draw_utils.py

import pandas as pd
import numpy as np
from matplotlib.colors import LinearSegmentedColormap

PAGE_MARKDOWN = """
 <style>
 .reportview-container {
 margin-top: -2em;
 }
 #MainMenu {visibility: hidden;}
 .stDeployButton {display:none;}
 footer {visibility: hidden;}
 #stDecoration {display:none;}
 </style>
"""

PAGE_INFO = """[![Dataset on HF](https://huggingface.co/datasets/huggingface/badges/resolve/main/dataset-on-hf-lg.svg)](https://huggingface.co/datasets/RMT-team/babilong) | [GitHub](https://github.com/booydar/babilong) | [Paper](https://arxiv.org/abs/2406.10149) | [HF Dataset](https://huggingface.co/datasets/RMT-team/babilong) | [HF Dataset 1k samples per task](https://huggingface.co/datasets/RMT-team/babilong-1k-samples) |"""

LENGTHS = ['0k', '1k', '2k', '4k', '8k', '16k', '32k', '64k', '128k', '512k', '1M', '2M', '10M']
LENGTHS_32k = ['0k', '1k', '2k', '4k', '8k', '16k', '32k']
LENGTHS_128k = ['0k', '1k', '2k', '4k', '8k', '16k', '32k', '64k', '128k']


def load_results():
    old_results_path = "data/leaderboard-v0_results.csv"
    new_results_path = "babilong/babilong_results/all_results.csv"
    old_results = pd.read_csv(old_results_path)
    new_results = pd.read_csv(new_results_path)

    def normalize_model_name(name):
        if '/' in name:
            name = name.split('/')[-1]
        return name.lower()

    old_results['normalized_name'] = old_results['model_name'].apply(normalize_model_name)
    new_results['normalized_name'] = new_results['model_name'].apply(normalize_model_name)
    # clean duplicate models in v0 results and new results
    duplicate_models = set(old_results['normalized_name']).intersection(set(new_results['normalized_name']))
    old_results_filtered = old_results[~old_results['normalized_name'].isin(duplicate_models)]
    res = pd.concat([old_results_filtered, new_results])
    res.drop('normalized_name', axis=1, inplace=True)

    res['task'] = res['task'].str.replace('avg', 'avg(qa1-5)')

    res.replace(-1, np.nan, inplace=True)
    res['≤32k'] = res[LENGTHS_32k].mean(axis=1)
    res['≤128k'] = res[LENGTHS_128k].mean(axis=1)

    # Calculate the maximum length with non-NaN values for each model
    res['max_eval_length_idx'] = res.apply(
        lambda row: max([LENGTHS.index(col) for col in LENGTHS if not pd.isna(row[col])], default=-1), axis=1)
    res['max_eval_length'] = res['max_eval_length_idx'].apply(lambda x: LENGTHS[x])

    # Sort first by max length (descending) and then by average score (descending)
    res.sort_values(['max_eval_length_idx', '≤128k'], ascending=[False, False], inplace=True)

    return res


# from pandas/io/formats/style.py
def relative_luminance(rgba) -> float:
    """
    Calculate relative luminance of a color.

    The calculation adheres to the W3C standards
    (https://www.w3.org/WAI/GL/wiki/Relative_luminance)

    Parameters
    ----------
    color : rgb or rgba tuple

    Returns
    -------
    float
        The relative luminance as a value from 0 to 1
    """
    r, g, b = (
        x / 12.92 if x <= 0.04045 else ((x + 0.055) / 1.055) ** 2.4
        for x in rgba[:3]
    )
    return 0.2126 * r + 0.7152 * g + 0.0722 * b


def style_dataframe(df):
    """
    Style a pandas DataFrame with a color gradient.
    """
    styled_df = df.copy()
    numeric_columns = styled_df.columns[1:]

    def color_scale(val):
        cmap = LinearSegmentedColormap.from_list('ryg', ["red", "yellow", "green"], N=256)

        if pd.isna(val):
            return 'background-color: white; color: white;'

        min_val = 0
        max_val = 100
        normalized = (val - min_val) / (max_val - min_val)
        rgba = cmap(normalized)
        text_color = 'white' if relative_luminance(rgba) < 0.408 else 'black'
        return f'background-color: rgba({rgba[0]*255},{rgba[1]*255},{rgba[2]*255},{rgba[3]}); color: {text_color}'

    styled_df = styled_df.style.map(color_scale, subset=numeric_columns)
    return styled_df