hparams.py

import tensorflow as tf
# NOTE: If you want full control for model architecture. please take a look
# at the code and change whatever you want. Some hyper parameters are hardcoded.

# Default hyperparameters:
hparams = tf.contrib.training.HParams(
    name="deepvoice3",

    # Text:
    # [en, jp]
    frontend='en',

    # Replace words to its pronunciation with fixed probability.
    # e.g., 'hello' to 'HH AH0 L OW1'
    # [en, jp]
    # en: Word -> pronunciation using CMUDict
    # jp: Word -> pronounciation usnig MeCab
    # [0 ~ 1.0]: 0 means no replacement happens.
    replace_pronunciation_prob=0.5,

    # Convenient model builder
    # [deepvoice3, deepvoice3_multispeaker, nyanko]
    # Definitions can be found at deepvoice3_pytorch/builder.py
    # deepvoice3: DeepVoice3 https://arxiv.org/abs/1710.07654
    # deepvoice3_multispeaker: Multi-speaker version of DeepVoice3
    # nyanko: https://arxiv.org/abs/1710.08969
    builder="deepvoice3",

    # Must be configured depends on the dataset and model you use
    n_speakers=1,
    speaker_embed_dim=16,

    # Audio:
    num_mels=80,
    fmin=125,
    fmax=7600,
    fft_size=1024,
    hop_size=256,
    sample_rate=22050,
    preemphasis=0.97,
    min_level_db=-100,
    ref_level_db=20,
    # whether to rescale waveform or not.
    # Let x is an input waveform, rescaled waveform y is given by:
    # y = x / np.abs(x).max() * rescaling_max
    rescaling=False,
    rescaling_max=0.999,
    # mel-spectrogram is normalized to [0, 1] for each utterance and clipping may
    # happen depends on min_level_db and ref_level_db, causing clipping noise.
    # If False, assertion is added to ensure no clipping happens.
    allow_clipping_in_normalization=True,

    # Model:
    downsample_step=4,  # must be 4 when builder="nyanko"
    outputs_per_step=1,  # must be 1 when builder="nyanko"
    embedding_weight_std=0.1,
    speaker_embedding_weight_std=0.01,
    padding_idx=0,
    # Maximum number of input text length
    # try setting larger value if you want to give very long text input
    max_positions=512,
    dropout=1 - 0.95,
    kernel_size=3,
    text_embed_dim=128,
    encoder_channels=256,
    decoder_channels=256,
    # Note: large converter channels requires significant computational cost
    converter_channels=256,
    query_position_rate=1.0,
    # can be computed by `compute_timestamp_ratio.py`.
    key_position_rate=1.385,  # 2.37 for jsut
    key_projection=False,
    value_projection=False,
    use_memory_mask=True,
    trainable_positional_encodings=False,
    freeze_embedding=False,
    # If True, use decoder's internal representation for postnet inputs,
    # otherwise use mel-spectrogram.
    use_decoder_state_for_postnet_input=True,

    # Data loader
    pin_memory=True,
    num_workers=2,  # Set it to 1 when in Windows (MemoryError, THAllocator.c 0x5)

    # Loss
    masked_loss_weight=0.5,  # (1-w)*loss + w * masked_loss
    priority_freq=3000,  # heuristic: priotrize [0 ~ priotiry_freq] for linear loss
    priority_freq_weight=0.0,  # (1-w)*linear_loss + w*priority_linear_loss
    # https://arxiv.org/pdf/1710.08969.pdf
    # Adding the divergence to the loss stabilizes training, expecially for
    # very deep (> 10 layers) networks.
    # Binary div loss seems has approx 10x scale compared to L1 loss, so I choose 0.1.
    binary_divergence_weight=0.1,  # set 0 to disable
    use_guided_attention=True,
    guided_attention_sigma=0.2,

    # Training:
    batch_size=16,
    adam_beta1=0.5,
    adam_beta2=0.9,
    adam_eps=1e-6,
    amsgrad=False,
    initial_learning_rate=5e-4,  # 0.001,
    lr_schedule="noam_learning_rate_decay",
    lr_schedule_kwargs={},
    nepochs=2000,
    weight_decay=0.0,
    clip_thresh=0.1,

    # Save
    checkpoint_interval=10000,
    eval_interval=10000,
    save_optimizer_state=True,

    # Eval:
    # this can be list for multple layers of attention
    # e.g., [True, False, False, False, True]
    force_monotonic_attention=True,
    # Attention constraint for incremental decoding
    window_ahead=3,
    # 0 tends to prevent word repretetion, but sometime causes skip words
    window_backward=1,
    power=1.4,  # Power to raise magnitudes to prior to phase retrieval

    # GC:
    # Forced garbage collection probability
    # Use only when MemoryError continues in Windows (Disabled by default)
    #gc_probability = 0.001,

    # json_meta mode only
    # 0: "use all",
    # 1: "ignore only unmatched_alignment",
    # 2: "fully ignore recognition",
    ignore_recognition_level=2,
    # when dealing with non-dedicated speech dataset(e.g. movie excerpts), setting min_text above 15 is desirable. Can be adjusted by dataset.
    min_text=20,
    # if true, data without phoneme alignment file(.lab) will be ignored
    process_only_htk_aligned=False,
)


def hparams_debug_string():
    values = hparams.values()
    hp = ['  %s: %s' % (name, values[name]) for name in sorted(values)]
    return 'Hyperparameters:\n' + '\n'.join(hp)