app.py

import json
import re
from pathlib import Path

import requests
import streamlit as st
import yaml
from huggingface_hub import hf_hub_download
from streamlit_ace import st_ace
from streamlit_tags import st_tags

# exact same regex as in the Hub server. Please keep in sync.
REGEX_YAML_BLOCK = re.compile(r"---[\n\r]+([\S\s]*?)[\n\r]+---[\n\r]")

with open("languages.json") as f:
    lang2name = json.load(f)


def try_parse_yaml(yaml_block):
    try:
        metadata = yaml.load(yaml_block, yaml.SafeLoader)
    except yaml.YAMLError as e:
        print("Error while parsing the metadata YAML:")
        if hasattr(e, "problem_mark"):
            if e.context is not None:
                st.error(
                    str(e.problem_mark)
                    + "\n  "
                    + str(e.problem)
                    + " "
                    + str(e.context)
                    + "\nPlease correct the README.md and retry."
                )
            else:
                st.error(
                    str(e.problem_mark)
                    + "\n  "
                    + str(e.problem)
                    + "\nPlease correct the README.md and retry."
                )
        else:
            st.error(
                "Something went wrong while parsing the metadata. "
                "Make sure it's written according to the YAML spec!"
            )
        return None
    return metadata


def main():
    st.markdown("# The 🤗 Speech Bench Metrics Editor")
    st.markdown("This tool will help you report the evaluation metrics for all of your speech recognition models. "
                "Follow the steps and watch your models appear on the [Speech Bench Leaderboard](https://huggingface.co/spaces/huggingface/hf-speech-bench)!")
    st.markdown("## 1. Load your model's metadata")
    st.markdown("Enter your model's path below.")
    model_id = st.text_input("", placeholder="<username>/<model>")
    if not model_id.strip():
        st.stop()
    try:
        readme_path = hf_hub_download(model_id, filename="README.md")
    except requests.exceptions.HTTPError:
        st.error(
            f"ERROR: https://huggingface.co/{model_id}/blob/main/README.md "
            f"not found, make sure you've entered a correct model path and created a model card for it!"
        )
        st.stop()

    content = Path(readme_path).read_text()
    match = REGEX_YAML_BLOCK.search(content)
    if match:
        meta_yaml = match.group(1)
    else:
        st.error(
            "ERROR: Couldn't find the metadata section inside your model's `README.md`. Do you have some basic metadata "
            "enclosed in `---` as described in [the Hub documentation](https://huggingface.co/docs/hub/model-repos#model-card-metadata)?"
        )
        st.stop()

    metadata = try_parse_yaml(meta_yaml)
    if metadata is None:
        st.stop()
        return
    else:
        st.success("Successfully loaded the metadata!")
    with st.expander("Inspect the parsed metadata for debugging"):
        st.json(metadata)

    st.markdown("## 2. Edit the data")
    
    if "tags" not in metadata:
        metadata["tags"] = []
    metadata["tags"].append("hf-asr-leaderboard")

    ############################
    # LANGUAGES
    ############################
    st.markdown("### Language(s)")
    st.markdown(
        "For each spoken language that your model handles, enter an "
        "[ISO 639-1](https://en.wikipedia.org/wiki/List_of_ISO_639-1_codes) language code, or "
        "find an appropriate alternative from "
        "[our list here](https://huggingface.co/spaces/huggingface/hf-speech-bench/blob/main/languages.json). "
        "When in doubt, use the most generic language code, e.g. `en` instead of `en-GB` and `en-US`."
    )
    st.markdown("*Example*: `en, gsw, pt-BR`")
    metadata["language"] = metadata["language"] if "language" in metadata else []
    metadata["language"] = (
        metadata["language"]
        if isinstance(metadata["language"], list)
        else [metadata["language"]]
    )

    languages = st_tags(
        label="", text="add more if needed, and press enter", value=metadata["language"],
        key=model_id+"_langs"
    )
    if not languages:
        st.stop()
        return
    lang_names = [lang2name[lang] if lang in lang2name else lang for lang in languages]
    st.markdown("These languages will be parsed by the leaderboard as: ")
    st.code(", ".join(lang_names))
    metadata["language"] = languages

    ############################
    # TRAIN DATASETS
    ############################
    st.markdown("### Training dataset(s)")
    st.markdown(
        "List the datasets that your model was **trained** on. "
        "If the datasets aren't published on the Hub yet, just add their names anyway."
    )
    st.markdown(
        "*Example*: `librispeech_asr, mozilla-foundation/common_voice_8_0, my_custom_youtube_dataset`"
    )

    if "datasets" not in metadata:
        metadata["datasets"] = []

    train_datasets = st_tags(
        label="", text="add more if needed, and press enter", value=metadata["datasets"],
        key=model_id+"_train_dataset"
    )
    if not train_datasets:
        st.stop()
        return
    if "common_voice" in train_datasets:
        st.warning(
            "WARNING: `common_voice` is deprecated, please replace it with its equivalent: "
            "`mozilla-foundation/common_voice_6_1`"
        )
    metadata["datasets"] = train_datasets

    ############################
    # MODEL NAME
    ############################
    st.markdown("### Model name")
    st.markdown("Enter a pretty name for your model.")
    st.markdown("*Example*: `XLS-R Wav2Vec2 LM Spanish by Jane Doe`")

    if "model-index" not in metadata:
        metadata["model-index"] = [{}]
    if "name" not in ["model-index"][0]:
        metadata["model-index"][0]["name"] = model_id.split("/")[-1]
    model_name = st.text_input("", value=metadata["model-index"][0]["name"])
    if not model_name:
        st.stop()
        return
    metadata["model-index"][0]["name"] = model_name

