vllm_omni.diffusion.models.soulx_singer.modules.preprocess.rosvot ¶

device = device

fmax = hparams['fmax']

fmin = hparams['fmin']

hann_window = torch.hann_window(self.win_size).to(
    self.device
)

hop_size = hparams['hop_size']

mel_basis = mel.float().to(self.device)

n_fft = hparams['fft_size']

num_mels = hparams['audio_num_mel_bins']

sampling_rate = hparams['audio_sample_rate']

win_size = hparams['win_size']

cond_encoder = ConvBlocks(
    hidden_size,
    out_dims=hidden_size,
    dilations=None,
    kernel_size=3,
    layers_in_block=1,
    c_multiple=1,
    post_net_kernel=3,
)

hidden_size = hparams['hidden_size']

hparams = hparams.copy()

mel_encoder = ConvBlocks(
    hidden_size,
    out_dims=hidden_size,
    dilations=None,
    kernel_size=3,
    layers_in_block=2,
    c_multiple=1,
    post_net_kernel=3,
)

mel_proj = nn.Conv1d(
    hparams["use_mel_bins"],
    hidden_size,
    kernel_size=3,
    padding=1,
)

net = BackboneNet(hparams)

note_bd_min_gap = round(
    hparams.get("note_bd_min_gap", 100)
    * hparams["audio_sample_rate"]
    / 1000
    / hparams["hop_size"]
)

note_bd_out = nn.Linear(hidden_size, 1)

note_bd_ref_min_gap = round(
    hparams.get("note_bd_ref_min_gap", 50)
    * hparams["audio_sample_rate"]
    / 1000
    / hparams["hop_size"]
)

note_bd_temperature = max(
    1e-07, hparams.get("note_bd_temperature", 1.0)
)

note_bd_threshold = hparams.get('note_bd_threshold', 0.5)

pitch_decoder = PitchDecoder(hparams)

pitch_embed = Embedding(300, hidden_size, 0, 'kaiming')

use_pitch = hparams.get('use_pitch_embed', True)

use_wbd = hparams.get('use_wbd', True)

uv_embed = Embedding(3, hidden_size, 0, 'kaiming')

word_bd_embed = Embedding(3, hidden_size, 0, 'kaiming')

dropout = hparams.get('dropout', 0.0)

hidden_size = hparams['hidden_size']

multihead_dot_attn = nn.Linear(
    hidden_size, self.pitch_attn_num_head
)

note_bd_out = nn.Linear(hidden_size, 1)

note_bd_temperature = max(
    1e-07, hparams.get("note_bd_temperature", 1.0)
)

note_num = hparams.get('note_num', 100)

note_start = hparams.get('note_start', 30)

pitch_attn_num_head = hparams.get('pitch_attn_num_head', 1)

pitch_out = nn.Linear(
    hidden_size, hparams.get("note_num", 100) + 4
)

pitch_temperature = max(
    1e-07, hparams.get("note_pitch_temperature", 1.0)
)

post = ConvBlocks(
    hidden_size,
    out_dims=hidden_size,
    dilations=None,
    kernel_size=3,
    layers_in_block=1,
    c_multiple=1,
    post_net_kernel=3,
)

hparams = load_rosvot_config(
    resolved_config, hparams_str=f"note_bd_threshold={the}"
)

mel_net = MelNet(self.hparams)

midi = MidiExtractor(self.hparams)

pe = (
    pe
    if pe is not None
    and self.hparams.get("use_pitch_embed", False)
    else None
)

support_audio_input: bool = True

verbose = verbose