pipeline.py

from diffusers import DiffusionPipeline
import os
import sys
from huggingface_hub import HfApi, hf_hub_download
# from .tools import build_dataset_json_from_list
import torch

class MOSDiffusionPipeline(DiffusionPipeline):

    def __init__(self, config_yaml, list_inference, reload_from_ckpt=None, base_folder=None):
        """
        Initialize the MOS Diffusion pipeline and download the necessary files/folders.

        Args:
            config_yaml (str): Path to the YAML configuration file.
            list_inference (str): Path to the file containing inference prompts.
            reload_from_ckpt (str, optional): Checkpoint path to reload from.
            base_folder (str, optional): Base folder to store downloaded files. Defaults to the current working directory.
        """
        super().__init__()


        self.base_folder = base_folder if base_folder else os.getcwd()
        self.repo_id = "jadechoghari/qa-mdt" 
        self.config_yaml = config_yaml
        self.list_inference = list_inference
        self.reload_from_ckpt = reload_from_ckpt
        config_yaml_path = os.path.join(self.config_yaml)
        self.configs = self.load_yaml(config_yaml_path)
        if self.reload_from_ckpt is not None:
            self.configs["reload_from_ckpt"] = self.reload_from_ckpt

        self.dataset_key = build_dataset_json_from_list(self.list_inference)
        self.exp_name = os.path.basename(self.config_yaml.split(".")[0])
        self.exp_group_name = os.path.basename(os.path.dirname(self.config_yaml))

    def download_required_folders(self):
        """
        Downloads the necessary folders from the Hugging Face Hub if they are not already available locally.
        """
        api = HfApi()

        files = api.list_repo_files(repo_id=self.repo_id)

        required_folders = ["audioldm_train", "checkpoints", "infer", "log", "taming", "test_prompts"]

        files_to_download = [f for f in files if any(f.startswith(folder) for folder in required_folders)]

        for file in files_to_download:
            local_file_path = os.path.join(self.base_folder, file)
            if not os.path.exists(local_file_path):
                downloaded_file = hf_hub_download(repo_id=self.repo_id, filename=file)

                os.makedirs(os.path.dirname(local_file_path), exist_ok=True)

                os.rename(downloaded_file, local_file_path)

        sys.path.append(self.base_folder)

    def load_yaml(self, yaml_path):
        """
        Helper method to load the YAML configuration.
        """
        import yaml
        with open(yaml_path, "r") as f:
            return yaml.safe_load(f)


    @torch.no_grad()
    def __call__(self, prompt: str):
        """
        Run the MOS Diffusion Pipeline. This method calls the infer function from infer_mos5.py.
        """
        from .infer.infer_mos5 import infer
        dataset_key = self.build_dataset_json_from_prompt(prompt)

        # we run inference with the prompt - configs - and other settings
        infer(
            dataset_key=dataset_key,
            configs=self.configs,
            config_yaml_path=self.config_yaml,
            exp_group_name="qa_mdt",
            exp_name="mos_as_token"
        )

    def build_dataset_json_from_prompt(self, prompt: str):
        """
        Build dataset_key dynamically from the provided prompt.
        """
        # for simplicity let's just return the prompt as the dataset_key
        data = [{"wav": "", "caption": prompt}]  # no wav file, just the caption (prompt)
        return {"data": data}
        

# Example of how to use the pipeline
if __name__ == "__main__":
    pipe = MOSDiffusionPipeline()
    result = pipe("Generate a description of a sunny day.")
    print(result)