[TTS]【Hackathon + No.190】 + 模型复现:iSTFTNet (#3006)
* iSTFTNet implementation based on hifigan, not affect the function and execution of HIFIGAN * modify the comment in iSTFT.yaml * add the comments in hifigan * iSTFTNet implementation based on hifigan, not affect the function and execution of HIFIGAN * modify the comment in iSTFT.yaml * add the comments in hifigan * add iSTFTNet.md * modify the format of iSTFTNet.md * modify iSTFT.yaml and hifigan.py * Format code using pre-commit * modify hifigan.py,delete the unused self.istft_layer_id , move the self.output_conv behind else, change conv_post to output_conv * update iSTFTNet_csmsc_ckpt.zip download link * modify iSTFTNet.md * modify hifigan.py and iSTFT.yaml * modify iSTFTNet.mdpull/3155/head
parent
72aa19c32c
commit
df37798598
@ -0,0 +1,174 @@
|
||||
# This is the configuration file for CSMSC dataset.
|
||||
# This configuration is based on HiFiGAN V1, which is an official configuration.
|
||||
# But I found that the optimizer setting does not work well with my implementation.
|
||||
# So I changed optimizer settings as follows:
|
||||
# - AdamW -> Adam
|
||||
# - betas: [0.8, 0.99] -> betas: [0.5, 0.9]
|
||||
# - Scheduler: ExponentialLR -> MultiStepLR
|
||||
# To match the shift size difference, the upsample scales is also modified from the original 256 shift setting.
|
||||
|
||||
###########################################################
|
||||
# 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:
|
||||
use_istft: True # Use iSTFTNet.
|
||||
istft_layer_id: 2 # Use istft after istft_layer_id layers of upsample layer if use_istft=True.
|
||||
n_fft: 2048 # FFT size (samples) in feature extraction.
|
||||
win_length: 1200 # Window length (samples) in feature extraction.
|
||||
in_channels: 80 # Number of input channels.
|
||||
out_channels: 1 # Number of output channels.
|
||||
channels: 512 # Number of initial channels.
|
||||
kernel_size: 7 # Kernel size of initial and final conv layers.
|
||||
upsample_scales: [5, 5, 4, 3] # Upsampling scales.
|
||||
upsample_kernel_sizes: [10, 10, 8, 6] # Kernel size for upsampling layers.
|
||||
resblock_kernel_sizes: [3, 7, 11] # Kernel size for residual blocks.
|
||||
resblock_dilations: # Dilations for residual blocks.
|
||||
- [1, 3, 5]
|
||||
- [1, 3, 5]
|
||||
- [1, 3, 5]
|
||||
use_additional_convs: True # Whether to use additional conv layer in residual blocks.
|
||||
bias: True # Whether to use bias parameter in conv.
|
||||
nonlinear_activation: "leakyrelu" # Nonlinear activation type.
|
||||
nonlinear_activation_params: # Nonlinear activation paramters.
|
||||
negative_slope: 0.1
|
||||
use_weight_norm: True # Whether to apply weight normalization.
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
###########################################################
|
||||
# DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
|
||||
###########################################################
|
||||
discriminator_params:
|
||||
scales: 3 # Number of multi-scale discriminator.
|
||||
scale_downsample_pooling: "AvgPool1D" # Pooling operation for scale discriminator.
|
||||
scale_downsample_pooling_params:
|
||||
kernel_size: 4 # Pooling kernel size.
|
||||
stride: 2 # Pooling stride.
|
||||
padding: 2 # Padding size.
|
||||
scale_discriminator_params:
|
||||
in_channels: 1 # Number of input channels.
|
||||
out_channels: 1 # Number of output channels.
|
||||
kernel_sizes: [15, 41, 5, 3] # List of kernel sizes.
|
||||
channels: 128 # Initial number of channels.
|
||||
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
|
||||
max_groups: 16 # Maximum number of groups in downsampling conv layers.
|
||||
bias: True
|
||||
downsample_scales: [4, 4, 4, 4, 1] # Downsampling scales.
|
||||
nonlinear_activation: "leakyrelu" # Nonlinear activation.
|
||||
nonlinear_activation_params:
|
||||
negative_slope: 0.1
|
||||
follow_official_norm: True # Whether to follow the official norm setting.
|
||||
periods: [2, 3, 5, 7, 11] # List of period for multi-period discriminator.
|
||||
period_discriminator_params:
|
||||
in_channels: 1 # Number of input channels.
|
||||
out_channels: 1 # Number of output channels.
|
||||
kernel_sizes: [5, 3] # List of kernel sizes.
|
||||
channels: 32 # Initial number of channels.
|
||||
downsample_scales: [3, 3, 3, 3, 1] # Downsampling scales.
|
||||
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
|
||||
bias: True # Whether to use bias parameter in conv layer."
|
||||
nonlinear_activation: "leakyrelu" # Nonlinear activation.
|
||||
nonlinear_activation_params: # Nonlinear activation paramters.
|
||||
negative_slope: 0.1
|
||||
use_weight_norm: True # Whether to apply weight normalization.
|
||||
use_spectral_norm: False # Whether to apply spectral normalization.
|
||||
|
||||
|
||||
###########################################################
|
||||
# STFT LOSS SETTING #
|
||||
###########################################################
|
||||
use_stft_loss: False # Whether to use multi-resolution STFT loss.
|
||||
use_mel_loss: True # Whether to use Mel-spectrogram loss.
|
||||
mel_loss_params:
|
||||
fs: 24000
|
||||
fft_size: 2048
|
||||
hop_size: 300
|
||||
win_length: 1200
|
||||
window: "hann"
|
||||
num_mels: 80
|
||||
fmin: 0
|
||||
fmax: 12000
|
||||
log_base: null
|
||||
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.
|
||||
use_feat_match_loss: True
|
||||
feat_match_loss_params:
|
||||
average_by_discriminators: False # Whether to average loss by #discriminators.
|
||||
average_by_layers: False # Whether to average loss by #layers in each discriminator.
|
||||
include_final_outputs: False # Whether to include final outputs in feat match loss calculation.
|
||||
|
||||
###########################################################
|
||||
# ADVERSARIAL LOSS SETTING #
|
||||
###########################################################
|
||||
lambda_aux: 45.0 # Loss balancing coefficient for STFT loss.
|
||||
lambda_adv: 1.0 # Loss balancing coefficient for adversarial loss.
|
||||
lambda_feat_match: 2.0 # Loss balancing coefficient for feat match loss..
|
||||
|
||||
###########################################################
|
||||
# DATA LOADER SETTING #
|
||||
###########################################################
|
||||
batch_size: 16 # Batch size.
|
||||
batch_max_steps: 8400 # Length of each audio in batch. Make sure dividable by hop_size.
|
||||
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: 2.0e-4 # Generator's learning rate.
|
||||
gamma: 0.5 # Generator's scheduler gamma.
|
||||
milestones: # At each milestone, lr will be multiplied by gamma.
|
||||
- 200000
|
||||
- 400000
|
||||
- 600000
|
||||
- 800000
|
||||
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 #
|
||||
###########################################################
|
||||
generator_train_start_steps: 1 # Number of steps to start to train discriminator.
|
||||
discriminator_train_start_steps: 0 # Number of steps to start to train discriminator.
|
||||
train_max_steps: 2500000 # 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
|
Loading…
Reference in new issue