    ############################
    # EVAL RESULTS
    ############################
    st.markdown("### Evaluation results")
    st.markdown(
        "To edit the metrics, you can either use the YAML editor below, or add new metrics using the handy "
        "form under it."
    )
    if "results" not in metadata["model-index"][0]:
        metadata["model-index"][0]["results"] = []

    results_editor = st.empty()
    with results_editor:
        results_yaml = yaml.dump(
            metadata["model-index"][0]["results"], sort_keys=False, line_break="\n"
        )
        results_yaml = st_ace(value=results_yaml, language="yaml")
        metadata["model-index"][0]["results"] = try_parse_yaml(results_yaml)

    dataset_path_kwargs = {}
    dataset_name_kwargs = {}
    if (
        len(metadata["model-index"][0]["results"]) > 0
        and "dataset" in metadata["model-index"][0]["results"][0]
    ):
        if "type" in metadata["model-index"][0]["results"][0]["dataset"]:
            dataset_path_kwargs["value"] = metadata["model-index"][0]["results"][0][
                "dataset"
            ]["type"]
        if "name" in metadata["model-index"][0]["results"][0]["dataset"]:
            dataset_name_kwargs["value"] = metadata["model-index"][0]["results"][0][
                "dataset"
            ]["name"]

    with st.form(key="eval_form"):
        dataset_path = st.text_input(
            label="Dataset path / id",
            placeholder="mozilla-foundation/common_voice_8_0",
            **dataset_path_kwargs,
        )
        dataset_name = st.text_input(
            label="A pretty name for the dataset. Examples: 'Common Voice 9.0 (French)', 'LibriSpeech (clean)'",
            placeholder="Common Voice 9.0 (French)",
            **dataset_name_kwargs,
        )
        dataset_config = st.text_input(
            label="Dataset configuration. Examples: clean, other, en, pt-BR",
            placeholder="en",
        )
        dataset_language_kwargs = {"value": languages[0]} if len(languages) > 0 else {}
        dataset_language = st.text_input(
            label="Dataset language. Examples: en, pt-BR",
            placeholder="en",
            **dataset_language_kwargs
        )
        dataset_split = st.text_input(
            label="Dataset split. Examples: test, validation",
            value="test",
            placeholder="test",
        )
        metric2name = {"wer": "Word Error Rate", "cer": "Character Error Rate"}
        metric_type = st.selectbox(
            label="Metric",
            options=["wer", "cer"],
            format_func=lambda key: metric2name[key],
        )
        metric_name = st.text_input(
            label="A pretty name for the metric. Example: Test WER (+LM)",
            placeholder="Test WER",
            value="Test WER",
        )
        metric_value = st.text_input(
            label="Metric value. Use values in range 0.0 to 100.0.",
            placeholder="12.34",
        )
        # try:
        #    metric_value = float(metric_value)
        # except ValueError:
        #    st.error(
        #        f"Couldn't parse `{metric_value}`. Make sure it's a number from 0.0 to 100.0"
        #    )

        submitted = st.form_submit_button("Add metric")
        if (
            submitted
            and dataset_name
            and dataset_path
            and dataset_config
            and dataset_split
            and dataset_language
            and metric_name
            and metric_type
            and metric_value
        ):
            metric = {
                "name": metric_name,
                "type": metric_type,
                "value": metric_value,
            }
            # first, try to find an existing dataset+config record to add a new metric to it
            updated_existing = False
            for existing_result in metadata["model-index"][0]["results"]:
                existing_dataset = existing_result["dataset"]
                if (
                    existing_dataset["type"] == dataset_path
                    and "config" in existing_dataset
                    and existing_dataset["config"] == dataset_config
                    and "split" in existing_dataset
                    and existing_dataset["split"] == dataset_split
                ):
                    if "metrics" not in existing_result:
                        existing_result["metrics"] = []
                    existing_result["metrics"].append(metric)
                    updated_existing = True
                    break
            # if no dataset+config results found, create a new one
            if not updated_existing:
                result = {
                    "task": {
                        "name": "Automatic Speech Recognition",
                        "type": "automatic-speech-recognition",
                    },
                    "dataset": {
                        "name": dataset_name,
                        "type": dataset_path,
                        "config": dataset_config,
                        "split": dataset_split,
                        "args": {"language": dataset_language},
                    },
                    "metrics": [metric],
                }
                metadata["model-index"][0]["results"].append(result)

            # update the code editor
            with results_editor:
                results_yaml = yaml.dump(
                    metadata["model-index"][0]["results"],
                    sort_keys=False,
                    line_break="\n",
                )
                results_yaml = st_ace(value=results_yaml, language="yaml")
                metadata["model-index"][0]["results"] = try_parse_yaml(results_yaml)
            st.success(
                f"Added the metric for {dataset_path} - {dataset_config}! "
                f"Check the result in the YAML editor above."
            )
        elif submitted:
            st.error(
                f"Make sure that you've filled the whole form before clicking 'Add metric'!"
            )

    ############################
    # FINAL YAML
    ############################
    st.markdown("## 3. Copy the generated metadata")
    st.markdown(
        "Copy the YAML from below and replace the metadata at the top of your model's README.md here: "
        f"https://huggingface.co/{model_id}/edit/main/README.md"
    )
    st.markdown("For more info on the metadata schema please refer to "
                "https://raw.githubusercontent.com/huggingface/hub-docs/main/modelcard.md")
    
    new_yaml = yaml.dump(metadata, sort_keys=False, line_break="\n")
    st.markdown(f"```yaml\n---\n{new_yaml}---\n```")


if __name__ == "__main__":
    main()