hparams.py

from text import symbols


class dotdict(dict):
    __getattr__ = dict.get
    __setattr__ = dict.__setitem__
    __delattr__ = dict.__delitem__


def create_hparams(hparams_string=None, verbose=False):
    """Create model hyperparameters. Parse nondefault from given string."""

    hparams = {
        ################################
        # Experiment Parameters        #
        ################################
        'epochs': 500,
        'iters_per_checkpoint': 1000,
        'seed': 1234,
        'dynamic_loss_scaling': True,
        'fp16_run': True,
        'distributed_run': False,
        'dist_backend': "nccl",
        'dist_url': "tcp://localhost:54321",
        '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': ['german_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,

        # Encoder parameters
        'encoder_kernel_size': 5,
        'encoder_n_convolutions': 3,
        'encoder_embedding_dim': 512,

        # Decoder parameters
        'n_frames_per_step': 1,  # currently only 1 is supported
        'decoder_rnn_dim': 1024,
        'prenet_dim': 256,
        'max_decoder_steps': 1000,
        'gate_threshold': 0.5,
        'p_attention_dropout': 0.1,
        'p_decoder_dropout': 0.1,

        # 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,
        'gamma': 0.995,
        'grad_clip_thresh': 1.0,
        'batch_size': 56,  # TODO: should be higher
        'mask_padding': True  # set model's padded outputs to padded values
    }

    if hparams_string:
        print('Parsing command line hparams: %s', hparams_string)
        hparams.parse(hparams_string)

    if verbose:
        print('Final parsed hparams: %s', hparams.values())

    return dotdict(hparams)