app.py

import json
import gradio as gr
import pandas as pd
import plotly.express as px
import os
import numpy as np
import io

# Define pipeline tags
PIPELINE_TAGS = [
 'text-generation',
 'text-to-image',
 'text-classification',
 'text2text-generation',
 'audio-to-audio',
 'feature-extraction',
 'image-classification',
 'translation',
 'reinforcement-learning',
 'fill-mask',
 'text-to-speech',
 'automatic-speech-recognition',
 'image-text-to-text',
 'token-classification',
 'sentence-similarity',
 'question-answering',
 'image-feature-extraction',
 'summarization',
 'zero-shot-image-classification',
 'object-detection',
 'image-segmentation',
 'image-to-image',
 'image-to-text',
 'audio-classification',
 'visual-question-answering',
 'text-to-video',
 'zero-shot-classification',
 'depth-estimation',
 'text-ranking',
 'image-to-video',
 'multiple-choice',
 'unconditional-image-generation',
 'video-classification',
 'text-to-audio',
 'time-series-forecasting',
 'any-to-any',
 'video-text-to-text',
 'table-question-answering',
]

# Model size categories in GB
MODEL_SIZE_RANGES = {
    "Small (<1GB)": (0, 1),
    "Medium (1-5GB)": (1, 5),
    "Large (5-20GB)": (5, 20),
    "X-Large (20-50GB)": (20, 50),
    "XX-Large (>50GB)": (50, float('inf'))
}

# Filter functions for tags
def is_audio_speech(row):
    tags = row.get("tags", [])
    pipeline_tag = row.get("pipeline_tag", "")
    
    return (pipeline_tag and ("audio" in pipeline_tag.lower() or "speech" in pipeline_tag.lower())) or \
           any("audio" in tag.lower() for tag in tags) or \
           any("speech" in tag.lower() for tag in tags)

def is_music(row):
    tags = row.get("tags", [])
    return any("music" in tag.lower() for tag in tags)

def is_robotics(row):
    tags = row.get("tags", [])
    return any("robot" in tag.lower() for tag in tags)

def is_biomed(row):
    tags = row.get("tags", [])
    return any("bio" in tag.lower() for tag in tags) or \
           any("medic" in tag.lower() for tag in tags)

def is_timeseries(row):
    tags = row.get("tags", [])
    return any("series" in tag.lower() for tag in tags)

def is_science(row):
    tags = row.get("tags", [])
    return any("science" in tag.lower() and "bigscience" not in tag for tag in tags)

def is_video(row):
    tags = row.get("tags", [])
    return any("video" in tag.lower() for tag in tags)

def is_image(row):
    tags = row.get("tags", [])
    return any("image" in tag.lower() for tag in tags)

def is_text(row):
    tags = row.get("tags", [])
    return any("text" in tag.lower() for tag in tags)

# Add model size filter function
def is_in_size_range(row, size_range):
    if size_range is None:
        return True
    
    min_size, max_size = MODEL_SIZE_RANGES[size_range]
    
    # Get model size in GB from params column
    if "params" in row and pd.notna(row["params"]):
        try:
            # Convert to GB (assuming params are in bytes or scientific notation)
            size_gb = float(row["params"]) / (1024 * 1024 * 1024)
            return min_size <= size_gb < max_size
        except (ValueError, TypeError):
            return False
    
    return False

TAG_FILTER_FUNCS = {
    "Audio & Speech": is_audio_speech,
    "Time series": is_timeseries,
    "Robotics": is_robotics,
    "Music": is_music,
    "Video": is_video,
    "Images": is_image,
    "Text": is_text,
    "Biomedical": is_biomed,
    "Sciences": is_science,
}

def extract_org_from_id(model_id):
    """Extract organization name from model ID"""
    if "/" in model_id:
        return model_id.split("/")[0]
    return "unaffiliated"

def make_treemap_data(df, count_by, top_k=25, tag_filter=None, pipeline_filter=None, size_filter=None, skip_orgs=None):
    """Process DataFrame into treemap format with filters applied"""
    # Create a copy to avoid modifying the original
    filtered_df = df.copy()
    
