hparams.py

# -*- coding: utf-8 -*-

import tensorflow as tf
import numpy as np

hparams = tf.contrib.training.HParams(
    name="Tacotron-2",
    # tacotron hyper parameter
    cleaners='korean_cleaners',  # 'korean_cleaners'   or 'english_cleaners'
    skip_path_filter=False,  # npz파일에서 불필요한 것을 거르는 작업을 할지 말지 결정. receptive_field 보다 짧은 data를 걸러야 하기 때문에 해 줘야 한다.
    use_lws=False,
    # Audio
    sample_rate=24000,  #
    # shift can be specified by either hop_size(우선) or frame_shift_ms
    hop_size=300,  # frame_shift_ms = 12.5ms
    fft_size=2048,  # n_fft. 주로 1024로 되어있는데, tacotron에서 2048사용
    win_size=1200,  # 50ms
    num_mels=80,
    # Spectrogram Pre-Emphasis (Lfilter: Reduce spectrogram noise and helps model certitude levels. Also allows for better G&L phase reconstruction)
    preemphasize=True,  # whether to apply filter
    preemphasis=0.97,
    min_level_db=-100,
    ref_level_db=20,
    signal_normalization=True,
    # Whether to normalize mel spectrograms to some predefined range (following below parameters)
    allow_clipping_in_normalization=True,  # Only relevant if mel_normalization = True
    symmetric_mels=True,
    # Whether to scale the data to be symmetric around 0. (Also multiplies the output range by 2, faster and cleaner convergence)
    max_abs_value=4.,
    # max absolute value of data. If symmetric, data will be [-max, max] else [0, max] (Must not be too big to avoid gradient explosion, not too small for fast convergence)

    rescaling=True,
    rescaling_max=0.999,

    trim_silence=True,  # Whether to clip silence in Audio (at beginning and end of audio only, not the middle)
    # M-AILABS (and other datasets) trim params (there parameters are usually correct for any data, but definitely must be tuned for specific speakers)
    trim_fft_size=512,
    trim_hop_size=128,
    trim_top_db=23,

    clip_mels_length=True,
    # For cases of OOM (Not really recommended, only use if facing unsolvable OOM errors, also consider clipping your samples to smaller chunks)
    max_mel_frames=1000,
    # Only relevant when clip_mels_length = True, please only use after trying output_per_steps=3 and still getting OOM errors.

    l2_regularization_strength=0,  # Coefficient in the L2 regularization.
    sample_size=9000,  # Concatenate and cut audio samples to this many samples
    silence_threshold=0,
    # Volume threshold below which to trim the start and the end from the training set samples. e.g. 2

    filter_width=3,
    gc_channels=32,  # global_condition_vector의 차원. 이것 지정함으로써, global conditioning을 모델에 반영하라는 의미가 된다.

    input_type="raw",  # 'mulaw-quantize', 'mulaw', 'raw',   mulaw, raw 2가지는 scalar input
    scalar_input=True,  # input_type과 맞아야 함.

    dilations=[1, 2, 4, 8, 16, 32, 64, 128, 256, 512,
               1, 2, 4, 8, 16, 32, 64, 128, 256, 512],
    residual_channels=128,
    dilation_channels=256,
    quantization_channels=256,
    out_channels=30,  # discretized_mix_logistic_loss를 적용하기 때문에, 3의 배수
    skip_channels=128,
    use_biases=True,
    upsample_type='SubPixel',  # 'SubPixel', None
    upsample_factor=[12, 25],  # np.prod(upsample_factor) must equal to hop_size

    # wavenet training hp
    wavenet_batch_size=2,  # 16--> OOM. wavenet은 batch_size가 고정되어야 한다.
    store_metadata=False,
    num_steps=1000000,  # Number of training steps

    # Learning rate schedule
    wavenet_learning_rate=1e-3,  # wavenet initial learning rate
    wavenet_decay_rate=0.5,  # Only used with 'exponential' scheme. Defines the decay rate.
    wavenet_decay_steps=300000,  # Only used with 'exponential' scheme. Defines the decay steps.

    # Regularization parameters
    wavenet_clip_gradients=True,  # Whether the clip the gradients during wavenet training.

    # residual 결과를 sum할 때, 
    legacy=True,
    # Whether to use legacy mode: Multiply all skip outputs but the first one with sqrt(0.5) (True for more early training stability, especially for large models)

    # residual block내에서  x = (x + residual) * np.sqrt(0.5)
    residual_legacy=True,
    # Whether to scale residual blocks outputs by a factor of sqrt(0.5) (True for input variance preservation early in training and better overall stability)

    wavenet_dropout=0.05,

    optimizer='adam',
    momentum=0.9,  # 'Specify the momentum to be used by sgd or rmsprop optimizer. Ignored by the adam optimizer.
    max_checkpoints=3,  # 'Maximum amount of checkpoints that will be kept alive. Default: '

    ####################################
    ####################################
    ####################################
    # TACOTRON HYPERPARAMETERS

