app.py

import gradio as gr
import pandas as pd
import os
import zipfile
import base64

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):
    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; font-size:1.5em;">{"★" * score_int}</span>'

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

def extract_link_text(html_link):
    """Extracts the inner text from an HTML link."""
    start = html_link.find('>') + 1
    end = html_link.rfind('</a>')
    if start > 0 and end > start:
        return html_link[start:end]
    else:
        return html_link

def generate_html_table_from_df(df):
    """
    Given a dataframe with a numeric energy column (gpu_energy_numeric),
    generate an HTML table with three columns:
      - Model (the link, with a fixed width based on the longest model name)
      - GPU Energy (Wh) plus a horizontal bar whose width is proportional
        to the energy value relative to the maximum in the table.
      - Score (displayed as stars)
    """
    # Compute a static width (in pixels) for the Model column based on the longest model name.
    if not df.empty:
        max_length = max(len(extract_link_text(link)) for link in df['Model'])
    else:
        max_length = 10
    # Multiply by an estimated average character width (10 pixels) and add some extra padding.
    static_width = max_length * 10 + 16

    max_energy = df['gpu_energy_numeric'].max() if not df.empty else 1
    color_map = {"1": "black", "2": "black", "3": "black", "4": "black", "5": "black"}
    html = '<table style="width:100%; border-collapse: collapse; font-family: Inter, sans-serif;">'
    # Keep only one header (the one with hover text)
    html += '<thead><tr style="background-color: #f2f2f2;">'
    html += '<th style="text-align: left; padding: 8px;" title="Model name with link to Hugging Face">Model</th>'
    html += '<th style="text-align: left; padding: 8px;" title="GPU energy consumed in Watt-hours for 1,000 queries">GPU Energy (Wh)</th>'
    html += '<th style="text-align: left; padding: 8px;" title="5 is most efficient, 1 is least. Relative energy efficiency score relative to other models in task/class at the time of leaderboard launch">Score</th>'
    html += '</tr></thead>'
    html += '<tbody>'
    for _, row in df.iterrows():
        energy_numeric = row['gpu_energy_numeric']
        energy_str = f"{energy_numeric:.2f}"
        # Compute the relative width (as a percentage)
        bar_width = (energy_numeric / max_energy) * 100
        score_val = row['energy_score']
        bar_color = color_map.get(str(score_val), "gray")
        html += '<tr>'
        html += f'<td style="padding: 8px; width: {static_width}px;">{row["Model"]}</td>'
        html += (
            f'<td style="padding: 8px;">{energy_str}<br>'
            f'<div style="background-color: {bar_color}; width: {bar_width:.1f}%; height: 10px;"></div></td>'
        )
        html += f'<td style="padding: 8px;">{row["Score"]}</td>'
        html += '</tr>'
    html += '</tbody></table>'
    return html

# --- Function to zip all CSV files ---
def zip_csv_files():
    data_dir = "data/energy"
    zip_filename = "data.zip"
    with zipfile.ZipFile(zip_filename, "w", zipfile.ZIP_DEFLATED) as zipf:
        for filename in os.listdir(data_dir):
            if filename.endswith(".csv"):
                filepath = os.path.join(data_dir, filename)
                zipf.write(filepath, arcname=filename)
    return zip_filename

def get_zip_data_link():
    """Creates a data URI download link for the ZIP file."""
    zip_filename = zip_csv_files()
    with open(zip_filename, "rb") as f:
        data = f.read()
    b64 = base64.b64encode(data).decode()
    href = f'<a href="data:application/zip;base64,{b64}" download="data.zip" style="margin: 0 15px; text-decoration: none; font-weight: bold; font-size: 1.1em;">Download Data</a>'
    return href

# --- Modified functions to include a sort_order parameter ---
def get_model_names_html(task, sort_order="Low to High"):
    df = pd.read_csv('data/energy/' + task)
    if df.columns[0].startswith("Unnamed:"):
        df = df.iloc[:, 1:]
    df['energy_score'] = df['energy_score'].astype(int)
    # Convert kWh to Wh:
    df['gpu_energy_numeric'] = pd.to_numeric(df['total_gpu_energy'], errors='raise') * 1000
    df['Model'] = df['model'].apply(make_link)
    df['Score'] = df['energy_score'].apply(format_stars)
    ascending = True if sort_order == "Low to High" else False
    df = df.sort_values(by='gpu_energy_numeric', ascending=ascending)
    return generate_html_table_from_df(df)