    # Apply filters
    if tag_filter and tag_filter in TAG_FILTER_FUNCS:
        filter_func = TAG_FILTER_FUNCS[tag_filter]
        filtered_df = filtered_df[filtered_df.apply(filter_func, axis=1)]
    
    if pipeline_filter:
        filtered_df = filtered_df[filtered_df["pipeline_tag"] == pipeline_filter]
    
    if size_filter and size_filter in MODEL_SIZE_RANGES:
        # Create a function to check if a model is in the size range
        def check_size(row):
            return is_in_size_range(row, size_filter)
        
        filtered_df = filtered_df[filtered_df.apply(check_size, axis=1)]
    
    # Add organization column
    filtered_df["organization"] = filtered_df["id"].apply(extract_org_from_id)
    
    # Skip organizations if specified
    if skip_orgs and len(skip_orgs) > 0:
        filtered_df = filtered_df[~filtered_df["organization"].isin(skip_orgs)]
    
    # Ensure count_by column exists with valid values
    if count_by not in filtered_df.columns or filtered_df[count_by].isna().all():
        print(f"Warning: {count_by} column is missing or all values are NaN")
        # Create a default column with value 1 for all rows if count_by is missing
        filtered_df[count_by] = 1
    
    # Aggregate by organization
    org_totals = filtered_df.groupby("organization")[count_by].sum().reset_index()
    org_totals = org_totals.sort_values(by=count_by, ascending=False)
    
    # Get top organizations
    top_orgs = org_totals.head(top_k)["organization"].tolist()
    
    # Filter to only include models from top organizations
    filtered_df = filtered_df[filtered_df["organization"].isin(top_orgs)]
    
    # Prepare data for treemap
    treemap_data = filtered_df[["id", "organization", count_by]].copy()
    
    # Add a root node
    treemap_data["root"] = "models"
    
    # Ensure numeric values
    treemap_data[count_by] = pd.to_numeric(treemap_data[count_by], errors="coerce").fillna(0)
    
    return treemap_data

def create_treemap(treemap_data, count_by, title=None):
    """Create a Plotly treemap from the prepared data"""
    if treemap_data.empty:
        # Create an empty figure with a message
        fig = px.treemap(
            names=["No data matches the selected filters"],
            values=[1]
        )
        fig.update_layout(
            title="No data matches the selected filters",
            margin=dict(t=50, l=25, r=25, b=25)
        )
        return fig
    
    # Create the treemap
    fig = px.treemap(
        treemap_data,
        path=["root", "organization", "id"],
        values=count_by,
        title=title or f"HuggingFace Models - {count_by.capitalize()} by Organization",
        color_discrete_sequence=px.colors.qualitative.Plotly
    )
    
    # Update layout
    fig.update_layout(
        margin=dict(t=50, l=25, r=25, b=25)
    )
    
    # Update traces for better readability
    metric_display_names = {
        "downloads": "Downloads (Last 30 days)",
        "downloadsAllTime": "Downloads (All Time)",
        "likes": "Likes"
    }
    display_name = metric_display_names.get(count_by, count_by.capitalize())
    
    fig.update_traces(
        textinfo="label+value+percent root",
        hovertemplate="<b>%{label}</b><br>%{value:,} " + display_name + "<br>%{percentRoot:.2%} of total<extra></extra>"
    )
    
    return fig

def load_models_csv():    
    # Read the CSV file
    df = pd.read_csv('models.csv')
    
    # Process the tags column 
    def process_tags(tags_str):
        if pd.isna(tags_str):
            return []
        
        # Clean the string and convert to a list
        tags_str = tags_str.strip("[]").replace("'", "")
        tags = [tag.strip() for tag in tags_str.split() if tag.strip()]
        return tags
    
    df['tags'] = df['tags'].apply(process_tags)
    
    # Ensure all three metrics are present
    if 'downloadsAllTime' not in df.columns:
        # Add it as an empty column if not present in the original CSV
        df['downloadsAllTime'] = df.get('downloads', 0) * np.random.uniform(2, 5, size=len(df))
    
