hparams.py

from utils.hparams_repo import HParamsRepo

hparams = HParamsRepo(
    # Audio
    num_mels=80,
    num_freq=1025,
    sample_rate=24000,
    frame_length_ms=50,
    frame_shift_ms=12.5,
    preemphasis=0.97,
    min_level_db=-100,
    ref_level_db=20,
    max_abs_value=1.0,
    symmetric_specs=False,
    bits=9,
    mu_law=True,

    # Model
    mode = 'RAW', # RAW or MOL
    upsample_factors=(5, 6, 10),
    rnn_dims=512,
    fc_dims=512,
    compute_dims=128,
    res_out_dims=128,
    res_blocks=10,

    # Data loader
    pin_memory=True,
    num_workers=16,
    
    # Training
    batch_size=32,
    lr=1e-3,
    checkpoint_interval=10000,
    gen_at_checkpoint=5,          # number of samples to generate at each checkpoint
    nepochs=10000,                 # Total number of training steps
    seq_len=5,                 # must be a multiple of hop_length
    pad=2,

    # Optimizer
    adam_beta1=0.9,
    adam_beta2=0.999,
    weight_decay=0.0,
    clip_thresh=1.0,
    
    # Generate
    batched=False,                 # very fast (realtime+) single utterance batched generation
    target=8000,                 # target number of samples to be generated in each batch entry
    overlap=400,                  # number of samples for crossfading between batches
    power=1.5,
    griffin_lim_iters=60,
)