    # Training
    adam_beta1=0.9,
    adam_beta2=0.999,

    # Learning rate schedule
    tacotron_decay_learning_rate=True,  # boolean, determines if the learning rate will follow an exponential decay
    tacotron_start_decay=40000,  # Step at which learning decay starts
    tacotron_decay_steps=18000,  # Determines the learning rate decay slope (UNDER TEST)
    tacotron_decay_rate=0.5,  # learning rate decay rate (UNDER TEST)
    tacotron_initial_learning_rate=1e-3,  # starting learning rate
    tacotron_final_learning_rate=1e-4,  # minimal learning rate

    initial_data_greedy=True,
    initial_phase_step=8000,
    # 여기서 지정한 step 이전에는 data_dirs의 각각의 디렉토리에 대하여 같은 수의 example을 만들고, 이후, weght 비듈에 따라 ... 즉, 아래의 'main_data_greedy_factor'의 영향을 받는다.
    main_data_greedy_factor=0,
    main_data=[''],  # 이곳에 있는 directory 속에 있는 data는 가중치를 'main_data_greedy_factor' 만큼 더 준다.
    prioritize_loss=False,

    # Model
    model_type='single',  # [single, multi-speaker]
    speaker_embedding_size=16,

    embedding_size=512,  # 'ᄀ', 'ᄂ', 'ᅡ' 에 대한 embedding dim
    dropout_prob=0.5,

    reduction_factor=2,  # reduction_factor가 적으면 더 많은 iteration이 필요하므로, 더 많은 메모리가 필요하다.

    # Encoder
    enc_conv_num_layers=3,
    enc_conv_kernel_size=5,
    enc_conv_channels=512,
    tacotron_zoneout_rate=0.1,
    encoder_lstm_units=256,

    attention_type='loc_sen',  # 'loc_sen', 'bah_mon_norm'
    attention_size=128,

    # Attention mechanism
    smoothing=False,  # Whether to smooth the attention normalization function
    attention_dim=128,  # dimension of attention space
    attention_filters=32,  # number of attention convolution filters
    attention_kernel=(31,),  # kernel size of attention convolution
    cumulative_weights=True,
    # Whether to cumulate (sum) all previous attention weights or simply feed previous weights (Recommended: True)

    # Attention synthesis constraints
    # "Monotonic" constraint forces the model to only look at the forwards attention_win_size steps.
    # "Window" allows the model to look at attention_win_size neighbors, both forward and backward steps.
    synthesis_constraint=False,
    # Whether to use attention windows constraints in synthesis only (Useful for long utterances synthesis)
    synthesis_constraint_type='window',  # can be in ('window', 'monotonic').
    attention_win_size=7,
    # Side of the window. Current step does not count. If mode is window and attention_win_size is not pair, the 1 extra is provided to backward part of the window.

    # Loss params
    mask_encoder=True,
    # whether to mask encoder padding while computing location sensitive attention. Set to True for better prosody but slower convergence.

    # Decoder
    prenet_layers=[256, 256],  # number of layers and number of units of prenet
    decoder_layers=2,  # number of decoder lstm layers
    decoder_lstm_units=1024,  # number of decoder lstm units on each layer

    dec_prenet_sizes=[256, 256],  # number of layers and number of units of prenet

    # Residual postnet
    postnet_num_layers=5,  # number of postnet convolutional layers
    postnet_kernel_size=(5,),  # size of postnet convolution filters for each layer
    postnet_channels=512,  # number of postnet convolution filters for each layer

    # for linear mel spectrogrma
    post_bank_size=8,
    post_bank_channel_size=128,
    post_maxpool_width=2,
    post_highway_depth=4,
    post_rnn_size=128,
    post_proj_sizes=[256, 80],  # num_mels=80
    post_proj_width=3,

    tacotron_reg_weight=1e-6,  # regularization weight (for L2 regularization)
    inference_prenet_dropout=True,

    # Eval
    min_tokens=30,  # originally 50, 30 is good for korean,  text를 token으로 쪼갰을 때, 최소 길이 이상되어야 train에 사용
    min_n_frame=30 * 5,  # min_n_frame = reduction_factor * min_iters, reduction_factor와 곱해서 min_n_frame을 설정한다.
    max_n_frame=200 * 5,
    skip_inadequate=False,

    griffin_lim_iters=60,
    power=1.5,

)

if hparams.use_lws:
    # Does not work if fft_size is not multiple of hop_size!!
    # sample size = 20480, hop_size=256=12.5ms. fft_size는 window_size를 결정하는데, 2048을 시간으로 환산하면 2048/20480 = 0.1초=100ms
    hparams.sample_rate = 20480  # 

    # shift can be specified by either hop_size(우선) or frame_shift_ms
    hparams.hop_size = 256  # frame_shift_ms = 12.5ms
    hparams.frame_shift_ms = None  # hop_size=  sample_rate *  frame_shift_ms / 1000
    hparams.fft_size = 2048  # 주로 1024로 되어있는데, tacotron에서 2048사용==> output size = 1025
    hparams.win_size = None  # 256x4 --> 50ms



else:
    # 미리 정의되 parameter들로 부터 consistant하게 정의해 준다.
    hparams.num_freq = int(hparams.fft_size / 2 + 1)
    hparams.frame_shift_ms = hparams.hop_size * 1000.0 / hparams.sample_rate  # hop_size=  sample_rate *  frame_shift_ms / 1000
    hparams.frame_length_ms = hparams.win_size * 1000.0 / hparams.sample_rate


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