def get_all_model_names_html(sort_order="Low to High"):
    all_df = pd.DataFrame()
    for task in tasks:
        df = pd.read_csv('data/energy/' + task)
        if df.columns[0].startswith("Unnamed:"):
            df = df.iloc[:, 1:]
        df['energy_score'] = df['energy_score'].astype(int)
        df['gpu_energy_numeric'] = pd.to_numeric(df['total_gpu_energy'], errors='raise') * 1000
        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'])
    ascending = True if sort_order == "Low to High" else False
    all_df = all_df.sort_values(by='gpu_energy_numeric', ascending=ascending)
    return generate_html_table_from_df(all_df)

def get_text_generation_model_names_html(model_class, sort_order="Low to High"):
    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['energy_score'] = df['energy_score'].astype(int)
    df['gpu_energy_numeric'] = pd.to_numeric(df['total_gpu_energy'], errors='raise') * 1000
    df['Model'] = df['model'].apply(make_link)
    df['Score'] = df['energy_score'].apply(format_stars)
    ascending = True if sort_order == "Low to High" else False
    df = df.sort_values(by='gpu_energy_numeric', ascending=ascending)
    return generate_html_table_from_df(df)

# --- Update functions for dropdown changes ---

# For Text Generation, two dropdowns: model class and sort order.
def update_text_generation(selected_display, sort_order):
    mapping = {
        "A (Single Consumer GPU) <20B parameters": "A",
        "B (Single Cloud GPU) 20-66B parameters": "B",
        "C (Multiple Cloud GPUs) >66B parameters": "C"
    }
    model_class = mapping.get(selected_display, "A")
    return get_text_generation_model_names_html(model_class, sort_order)

# For the other tabs, each update function simply takes the sort_order.
def update_image_generation(sort_order):
    return get_model_names_html('image_generation.csv', sort_order)

def update_text_classification(sort_order):
    return get_model_names_html('text_classification.csv', sort_order)

def update_image_classification(sort_order):
    return get_model_names_html('image_classification.csv', sort_order)

def update_image_captioning(sort_order):
    return get_model_names_html('image_captioning.csv', sort_order)

def update_summarization(sort_order):
    return get_model_names_html('summarization.csv', sort_order)

def update_asr(sort_order):
    return get_model_names_html('asr.csv', sort_order)

def update_object_detection(sort_order):
    return get_model_names_html('object_detection.csv', sort_order)

def update_sentence_similarity(sort_order):
    return get_model_names_html('sentence_similarity.csv', sort_order)

def update_extractive_qa(sort_order):
    return get_model_names_html('question_answering.csv', sort_order)

def update_all_tasks(sort_order):
    return get_all_model_names_html(sort_order)

# --- Build the Gradio Interface ---

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;
}
""")

with demo:
    # Replace title with a centered logo and a centered subtitle.
    gr.HTML('<div style="text-align: center;"><img src="https://huggingface.co/spaces/bgamazay/Leaderboard_test/resolve/main/logo.png" alt="Logo"></div>')
    gr.Markdown('<p style="text-align: center;">Welcome to the leaderboard for the <a href="https://huggingface.co/AIEnergyScore">AI Energy Score Project!</a> — Select different tasks to see scored models.</p>')
    
    # Header links (using a row of components, including a Download Data link)
    with gr.Row():
        submission_link = gr.HTML('<a href="https://huggingface.co/spaces/AIEnergyScore/submission_portal" style="margin: 0 10px; text-decoration: none; font-weight: bold; font-size: 1.1em;">Submission Portal</a>')
        label_link = gr.HTML('<a href="https://huggingface.co/spaces/AIEnergyScore/Label" style="margin: 0 10px; text-decoration: none; font-weight: bold; font-size: 1.1em;">Label Generator</a>')
        faq_link = gr.HTML('<a href="https://huggingface.github.io/AIEnergyScore/#faq" style="margin: 0 10px; text-decoration: none; font-weight: bold; font-size: 1.1em;">FAQ</a>')
        documentation_link = gr.HTML('<a href="https://huggingface.github.io/AIEnergyScore/#documentation" style="margin: 0 10px; text-decoration: none; font-weight: bold; font-size: 1.1em;">Documentation</a>')
        download_link = gr.HTML(get_zip_data_link())
        community_link = gr.HTML('<a href="https://huggingface.co/spaces/AIEnergyScore/README/discussions" style="margin: 0 10px; text-decoration: none; font-weight: bold; font-size: 1.1em;">Community</a>')
    
