DL-Art-School/codes/models/tacotron2/hparams.py

88 lines
2.6 KiB
Python
Raw Normal View History

2021-07-14 06:08:42 +00:00
#import tensorflow as tf
from models.tacotron2.text import symbols
def create_hparams(hparams_string=None, verbose=False):
"""Create model hyperparameters. Parse nondefault from given string."""
hparams = dict(
################################
# Experiment Parameters #
################################
epochs=500,
iters_per_checkpoint=1000,
seed=1234,
dynamic_loss_scaling=True,
fp16_run=False,
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=['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,
# 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,
grad_clip_thresh=1.0,
batch_size=64,
mask_padding=True # set model's padded outputs to padded values
)
return hparams