hparams.py

from text.symbols import symbols
import torch


@torch.jit.ignore()
class Hparams():
    ################################
    # Experiment Parameters        #
    ################################
    epochs = 2000,
    iters_per_checkpoint = 1000,
    seed = 54734,
    dynamic_loss_scaling = True,
    fp16_run = True,
    distributed_run = True,
    dist_backend = "nccl",
    dist_url = "tcp://localhost:13314",
    cudnn_enabled = True,
    cudnn_benchmark = False,
    ignore_layers = ['embedding.weight'],

    ################################
    # Data Parameters             #
    ################################
    load_mel_from_disk = False,
    training_files = 'filelists/ljs_audio_text_train_filelist.txt',
    validation_files = 'filelists/ljs_audio_text_val_filelist.txt',
    text_cleaners = ['english_cleaners'],

    ################################
    # Audio Parameters             #
    ################################
    max_wav_value = 32768.0,
    sampling_rate = 22050,
    filter_length = 1024,
    hop_length = 256,
    win_length = 1024,
    n_mel_channels = 80,
    mel_fmin = 0.0,
    mel_fmax = 8000.0,

    ################################
    # Model Parameters             #
    ################################
    n_symbols = len(symbols),
    symbols_embedding_dim = 512,
    speakers_number = 2,
    # Encoder parameters
    encoder_kernel_size = 1,
    encoder_n_convolutions = 3,
    encoder_embedding_dim = 512,
    languages_number = 2,
    lang_embedding_dims = 10,
    speaker_embedding_dims = 32,
    # languages = ["korean","english"],
    bottleneck_dim = 3,
    generator_dim = 16,
    perfect_sampling = True,

    # Decoder parameters
    n_frames_per_step = 1,  # currently only 1 is supported
    decoder_rnn_dim = 1024,
    prenet_dim = 256,
    max_decoder_steps = 2000,
    gate_threshold = 0.5,
    p_attention_dropout = 0.1,
    p_decoder_dropout = 0.1,
    discriminative_embedding_dims = 32,
    # Attention parameters
    attention_rnn_dim = 1024,
    attention_dim = 128,

    # Location Layer parameters
    attention_location_n_filters = 32,
    attention_location_kernel_size = 31,

    # Mel-post processing network parameters
    postnet_embedding_dim = 512,
    postnet_kernel_size = 5,
    postnet_n_convolutions = 5,

    ################################
    # Optimization Hyperparameters #
    ################################
    use_saved_learning_rate = False,
    learning_rate = 1e-3,
    weight_decay = 1e-6,
    grad_clip_thresh = 1.0,
    batch_size = 64,
    mask_padding = True  # set model's padded outputs to padded values


@torch.jit.ignore()
def create_hparams(hparams=None):
    """Create model hyperparameters. Parse nondefault from given string."""
    hparams = Hparams()
    return hparams