You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
PaddleSpeech/examples/csmsc/voc4/conf/default.yaml

137 lines
6.8 KiB

# This is the configuration file for CSMSC dataset.This configuration is based
# on StyleMelGAN paper but uses MSE loss instead of Hinge loss. And I found that
# batch_size = 8 is also working good. So maybe if you want to accelerate the training,
# you can reduce the batch size (e.g. 8 or 16). Upsampling scales is modified to
# fit the shift size 300 pt.
# NOTE: batch_max_steps(24000) == prod(noise_upsample_scales)(80) * prod(upsample_scales)(300)
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # Sampling rate.
n_fft: 2048 # FFT size (samples).
n_shift: 300 # Hop size (samples). 12.5ms
win_length: 1200 # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation. (Hz)
fmax: 7600 # Maximum frequency in mel basis calculation. (Hz)
###########################################################
# GENERATOR NETWORK ARCHITECTURE SETTING #
###########################################################
generator_params:
in_channels: 128 # Number of input channels.
aux_channels: 80
channels: 64 # Initial number of channels for conv layers.
out_channels: 1 # Number of output channels.
kernel_size: 9 # Kernel size of initial and final conv layers.
dilation: 2
bias: True
noise_upsample_scales: [10, 2, 2, 2]
noise_upsample_activation: "leakyrelu"
noise_upsample_activation_params:
negative_slope: 0.2
upsample_scales: [5, 1, 5, 1, 3, 1, 2, 2, 1] # List of Upsampling scales. prod(upsample_scales) == n_shift
upsample_mode: "nearest"
gated_function: "softmax"
use_weight_norm: True # Whether to use weight normalization.
###########################################################
# DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
###########################################################
discriminator_params:
repeats: 4
window_sizes: [512, 1024, 2048, 4096]
pqmf_params:
- [1, None, None, None]
- [2, 62, 0.26700, 9.0]
- [4, 62, 0.14200, 9.0]
- [8, 62, 0.07949, 9.0]
discriminator_params:
out_channels: 1 # Number of output channels.
kernel_sizes: [5, 3] # List of kernel size.
channels: 16 # Number of channels of the initial conv layer.
max_downsample_channels: 512 # Maximum number of channels of downsampling layers.
bias: True
downsample_scales: [4, 4, 4, 1] # List of downsampling scales.
nonlinear_activation: "leakyrelu" # Nonlinear activation function.
nonlinear_activation_params: # Parameters of nonlinear activation function.
negative_slope: 0.2
use_weight_norm: True # Whether to use weight norm.
###########################################################
# STFT LOSS SETTING #
###########################################################
use_stft_loss: True
stft_loss_params:
fft_sizes: [1024, 2048, 512] # List of FFT size for STFT-based loss.
hop_sizes: [120, 240, 50] # List of hop size for STFT-based loss
win_lengths: [600, 1200, 240] # List of window length for STFT-based loss.
window: "hann" # Window function for STFT-based loss
lambda_aux: 1.0 # Loss balancing coefficient for aux loss.
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
lambda_adv: 1.0 # Loss balancing coefficient for adv loss.
generator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
discriminator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 32 # Batch size.
# batch_max_steps(24000) == prod(noise_upsample_scales)(80) * prod(upsample_scales)(300, n_shift)
batch_max_steps: 24000 # Length of each audio in batch. Make sure dividable by n_shift.
num_workers: 2 # Number of workers in DataLoader.
###########################################################
# OPTIMIZER & SCHEDULER SETTING #
###########################################################
generator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Generator's weight decay coefficient.
generator_scheduler_params:
learning_rate: 1.0e-4 # Generator's learning rate.
gamma: 0.5 # Generator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 100000
- 300000
- 500000
- 700000
- 900000
generator_grad_norm: -1 # Generator's gradient norm.
discriminator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Discriminator's weight decay coefficient.
discriminator_scheduler_params:
learning_rate: 2.0e-4 # Discriminator's learning rate.
gamma: 0.5 # Discriminator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
discriminator_grad_norm: -1 # Discriminator's gradient norm.
###########################################################
# INTERVAL SETTING #
###########################################################
discriminator_train_start_steps: 100000 # Number of steps to start to train discriminator.
train_max_steps: 1500000 # Number of training steps.
save_interval_steps: 5000 # Interval steps to save checkpoint.
eval_interval_steps: 1000 # Interval steps to evaluate the network.
###########################################################
# OTHER SETTING #
###########################################################
num_snapshots: 10 # max number of snapshots to keep while training
seed: 42 # random seed for paddle, random, and np.random