configuration_musicfm.py

from transformers import PretrainedConfig


class MusicFMConfig(PretrainedConfig):
    model_type = "musicfm"

    def __init__(
        self,
        num_codebooks: int = 1,
        codebook_dim: int = 16,
        codebook_size: int = 4096,
        features: list[str] = ["melspec_2048"],
        hop_length: int = 240,
        n_mels: int = 128,
        conv_dim: int = 512,
        encoder_dim: int = 1024,
        encoder_depth: int = 12,
        mask_hop: float = 0.4,
        mask_prob: float = 0.6,
        is_flash: bool = False,
        stat: dict[str, float] = {},
        **kwargs,
    ) -> None:
        super().__init__(**kwargs)

        self.num_codebooks = num_codebooks
        self.codebook_dim = codebook_dim
        self.codebook_size = codebook_size
        self.features = features
        self.hop_length = hop_length
        self.n_mels = n_mels
        self.conv_dim = conv_dim
        self.encoder_dim = encoder_dim
        self.encoder_depth = encoder_depth
        self.mask_hop = mask_hop
        self.mask_prob = mask_prob
        self.is_flash = is_flash
        self.stat = stat


class MusicFMInferenceConfig(MusicFMConfig):
    model_type = "musicfm_inference"

    def __init__(
        self,
        num_codebooks: int = 1,
        codebook_dim: int = 16,
        codebook_size: int = 4096,
        features: list[str] = ["melspec_2048"],
        hop_length: int = 240,
        n_mels: int = 128,
        conv_dim: int = 512,
        encoder_dim: int = 1024,
        encoder_depth: int = 12,
        mask_hop: float = 0.4,
        mask_prob: float = 0.6,
        is_flash: bool = False,
        layer_index: int = 9,
        stat: dict[str, float] = {},
        **kwargs,
    ) -> None:
        super().__init__(
            num_codebooks=num_codebooks,
            codebook_dim=codebook_dim,
            codebook_size=codebook_size,
            features=features,
            hop_length=hop_length,
            n_mels=n_mels,
            conv_dim=conv_dim,
            encoder_dim=encoder_dim,
            encoder_depth=encoder_depth,
            mask_hop=mask_hop,
            mask_prob=mask_prob,
            is_flash=is_flash,
            stat=stat,
            **kwargs,
        )

        self.layer_index = layer_index