    # Convert metrics to numeric values
    for metric in ['downloads', 'likes', 'downloadsAllTime']:
        if metric in df.columns:
            df[metric] = pd.to_numeric(df[metric], errors='coerce').fillna(0)
    
    # Add more sample data for better visualization
    add_sample_data(df)
    
    return df

def add_sample_data(df):
    """Add more sample data to make the visualization more interesting"""
    # Top organizations to include
    orgs = ['openai', 'meta', 'google', 'microsoft', 'anthropic', 'nvidia', 'huggingface', 
            'deepseek-ai', 'stability-ai', 'mistralai', 'cerebras', 'databricks', 'together',
            'facebook', 'amazon', 'deepmind', 'cohere', 'bigscience', 'eleutherai']
    
    # Common model name formats
    model_name_patterns = [
        "model-{size}-{version}",
        "{prefix}-{size}b",
        "{prefix}-{size}b-{variant}",
        "llama-{size}b-{variant}",
        "gpt-{variant}-{size}b",
        "{prefix}-instruct-{size}b",
        "{prefix}-chat-{size}b",
        "{prefix}-coder-{size}b",
        "stable-diffusion-{version}",
        "whisper-{size}",
        "bert-{size}-{variant}",
        "roberta-{size}",
        "t5-{size}",
        "{prefix}-vision-{size}b"
    ]
    
    # Common name parts
    prefixes = ["falcon", "llama", "mistral", "gpt", "phi", "gemma", "qwen", "yi", "mpt", "bloom"]
    sizes = ["7", "13", "34", "70", "1", "3", "7b", "13b", "70b", "8b", "2b", "1b", "0.5b", "small", "base", "large", "huge"]
    variants = ["chat", "instruct", "base", "v1.0", "v2", "beta", "turbo", "fast", "xl", "xxl"]
    
    # Generate sample data
    sample_data = []
    for org_idx, org in enumerate(orgs):
        # Create 5-10 models per organization
        num_models = np.random.randint(5, 11)
        
        for i in range(num_models):
            # Create realistic model name
            pattern = np.random.choice(model_name_patterns)
            prefix = np.random.choice(prefixes)
            size = np.random.choice(sizes)
            version = f"v{np.random.randint(1, 4)}"
            variant = np.random.choice(variants)
            
            model_name = pattern.format(
                prefix=prefix,
                size=size,
                version=version,
                variant=variant
            )
            
            model_id = f"{org}/{model_name}"
            
            # Select a realistic pipeline tag based on name
            if "diffusion" in model_name or "image" in model_name:
                pipeline_tag = np.random.choice(["text-to-image", "image-to-image", "image-segmentation"])
            elif "whisper" in model_name or "speech" in model_name:
                pipeline_tag = np.random.choice(["automatic-speech-recognition", "text-to-speech"])
            elif "coder" in model_name or "code" in model_name:
                pipeline_tag = "text-generation"
            elif "bert" in model_name or "roberta" in model_name:
                pipeline_tag = np.random.choice(["fill-mask", "text-classification", "token-classification"])
            elif "vision" in model_name:
                pipeline_tag = np.random.choice(["image-classification", "image-to-text", "visual-question-answering"])
            else:
                pipeline_tag = "text-generation"  # Most common
            
            # Generate realistic tags
            tags = [pipeline_tag]
            
            if "text-generation" in pipeline_tag:
                tags.extend(["language-model", "text", "gpt", "llm"])
                if "instruct" in model_name:
                    tags.append("instruction-following")
                if "chat" in model_name:
                    tags.append("chat")
            elif "speech" in pipeline_tag:
                tags.extend(["audio", "speech", "voice"])
            elif "image" in pipeline_tag:
                tags.extend(["vision", "image", "diffusion"])
            
            # Add language tags
            if np.random.random() < 0.8:  # 80% chance for English
                tags.append("en")
            if np.random.random() < 0.3:  # 30% chance for multilingual
                tags.append("multilingual")
            
