import gradio as gr
import pandas as pd
from fuzzywuzzy import fuzz

def load_leaderboard():
    imagenet_df = pd.read_csv('https://raw.githubusercontent.com/huggingface/pytorch-image-models/main/results/results-imagenet.csv')
    imagenet_real_df = pd.read_csv('https://raw.githubusercontent.com/huggingface/pytorch-image-models/main/results/results-imagenet-real.csv')
    imagenetv2_df = pd.read_csv('https://raw.githubusercontent.com/huggingface/pytorch-image-models/main/results/results-imagenetv2-matched-frequency.csv')
    sketch_df = pd.read_csv('https://raw.githubusercontent.com/huggingface/pytorch-image-models/main/results/results-sketch.csv')
    imagenet_a_df = pd.read_csv('https://raw.githubusercontent.com/huggingface/pytorch-image-models/main/results/results-imagenet-a.csv')
    imagenet_r_df = pd.read_csv('https://raw.githubusercontent.com/huggingface/pytorch-image-models/main/results/results-imagenet-r.csv')

    # columns to remove from each dataframe
    remove_column_names = ["top1_err", "top5_err", "top1_diff", "top5_diff", "rank_diff"]
    for remove_column_name in remove_column_names:
        if remove_column_name in imagenet_df.columns:
            imagenet_df = imagenet_df.drop(columns=remove_column_name)
        if remove_column_name in imagenet_real_df.columns:
            imagenet_real_df = imagenet_real_df.drop(columns=remove_column_name)
        if remove_column_name in imagenetv2_df.columns:
            imagenetv2_df = imagenetv2_df.drop(columns=remove_column_name)
        if remove_column_name in sketch_df.columns:
            sketch_df = sketch_df.drop(columns=remove_column_name)
        if remove_column_name in imagenet_a_df.columns:
            imagenet_a_df = imagenet_a_df.drop(columns=remove_column_name)
        if remove_column_name in imagenet_r_df.columns:
            imagenet_r_df = imagenet_r_df.drop(columns=remove_column_name)

    # Rename top1 and top5 columns to the name of the dataframe+top1/top5
    imagenet_df = imagenet_df.rename(columns={"top1": "imagenet_top1", "top5": "imagenet_top5"})
    imagenet_real_df = imagenet_real_df.rename(columns={"top1": "imagenet_real_top1", "top5": "imagenet_real_top5"})
    imagenetv2_df = imagenetv2_df.rename(columns={"top1": "imagenetv2_top1", "top5": "imagenetv2_top5"})
    sketch_df = sketch_df.rename(columns={"top1": "sketch_top1", "top5": "sketch_top5"})
    imagenet_a_df = imagenet_a_df.rename(columns={"top1": "imagenet_a_top1", "top5": "imagenet_a_top5"})
    imagenet_r_df = imagenet_r_df.rename(columns={"top1": "imagenet_r_top1", "top5": "imagenet_r_top5"})

    # Merge all dataframes
    result = pd.merge(imagenet_df, imagenet_real_df, on=['model', 'param_count', 'img_size', 'crop_pct', 'interpolation'], how='outer')
    result = pd.merge(result, imagenetv2_df, on=['model', 'param_count', 'img_size', 'crop_pct', 'interpolation'], how='outer')
    result = pd.merge(result, sketch_df, on=['model', 'param_count', 'img_size', 'crop_pct', 'interpolation'], how='outer')
    result = pd.merge(result, imagenet_a_df, on=['model', 'param_count', 'img_size', 'crop_pct', 'interpolation'], how='outer')
    result = pd.merge(result, imagenet_r_df, on=['model', 'param_count', 'img_size', 'crop_pct', 'interpolation'], how='outer')

    # Average top1 and top5 and add the average column after `model` column
    result['average_top1'] = result[['imagenet_top1', 'imagenet_real_top1', 'imagenetv2_top1', 'sketch_top1', 'imagenet_a_top1', 'imagenet_r_top1']].mean(axis=1)
    result['average_top5'] = result[['imagenet_top5', 'imagenet_real_top5', 'imagenetv2_top5', 'sketch_top5', 'imagenet_a_top5', 'imagenet_r_top5']].mean(axis=1)
    result = result[['model', 'average_top1', 'average_top5', 'param_count', 'img_size', 'crop_pct', 'interpolation', 'imagenet_top1', 'imagenet_top5', 'imagenet_real_top1', 'imagenet_real_top5', 'imagenetv2_top1', 'imagenetv2_top5', 'sketch_top1', 'sketch_top5', 'imagenet_a_top1', 'imagenet_a_top5', 'imagenet_r_top1', 'imagenet_r_top5']]
    result = result.sort_values(by='average_top1', ascending=False)

    # Round the values to 3 decimal places
    result = result.round(3)

    return result

global df
df = load_leaderboard()

def search_leaderboard(model_name):
    if not model_name:
        return df
    
    threshold = 95  # You can adjust this value to make the search more or less strict

    def calculate_similarity(row):
        similarity = fuzz.partial_ratio(model_name.lower(), row['model'].lower())
        return similarity if similarity >= threshold else 0

    # Add a new column for similarity scores
    df['similarity'] = df.apply(calculate_similarity, axis=1)

    # Filter and sort the dataframe
    filtered_df = df[df['similarity'] > 0].sort_values('similarity', ascending=False)

    # Remove the similarity column before returning
    filtered_df = filtered_df.drop('similarity', axis=1)

    return filtered_df

with gr.Blocks("Timm Leaderboard") as app:
    gr.HTML("<center><h1>PyTorch Image Models Leaderboard</h1></center>")
    gr.HTML("<p>This leaderboard is based on the results of the models from the <a href='https://github.com/huggingface/pytorch-image-models'>PyTorch Image Models</a> repository.</p>")
    with gr.Row():
        search_bar = gr.Textbox(lines=1, label="Search Model (You can press Enter to Search)", placeholder="Search for a model", scale=4)
        search_btn = gr.Button(value="Search", variant="primary", scale=1)
    leaderboard = gr.Dataframe(df)
    refresh_button = gr.Button(value="Refresh Leaderboard", variant="primary")
    refresh_button.click(load_leaderboard, outputs=[leaderboard])
    search_btn.click(search_leaderboard, inputs=[search_bar], outputs=[leaderboard])
    search_bar.submit(search_leaderboard, inputs=[search_bar], outputs=[leaderboard])

app.launch()