app.py

import gradio as gr
import pandas as pd
import plotly.express as px
import requests
import re
import os
import glob


# Download the main results file
def download_main_results():
    url = "https://github.com/huggingface/pytorch-image-models/raw/main/results/results-imagenet.csv"
    if not os.path.exists("results-imagenet.csv"):
        response = requests.get(url)
        with open("results-imagenet.csv", "wb") as f:
            f.write(response.content)


def download_github_csvs_api(
    repo="huggingface/pytorch-image-models",
    folder="results",
    filename_pattern=r"benchmark-.*\.csv",
    output_dir="benchmarks",
):
    """Download benchmark CSV files from GitHub API."""
    api_url = f"https://api.github.com/repos/{repo}/contents/{folder}"
    r = requests.get(api_url)
    if r.status_code != 200:
        return []

    files = r.json()
    matched_files = [f["name"] for f in files if re.match(filename_pattern, f["name"])]

    if not matched_files:
        return []

    raw_base = f"https://raw.githubusercontent.com/{repo}/main/{folder}/"
    os.makedirs(output_dir, exist_ok=True)

    for fname in matched_files:
        raw_url = raw_base + fname
        out_path = os.path.join(output_dir, fname)

        if not os.path.exists(out_path):  # Only download if not exists
            resp = requests.get(raw_url)
            if resp.ok:
                with open(out_path, "wb") as f:
                    f.write(resp.content)

    return matched_files


def load_main_data():
    """Load the main ImageNet results."""
    download_main_results()
    df_results = pd.read_csv("results-imagenet.csv")
    df_results["model_org"] = df_results["model"]
    df_results["model"] = df_results["model"].str.split(".").str[0]
    return df_results


def get_data(benchmark_file, df_results):
    """Process benchmark data and merge with main results."""
    pattern = (
        r"^(?:"
        r"eva|"
        r"maxx?vit(?:v2)?|"
        r"coatnet|coatnext|"
        r"convnext(?:v2)?|"
        r"beit(?:v2)?|"
        r"efficient(?:net(?:v2)?|former(?:v2)?|vit)|"
        r"regnet[xyvz]?|"
        r"levit|"
        r"mobilenet(?:v\d*)?|"
        r"vitd?|"
        r"swin(?:v2)?"
        r")$"
    )

    if not os.path.exists(benchmark_file):
        return pd.DataFrame()

    df = pd.read_csv(benchmark_file).merge(df_results, on="model")
    df["secs"] = 1.0 / df["infer_samples_per_sec"]
    df["family"] = df.model.str.extract("^([a-z]+?(?:v2)?)(?:\d|_|$)")
    df = df[~df.model.str.endswith("gn")]
    df.loc[df.model.str.contains("resnet.*d"), "family"] = (
        df.loc[df.model.str.contains("resnet.*d"), "family"] + "d"
    )
    return df[df.family.str.contains(pattern)]


def create_plot(benchmark_file, x_axis, y_axis, selected_families, log_x, log_y):
    """Create the scatter plot based on user selections."""
    df_results = load_main_data()
    df = get_data(benchmark_file, df_results)

    if df.empty:
        return None

    # Filter by selected families
    if selected_families:
        df = df[df["family"].isin(selected_families)]

    if df.empty:
        return None

    # Create the plot
    fig = px.scatter(
        df,
        width=1000,
        height=800,
        x=x_axis,
        y=y_axis,
        size=df['infer_img_size']**2,
        log_x=log_x,
        log_y=log_y,
        color="family",
        hover_name="model_org",
        hover_data=["infer_samples_per_sec", "infer_img_size"],
        title=f"Model Performance: {y_axis} vs {x_axis}",
    )

    return fig


def setup_interface():
    """Set up the Gradio interface."""
    # Download benchmark files
    downloaded_files = download_github_csvs_api()

    # Get available benchmark files
    benchmark_files = glob.glob("benchmarks/benchmark-*.csv")
    if not benchmark_files:
        benchmark_files = ["No benchmark files found"]

    # Load sample data to get families and columns
    df_results = load_main_data()

    # Relevant columns for plotting
    plot_columns = [
        "top1",
        "top5",
        "infer_samples_per_sec",
        "secs",
        "param_count_x",
        "infer_img_size",
    ]

    # Get families from a sample file (if available)
    families = []
    if benchmark_files and benchmark_files[0] != "No benchmark files found":
        sample_df = get_data(benchmark_files[0], df_results)
        if not sample_df.empty:
            families = sorted(sample_df["family"].unique().tolist())

    return benchmark_files, plot_columns, families


# Initialize the interface
benchmark_files, plot_columns, families = setup_interface()

# Create the Gradio interface
with gr.Blocks(title="Image Model Performance Analysis") as demo:
    gr.Markdown("# Image Model Performance Analysis")
    gr.Markdown(
        "Analyze and visualize performance metrics of different image models based on benchmark data."
    )

    with gr.Row():
        with gr.Column(scale=1):

            # Set preferred default file
            preferred_file = (
                "benchmarks/benchmark-infer-amp-nhwc-pt240-cu124-rtx3090.csv"
            )
            default_file = (
                preferred_file
                if preferred_file in benchmark_files
                else (benchmark_files[0] if benchmark_files else None)
            )

            benchmark_dropdown = gr.Dropdown(
                choices=benchmark_files,
                value=default_file,
                label="Select Benchmark File",
            )

            x_axis_radio = gr.Radio(choices=plot_columns, value="secs", label="X-axis")

            y_axis_radio = gr.Radio(choices=plot_columns, value="top1", label="Y-axis")

            family_checkboxes = gr.CheckboxGroup(
                choices=families, value=families, label="Select Model Families"
            )

            log_x_checkbox = gr.Checkbox(value=True, label="Log scale X-axis")

            log_y_checkbox = gr.Checkbox(value=False, label="Log scale Y-axis")

            update_button = gr.Button("Update Plot", variant="primary")

        with gr.Column(scale=2):
            plot_output = gr.Plot()
            gr.Markdown("The benchmark data comes from the [pytorch-image-models](https://github.com/huggingface/pytorch-image-models) repository by [Ross Wightman](https://huggingface.co/rwightman).")
            gr.Markdown("Based on the original notebook by [Jeremy Howard](https://huggingface.co/jph00).")
            gr.Markdown("Read more about the project on my blog [dronelab.dev](https://dronelab.dev/posts/which-image-models-are-best-updated/).")
    

    # Update plot when button is clicked
    update_button.click(
        fn=create_plot,
        inputs=[
            benchmark_dropdown,
            x_axis_radio,
            y_axis_radio,
            family_checkboxes,
            log_x_checkbox,
            log_y_checkbox,
        ],
        outputs=plot_output,
    )

    # Auto-update when benchmark file changes
    def update_families(benchmark_file):
        if not benchmark_file or benchmark_file == "No benchmark files found":
            return gr.CheckboxGroup(choices=[], value=[])

        df_results = load_main_data()
        df = get_data(benchmark_file, df_results)
        if df.empty:
            return gr.CheckboxGroup(choices=[], value=[])

        new_families = sorted(df["family"].unique().tolist())
        return gr.CheckboxGroup(choices=new_families, value=new_families)

    benchmark_dropdown.change(
        fn=update_families, inputs=benchmark_dropdown, outputs=family_checkboxes
    )

    # Load initial plot
    demo.load(
        fn=create_plot,
        inputs=[
            benchmark_dropdown,
            x_axis_radio,
            y_axis_radio,
            family_checkboxes,
            log_x_checkbox,
            log_y_checkbox,
        ],
        outputs=plot_output,
    )

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