import requests
import pandas as pd
from tqdm.auto import tqdm
import streamlit as st
from huggingface_hub import HfApi, hf_hub_download
from huggingface_hub.repocard import metadata_load


def make_clickable(model_name):
    link = "https://huggingface.co/" + model_name
    return f'<a target="_blank" href="{link}">{model_name}</a>'


def get_model_ids():
    api = HfApi()
    # TODO: switch to hf-leaderboards for the final version.
    models = api.list_models(filter="hf-asr-leaderboard")
    model_ids = [x.modelId for x in models]
    return model_ids


def get_metadata(model_id):
    try:
        readme_path = hf_hub_download(model_id, filename="README.md")
        return metadata_load(readme_path)
    except requests.exceptions.HTTPError:
        # 404 README.md not found
        return None


def parse_metric_value(value):
    if isinstance(value, str):
        "".join(value.split("%"))
        try:
            value = float(value)
        except:  # noqa: E722
            value = None
    elif isinstance(value, list):
        if len(value) > 0:
            value = value[0]
        else:
            value = None
    value = round(value, 2) if value is not None else None
    return value


def parse_metrics_rows(meta):
    if "model-index" not in meta:
        return None
    for result in meta["model-index"][0]["results"]:
        if "dataset" not in result or "metrics" not in result:
            continue
        dataset = result["dataset"]["type"]
        if "args" not in result["dataset"]:
            continue
        row = {"dataset": dataset}
        for metric in result["metrics"]:
            type = metric["type"].lower().strip()
            value = parse_metric_value(metric["value"])
            if value is None:
                continue
            if type not in row or value < row[type]:
                # overwrite the metric if the new value is lower (e.g. with LM)
                row[type] = value
        yield row


@st.cache(ttl=600)
def get_data():
    data = []
    model_ids = get_model_ids()
    for model_id in tqdm(model_ids):
        meta = get_metadata(model_id)
        if meta is None:
            continue
        for row in parse_metrics_rows(meta):
            if row is None:
                continue
            row["model_id"] = model_id
            data.append(row)
    return pd.DataFrame.from_records(data)


dataframe = get_data()
selectable_datasets = list(set(dataframe.dataset.tolist()))

st.markdown("# 🤗 Leaderboards")

dataset = st.sidebar.selectbox(
    "Dataset",
    selectable_datasets,
    index=selectable_datasets.index("common_voice"),
)

dataset_df = dataframe[dataframe.dataset == dataset]
dataset_df = dataset_df.dropna(axis="columns", how="all")

metric = st.sidebar.selectbox(
    "Metric",
    list(filter(lambda column: column not in ("model_id", "dataset"), dataset_df.columns)),
)

dataset_df = dataset_df.filter(["model_id", metric])
dataset_df = dataset_df.dropna()
dataset_df = dataset_df.sort_values(by=metric, ascending=False)

st.markdown(
    "Please click on the model's name to be redirected to its model card which includes documentation and examples on how to use it."
)

# display the model ranks
dataset_df = dataset_df.reset_index(drop=True)
dataset_df.index += 1

# turn the model ids into clickable links
dataset_df["model_id"] = dataset_df["model_id"].apply(make_clickable)

table_html = dataset_df.to_html(escape=False)
table_html = table_html.replace("<th>", '<th align="left">')  # left-align the headers
st.write(table_html, unsafe_allow_html=True)

st.markdown(
    "Want to beat the Leaderboard? Don't see your model here? Simply add the `hf-leaderboards` tag to your model card alongside your evaluation metrics. See [this commit](https://huggingface.co/facebook/wav2vec2-base-960h/commit/88338305603a4d8db25aca96e669beb5f7dc65cb) as an example."
)