rasahq--rasa/.github/scripts/download_pretrained.py

import argparse
import logging
import time
from typing import List, NamedTuple, Optional, Text

from transformers import AutoTokenizer, TFAutoModel

import rasa.shared.utils.io
from rasa.nlu.utils.hugging_face.registry import (
    model_weights_defaults,
    model_class_dict,
)

logger = logging.getLogger(__name__)

COMP_NAME = "LanguageModelFeaturizer"
DEFAULT_MODEL_NAME = "bert"


class LmfSpec(NamedTuple):
    """Holds information about the LanguageModelFeaturizer."""

    model_name: Text
    model_weights: Text
    cache_dir: Optional[Text] = None


def get_model_name_and_weights_from_config(
    config_path: str,
) -> List[LmfSpec]:
    config = rasa.shared.utils.io.read_config_file(config_path)
    logger.info(config)
    steps = config.get("pipeline", [])

    # Look for LanguageModelFeaturizer steps
    steps = list(filter(lambda x: x["name"] == COMP_NAME, steps))

    lmf_specs = []
    for lmfeat_step in steps:
        if "model_name" not in lmfeat_step:
            if "model_weights" in lmfeat_step:
                model_weights = lmfeat_step["model_weights"]
                raise KeyError(
                    "When model_name is not given, then model_weights cannot be set. "
                    f"Here, model_weigths is set to {model_weights}"
                )
            model_name = DEFAULT_MODEL_NAME
            model_weights = model_weights_defaults[DEFAULT_MODEL_NAME]
        else:
            model_name = lmfeat_step["model_name"]

            if model_name not in model_class_dict:
                raise KeyError(
                    f"'{model_name}' not a valid model name. Choose from "
                    f"{str(list(model_class_dict.keys()))} or create"
                    f"a new class inheriting from this class to support your model."
                )

            model_weights = lmfeat_step.get("model_weights")
            if not model_weights:
                logger.info(
                    f"Model weights not specified. Will choose default model "
                    f"weights: {model_weights_defaults[model_name]}"
                )
                model_weights = model_weights_defaults[model_name]
        cache_dir = lmfeat_step.get("cache_dir", None)
        lmf_specs.append(LmfSpec(model_name, model_weights, cache_dir))

    return lmf_specs


def instantiate_to_download(comp: LmfSpec) -> None:
    """Instantiates Auto class instances, but only to download."""

    _ = AutoTokenizer.from_pretrained(comp.model_weights, cache_dir=comp.cache_dir)
    logger.info("Done with AutoTokenizer, now doing TFAutoModel")
    _ = TFAutoModel.from_pretrained(comp.model_weights, cache_dir=comp.cache_dir)


def download(config_path: str):
    lmf_specs = get_model_name_and_weights_from_config(config_path)

    if not lmf_specs:
        logger.info(f"No {COMP_NAME} found, therefore, skipping download")
        return

    for lmf_spec in lmf_specs:
        logger.info(
            f"model_name: {lmf_spec.model_name}, "
            f"model_weights: {lmf_spec.model_weights}, "
            f"cache_dir: {lmf_spec.cache_dir}"
        )
        start = time.time()

        instantiate_to_download(lmf_spec)

        duration_in_sec = time.time() - start
        logger.info(f"Instantiating Auto classes takes {duration_in_sec:.2f}seconds")


def create_argument_parser() -> argparse.ArgumentParser:
    """Downloads pretrained models, i.e., Huggingface weights."""
    parser = argparse.ArgumentParser(
        description="Downloads pretrained models, i.e., Huggingface weights, "
        "e.g. path to bert_diet_responset2t.yml"
    )
    parser.add_argument(
        "-c",
        "--config",
        type=str,
        required=True,
        help="The path to the config yaml file.",
    )
    return parser


if __name__ == "__main__":
    arg_parser = create_argument_parser()
    cmdline_args = arg_parser.parse_args()
    download(cmdline_args.config)