PaddleSpeech/examples/csmsc/voc3/conf/finetune.yaml

# This is the hyperparameter configuration file for MelGAN.
# Please make sure this is adjusted for the CSMSC dataset. If you want to
# apply to the other dataset, you might need to carefully change some parameters.
# This configuration requires ~ 8GB memory and will finish within 7 days on Titan V.

# This configuration is based on full-band MelGAN but the hop size and sampling
# rate is different from the paper (16kHz vs 24kHz). The number of iteraions
# is not shown in the paper so currently we train 1M iterations (not sure enough
# to converge). 

###########################################################
#                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: 80               # Number of input channels.
    out_channels: 4               # Number of output channels.
    kernel_size: 7                # Kernel size of initial and final conv layers.
    channels: 384                 # Initial number of channels for conv layers.
    upsample_scales: [5, 5, 3]    # List of Upsampling scales. prod(upsample_scales) x out_channels == n_shift
    stack_kernel_size: 3          # Kernel size of dilated conv layers in residual stack.
    stacks: 4                     # Number of stacks in a single residual stack module.
    use_weight_norm: True         # Whether to use weight normalization.
    use_causal_conv: False        # Whether to use causal convolution.
    use_final_nonlinear_activation: True


###########################################################
#       DISCRIMINATOR NETWORK ARCHITECTURE SETTING        #
###########################################################
discriminator_params:
    in_channels: 1                    # Number of input channels.
    out_channels: 1                   # Number of output channels.
    scales: 3                         # Number of multi-scales.
    downsample_pooling: "AvgPool1D"   # Pooling type for the input downsampling.
    downsample_pooling_params:        # Parameters of the above pooling function.
        kernel_size: 4
        stride: 2
        padding: 1
        exclusive: True
    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.
    downsample_scales: [4, 4, 4]      # 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
use_subband_stft_loss: True
subband_stft_loss_params:
    fft_sizes: [384, 683, 171]  # List of FFT size for STFT-based loss.
    hop_sizes: [30, 60, 10]     # List of hop size for STFT-based loss.
    win_lengths: [150, 300, 60] # List of window length for STFT-based loss.
    window: "hann"              # Window function for STFT-based loss

###########################################################
#               ADVERSARIAL LOSS SETTING                  #
###########################################################
use_feat_match_loss: False # Whether to use feature matching loss.
lambda_adv: 2.5            # Loss balancing coefficient for adversarial loss.

###########################################################
#                  DATA LOADER SETTING                    #
###########################################################
batch_size: 64             # Batch size.
batch_max_steps: 16200     # 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:
    epsilon: 1.0e-7                     # Generator's epsilon.
    weight_decay: 0.0                   # Generator's weight decay coefficient.

generator_grad_norm: -1                 # Generator's gradient norm.
generator_scheduler_params:
    learning_rate: 1.0e-3               # Generator's learning rate.
    gamma: 0.5                          # Generator's scheduler gamma.
    milestones:                         # At each milestone, lr will be multiplied by gamma.
        - 100000
        - 200000
        - 300000
        - 400000
        - 500000
        - 600000
discriminator_optimizer_params:
    epsilon: 1.0e-7                         # Discriminator's epsilon.
    weight_decay: 0.0                       # Discriminator's weight decay coefficient.
  
discriminator_grad_norm: -1                 # Discriminator's gradient norm.
discriminator_scheduler_params:
    learning_rate: 1.0e-3                   # Discriminator's learning rate.
    gamma: 0.5                              # Discriminator's scheduler gamma.
    milestones:                             # At each milestone, lr will be multiplied by gamma.
        - 100000
        - 200000
        - 300000
        - 400000
        - 500000
        - 600000

###########################################################
#                    INTERVAL SETTING                     #
###########################################################
discriminator_train_start_steps: 200000 # Number of steps to start to train discriminator.
train_max_steps: 2000000                # Number of training steps.
save_interval_steps: 1000               # 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