src/results.py

import asyncio

import gradio as gr
import pandas as pd
from huggingface_hub import HfFileSystem

import src.constants as constants
from src.hub import load_file


def fetch_result_paths():
    fs = HfFileSystem()
    paths = fs.glob(f"{constants.RESULTS_DATASET_ID}/**/**/*.json")
    return paths


def sort_result_paths_per_model(paths):
    from collections import defaultdict

    d = defaultdict(list)
    for path in paths:
        model_id, _ = path[len(constants.RESULTS_DATASET_ID) + 1 :].rsplit("/", 1)
        d[model_id].append(path)
    return {model_id: sorted(paths) for model_id, paths in d.items()}


def update_load_results_component():
    return (gr.Button("Load", interactive=True),) * 2


async def load_results_dataframe(model_id, result_paths_per_model=None):
    if not model_id or not result_paths_per_model:
        return
    result_paths = result_paths_per_model[model_id]
    results = await asyncio.gather(*[load_file(path) for path in result_paths])
    data = {"results": {}, "configs": {}}
    for result in results:
        data["results"].update(result["results"])
        data["configs"].update(result["configs"])
        model_name = result.get("model_name", "Model")
    df = pd.json_normalize([data])
    # df.columns = df.columns.str.split(".")  # .split return a list instead of a tuple
    return df.set_index(pd.Index([model_name])).reset_index()


async def load_results_dataframes(*model_ids, result_paths_per_model=None):
    result = await asyncio.gather(
        *[load_results_dataframe(model_id, result_paths_per_model) for model_id in model_ids]
    )
    return result


def display_results(task, hide_std_errors, show_only_differences, *dfs):
    dfs = [df.set_index("index") for df in dfs if "index" in df.columns]
    if not dfs:
        return None, None
    df = pd.concat(dfs)
    df = df.T.rename_axis(columns=None)
    return (
        display_tab("results", df, task, hide_std_errors=hide_std_errors),
        display_tab("configs", df, task, show_only_differences=show_only_differences),
    )


def display_tab(tab, df, task, hide_std_errors=True, show_only_differences=False):
    if show_only_differences:
        any_difference = df.ne(df.iloc[:, 0], axis=0).any(axis=1)
    df = df.style.format(escape="html", na_rep="")
    # Hide rows
    df.hide(
        [
            row
            for row in df.index
            if (
                not row.startswith(f"{tab}.")
                or row.startswith(f"{tab}.leaderboard.")
                or row.endswith(".alias")
                or (
                    not row.startswith(f"{tab}.{task}")
                    if task != "All"
                    else row.startswith(f"{tab}.leaderboard_arc_challenge")
                )
                # Hide std errors
                or (hide_std_errors and row.endswith("_stderr,none"))
                # Hide non-different rows
                or (show_only_differences and not any_difference[row])
            )
        ],
        axis="index",
    )
    # Color metric result cells
    idx = pd.IndexSlice
    colored_rows = idx[
        [
            row
            for row in df.index
            if row.endswith("acc,none") or row.endswith("acc_norm,none") or row.endswith("exact_match,none")
        ]
    ]  # Apply only on numeric cells, otherwise the background gradient will not work
    subset = idx[colored_rows, idx[:]]
    df.background_gradient(cmap="Greens", vmin=0, vmax=1, subset=subset, axis=None)
    # Format index values: remove prefix and suffix
    start = len(f"{tab}.leaderboard_") if task == "All" else len(f"{tab}.{task} ")
    df.format_index(lambda idx: idx[start:].removesuffix(",none"), axis="index")
    return df.to_html()


def update_tasks_component():
    return (
        gr.Radio(
            ["All"] + list(constants.TASKS.values()),
            label="Tasks",
            info="Evaluation tasks to be displayed",
            value="All",
            visible=True,
        ),
    ) * 2


def clear_results():
    # model_id_1, model_id_2, dataframe_1, dataframe_2, load_results_btn, load_configs_btn, results_task, configs_task
    return (
        None,
        None,
        None,
        None,
        *(gr.Button("Load", interactive=False),) * 2,
        *(
            gr.Radio(
                ["All"] + list(constants.TASKS.values()),
                label="Tasks",
                info="Evaluation tasks to be displayed",
                value="All",
                visible=False,
            ),
        )
        * 2,
    )


def display_loading_message_for_results():
    return ("<h3 style='text-align: center;'>Loading...</h3>",) * 2