    with gr.Tabs():
        # --- Text Generation Tab ---
        with gr.TabItem("Text Generation 💬"):
            with gr.Row():
                model_class_options = [
                    "A (Single Consumer GPU) <20B parameters",
                    "B (Single Cloud GPU) 20-66B parameters",
                    "C (Multiple Cloud GPUs) >66B parameters"
                ]
                model_class_dropdown = gr.Dropdown(
                    choices=model_class_options,
                    label="Select Model Class",
                    value=model_class_options[0]
                )
                sort_dropdown_tg = gr.Dropdown(
                    choices=["Low to High", "High to Low"],
                    label="Sort",
                    value="Low to High"
                )
            tg_table = gr.HTML(get_text_generation_model_names_html("A", "Low to High"))
            # When either dropdown changes, update the table.
            model_class_dropdown.change(fn=update_text_generation, inputs=[model_class_dropdown, sort_dropdown_tg], outputs=tg_table)
            sort_dropdown_tg.change(fn=update_text_generation, inputs=[model_class_dropdown, sort_dropdown_tg], outputs=tg_table)
        
        # --- Image Generation Tab ---
        with gr.TabItem("Image Generation 📷"):
            sort_dropdown_img = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            img_table = gr.HTML(get_model_names_html('image_generation.csv', "Low to High"))
            sort_dropdown_img.change(fn=update_image_generation, inputs=sort_dropdown_img, outputs=img_table)
        
        # --- Text Classification Tab ---
        with gr.TabItem("Text Classification 🎭"):
            sort_dropdown_tc = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            tc_table = gr.HTML(get_model_names_html('text_classification.csv', "Low to High"))
            sort_dropdown_tc.change(fn=update_text_classification, inputs=sort_dropdown_tc, outputs=tc_table)
        
        # --- Image Classification Tab ---
        with gr.TabItem("Image Classification 🖼️"):
            sort_dropdown_ic = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            ic_table = gr.HTML(get_model_names_html('image_classification.csv', "Low to High"))
            sort_dropdown_ic.change(fn=update_image_classification, inputs=sort_dropdown_ic, outputs=ic_table)
        
        # --- Image Captioning Tab ---
        with gr.TabItem("Image Captioning 📝"):
            sort_dropdown_icap = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            icap_table = gr.HTML(get_model_names_html('image_captioning.csv', "Low to High"))
            sort_dropdown_icap.change(fn=update_image_captioning, inputs=sort_dropdown_icap, outputs=icap_table)
        
        # --- Summarization Tab ---
        with gr.TabItem("Summarization 📃"):
            sort_dropdown_sum = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            sum_table = gr.HTML(get_model_names_html('summarization.csv', "Low to High"))
            sort_dropdown_sum.change(fn=update_summarization, inputs=sort_dropdown_sum, outputs=sum_table)
        
        # --- Automatic Speech Recognition Tab ---
        with gr.TabItem("Automatic Speech Recognition 💬"):
            sort_dropdown_asr = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            asr_table = gr.HTML(get_model_names_html('asr.csv', "Low to High"))
            sort_dropdown_asr.change(fn=update_asr, inputs=sort_dropdown_asr, outputs=asr_table)
        
        # --- Object Detection Tab ---
        with gr.TabItem("Object Detection 🚘"):
            sort_dropdown_od = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            od_table = gr.HTML(get_model_names_html('object_detection.csv', "Low to High"))
            sort_dropdown_od.change(fn=update_object_detection, inputs=sort_dropdown_od, outputs=od_table)
        
        # --- Sentence Similarity Tab ---
        with gr.TabItem("Sentence Similarity 📚"):
            sort_dropdown_ss = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            ss_table = gr.HTML(get_model_names_html('sentence_similarity.csv', "Low to High"))
            sort_dropdown_ss.change(fn=update_sentence_similarity, inputs=sort_dropdown_ss, outputs=ss_table)
        
        # --- Extractive QA Tab ---
        with gr.TabItem("Extractive QA ❔"):
            sort_dropdown_qa = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            qa_table = gr.HTML(get_model_names_html('question_answering.csv', "Low to High"))
            sort_dropdown_qa.change(fn=update_extractive_qa, inputs=sort_dropdown_qa, outputs=qa_table)
        
        # --- All Tasks Tab ---
        with gr.TabItem("All Tasks 💡"):
            sort_dropdown_all = gr.Dropdown(
                choices=["Low to High", "High to Low"],
                label="Sort",
                value="Low to High"
            )
            all_table = gr.HTML(get_all_model_names_html("Low to High"))
            sort_dropdown_all.change(fn=update_all_tasks, inputs=sort_dropdown_all, outputs=all_table)
    
    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()