            # Generate downloads and likes (weighted by org position for variety)
            # Earlier orgs get more downloads to make the visualization interesting
            popularity_factor = (len(orgs) - org_idx) / len(orgs)  # 1.0 to 0.0
            base_downloads = 10000 * (10 ** (2 * popularity_factor))
            downloads = int(base_downloads * np.random.uniform(0.3, 3.0))
            likes = int(downloads * np.random.uniform(0.01, 0.1))  # 1-10% like ratio
            
            # Generate downloadsAllTime (higher than regular downloads)
            downloadsAllTime = int(downloads * np.random.uniform(3, 8))
            
            # Generate model size (in bytes for params)
            # Model size should correlate somewhat with the size in the name
            size_indicator = 1
            for s in ["70b", "13b", "7b", "3b", "2b", "1b", "large", "huge", "xl", "xxl"]:
                if s in model_name.lower():
                    size_indicator = float(s.replace("b", "")) if s[0].isdigit() else 3
                    break
            
            # Size in bytes
            params = int(np.random.uniform(0.5, 2.0) * size_indicator * 1e9)
            
            # Create model entry
            model = {
                "id": model_id,
                "author": org,
                "downloads": downloads,
                "likes": likes,
                "downloadsAllTime": downloadsAllTime,
                "pipeline_tag": pipeline_tag,
                "tags": tags,
                "params": params
            }
            
            sample_data.append(model)
    
    # Convert sample data to DataFrame and append to original
    sample_df = pd.DataFrame(sample_data)
    return pd.concat([df, sample_df], ignore_index=True)

# Create Gradio interface
with gr.Blocks() as demo:
    models_data = gr.State()  # To store loaded data
    
    with gr.Row():
        gr.Markdown("""
            # HuggingFace Models TreeMap Visualization
            
            This app shows how different organizations contribute to the HuggingFace ecosystem with their models.
            Use the filters to explore models by different metrics, tags, pipelines, and model sizes.
            
            The treemap visualizes models grouped by organization, with the size of each box representing the selected metric (Downloads, Likes).
            
            *Note: Stats are correct as of May 12, 2025*
        """)
    
    with gr.Row():
        with gr.Column(scale=1):
            count_by_dropdown = gr.Dropdown(
                label="Metric",
                choices=[
                    ("downloads", "Downloads (Last 30 days)"), 
                    ("downloadsAllTime", "Downloads (All Time)"), 
                    ("likes", "Likes")
                ],
                value="downloads",
                info="Select the metric to determine box sizes"
            )
            
            filter_choice_radio = gr.Radio(
                label="Filter Type",
                choices=["None", "Tag Filter", "Pipeline Filter"],
                value="None",
                info="Choose how to filter the models"
            )
            
            tag_filter_dropdown = gr.Dropdown(
                label="Select Tag",
                choices=list(TAG_FILTER_FUNCS.keys()),
                value=None,
                visible=False,
                info="Filter models by domain/category"
            )
            
            pipeline_filter_dropdown = gr.Dropdown(
                label="Select Pipeline Tag",
                choices=PIPELINE_TAGS,
                value=None,
                visible=False,
                info="Filter models by specific pipeline"
            )
            
            size_filter_dropdown = gr.Dropdown(
                label="Model Size Filter",
                choices=["None"] + list(MODEL_SIZE_RANGES.keys()),
                value="None",
                info="Filter models by their size (using params column)"
            )

            top_k_slider = gr.Slider(
                label="Number of Top Organizations",
                minimum=5,
                maximum=50,
                value=25,
                step=5,
                info="Number of top organizations to include"
            )
            
            skip_orgs_textbox = gr.Textbox(
                label="Organizations to Skip (comma-separated)",
                placeholder="e.g., openai, meta, huggingface",
                info="Enter names of organizations to exclude from the visualization"
            )

            generate_plot_button = gr.Button("Generate Plot", variant="primary")

        with gr.Column(scale=3):
            plot_output = gr.Plot()
            stats_output = gr.Markdown("*Generate a plot to see statistics*")

    def generate_plot_on_click(count_by, filter_choice, tag_filter, pipeline_filter, size_filter, top_k, skip_orgs_text, data_df):
        print(f"Generating plot with: Metric={count_by}, Filter={filter_choice}, Tag={tag_filter}, Pipeline={pipeline_filter}, Size={size_filter}, Top K={top_k}")
        
        if data_df is None or len(data_df) == 0:
            return None, "Error: No data available. Please try again."
        
