: "lightgreen",
        5: "green"
    }
    
    # Create a horizontal scatter plot: x is the energy, y is the model.
    fig = px.scatter(
        df,
        x="GPU Energy (Wh)",
        y="model",
        custom_data=['energy_score'],
        height=500,
        width=800,
        color="energy_score",
        color_discrete_map=color_map
    )
    fig.update_traces(
        hovertemplate="<br>".join([
            "Model: %{y}",
            "GPU Energy (Wh): %{x}",
            "Energy Score: %{customdata[0]}"
        ])
    )
    fig.update_layout(xaxis_title="GPU Energy (Wh)", yaxis_title="Model")
    return fig

def get_all_plots():
    """
    Combine data from all tasks and return a scatter plot.
    Duplicate models are dropped.
    """
    all_df = pd.DataFrame()
    for task in tasks:
        df = pd.read_csv('data/energy/' + task)
        df['energy_score'] = df['energy_score'].astype(int)
        df['GPU Energy (Wh)'] = df['total_gpu_energy'].round(4)
        all_df = pd.concat([all_df, df], ignore_index=True)
    all_df = all_df.drop_duplicates(subset=['model'])
    
    color_map = {
        1: "red",
        2: "orange",
        3: "yellow",
        4: "lightgreen",
        5: "green"
    }
    fig = px.scatter(
        all_df,
        x="GPU Energy (Wh)",
        y="model",
        custom_data=['energy_score'],
        height=500,
        width=800,
        color="energy_score",
        color_discrete_map=color_map
    )
    fig.update_traces(
        hovertemplate="<br>".join([
            "Model: %{y}",
            "GPU Energy (Wh): %{x}",
            "Energy Score: %{customdata[0]}"
        ])
    )
    fig.update_layout(xaxis_title="GPU Energy (Wh)", yaxis_title="Model")
    return fig

def get_model_names(task):
    """
    For a given task, load the energy CSV and return a dataframe with the following columns:
      - Model (a markdown link)
      - GPU Energy (Wh)
      - Score (a star rating based on energy_score)
    For text_generation.csv only, also add the "Class" column from the CSV.
    The final order is: Model, GPU Energy (Wh), Score, [Class].
    """
    df = pd.read_csv('data/energy/' + task)
    df['energy_score'] = df['energy_score'].astype(int)
    df['GPU Energy (Wh)'] = df['total_gpu_energy'].round(4)
    df['Model'] = df['model'].apply(make_link)
    df['Score'] = df['energy_score'].apply(format_stars)
    
    # If this CSV contains a "class" column (e.g., for Text Generation), add it.
    if 'class' in df.columns:
        df['Class'] = df['class']
        df = df[['Model', 'GPU Energy (Wh)', 'Score', 'Class']]
    else:
        df = df[['Model', 'GPU Energy (Wh)', 'Score']]
        
    df = df.sort_values(by='GPU Energy (Wh)')
    return df

def get_all_model_names():
    """
    Combine data from all tasks and return a leaderboard table with:
      - Model, GPU Energy (Wh), Score
    Duplicate models are dropped.
    """
    all_df = pd.DataFrame()
    for task in tasks:
        df = pd.read_csv('data/energy/' + task)
        df['energy_score'] = df['energy_score'].astype(int)
        df['GPU Energy (Wh)'] = df['total_gpu_energy'].round(4)
        df['Model'] = df['model'].apply(make_link)
        df['Score'] = df['energy_score'].apply(format_stars)
        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='GPU Energy (Wh)')
    return all_df[['Model', 'GPU Energy (Wh)', 'Score']]

# Build the Gradio interface.
demo = gr.Blocks()

with demo:
    gr.Markdown(
        """# AI Energy Score Leaderboard
### Welcome to the leaderboard for the [AI Energy Score Project!](https://huggingface.co/AIEnergyScore)
Click through the tasks below to see how different models measure up in terms of energy efficiency."""
    )
    gr.Markdown(
        """Test your own models via the [submission portal](https://huggingface.co/spaces/AIEnergyScore/submission_portal)"""
    )
    
    with gr.Tabs():
        with gr.TabItem("Text Generation 💬"):
            with gr.Row():
                with gr.Column(scale=1.3):
                    plot = gr.Plot(get_plots('text_generation.csv'))
                with gr.Column(scale=1):
                    # For text generation, the CSV is assumed to have a "class" column.
                    table = gr.Dataframe(get_model_names('text_generation.csv'), datatype="markdown")
                    
        with gr.TabItem("Image Generation 📷"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('image_generation.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('image_generation.csv'), datatype="markdown")
                    
        with gr.TabItem("Text Classification 🎭"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('text_classification.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('text_classification.csv'), datatype="markdown")
                    
        with gr.TabItem("Image Classification 🖼️"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('image_classification.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('image_classification.csv'), datatype="markdown")
                    
        with gr.TabItem("Image Captioning 📝"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('image_captioning.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('image_captioning.csv'), datatype="markdown")
                    
        with gr.TabItem("Summarization 📃"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('summarization.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('summarization.csv'), datatype="markdown")
                    
        with gr.TabItem("Automatic Speech Recognition 💬"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('asr.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('asr.csv'), datatype="markdown")
                    
        with gr.TabItem("Object Detection 🚘"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('object_detection.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('object_detection.csv'), datatype="markdown")
                    
        with gr.TabItem("Sentence Similarity 📚"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('sentence_similarity.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('sentence_similarity.csv'), datatype="markdown")
                    
        with gr.TabItem("Extractive QA ❔"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_plots('question_answering.csv'))
                with gr.Column():
                    table = gr.Dataframe(get_model_names('question_answering.csv'), datatype="markdown")
                    
        with gr.TabItem("All Tasks 💡"):
            with gr.Row():
                with gr.Column():
                    plot = gr.Plot(get_all_plots)
                with gr.Column():
                    table = gr.Dataframe(get_all_model_names, datatype="markdown")
                    
    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()