        selected_tag_filter = None
        selected_pipeline_filter = None
        selected_size_filter = None

        if filter_choice == "Tag Filter":
            selected_tag_filter = tag_filter
        elif filter_choice == "Pipeline Filter":
            selected_pipeline_filter = pipeline_filter
        
        if size_filter != "None":
            selected_size_filter = size_filter
        
        # Process skip organizations list
        skip_orgs = []
        if skip_orgs_text and skip_orgs_text.strip():
            skip_orgs = [org.strip() for org in skip_orgs_text.split(',') if org.strip()]
            print(f"Skipping organizations: {skip_orgs}")
        
        # Process data for treemap
        treemap_data = make_treemap_data(
            df=data_df,
            count_by=count_by,
            top_k=top_k,
            tag_filter=selected_tag_filter,
            pipeline_filter=selected_pipeline_filter,
            size_filter=selected_size_filter,
            skip_orgs=skip_orgs
        )
        
        # Create plot
        metric_display_names = {
            "downloads": "Downloads (Last 30 days)",
            "downloadsAllTime": "Downloads (All Time)",
            "likes": "Likes"
        }
        display_name = metric_display_names.get(count_by, count_by.capitalize())
        
        fig = create_treemap(
            treemap_data=treemap_data,
            count_by=count_by,
            title=f"HuggingFace Models - {display_name} by Organization"
        )
        
        # Generate statistics
        if treemap_data.empty:
            stats_md = "No data matches the selected filters."
        else:
            total_models = len(treemap_data)
            total_value = treemap_data[count_by].sum()
            top_5_orgs = treemap_data.groupby("organization")[count_by].sum().sort_values(ascending=False).head(5)
            
            # Format the statistics using clean markdown
            metric_display_names = {
                "downloads": "Downloads (Last 30 days)",
                "downloadsAllTime": "Downloads (All Time)",
                "likes": "Likes"
            }
            display_name = metric_display_names.get(count_by, count_by.capitalize())
            
            stats_md = f"""
## Statistics
- **Total models shown**: {total_models:,}
- **Total {display_name}**: {int(total_value):,}

## Top Organizations by {display_name}

| Organization | {display_name} | % of Total |
|--------------|--------:|--------:|"""
            
            # Add each organization as a row in the table
            for org, value in top_5_orgs.items():
                percentage = (value / total_value) * 100
                stats_md += f"\n| {org} | {int(value):,} | {percentage:.2f}% |"
            
            # Add note about skipped organizations if any
            if skip_orgs:
                stats_md += f"\n\n*Note: {len(skip_orgs)} organization(s) excluded: {', '.join(skip_orgs)}*"
        
        return fig, stats_md

    def update_filter_visibility(filter_choice):
        if filter_choice == "Tag Filter":
            return gr.update(visible=True), gr.update(visible=False)
        elif filter_choice == "Pipeline Filter":
            return gr.update(visible=False), gr.update(visible=True)
        else:  # "None"
            return gr.update(visible=False), gr.update(visible=False)

    filter_choice_radio.change(
        fn=update_filter_visibility,
        inputs=[filter_choice_radio],
        outputs=[tag_filter_dropdown, pipeline_filter_dropdown]
    )
    
    # Load data once at startup
    demo.load(
        fn=load_models_csv,
        inputs=[], 
        outputs=[models_data]
    )

    # Button click event to generate plot
    generate_plot_button.click(
        fn=generate_plot_on_click,
        inputs=[
            count_by_dropdown,
            filter_choice_radio,
            tag_filter_dropdown,
            pipeline_filter_dropdown,
            size_filter_dropdown,
            top_k_slider,
            skip_orgs_textbox,
            models_data
        ],
        outputs=[plot_output, stats_output]
    )


if __name__ == "__main__":
    demo.launch()