|
| 1 | +# This is the hyperparameter configuration file for Parallel WaveGAN. |
| 2 | +# Please make sure this is adjusted for the VCTK corpus. If you want to |
| 3 | +# apply to the other dataset, you might need to carefully change some parameters. |
| 4 | +# This configuration requires 12 GB GPU memory and takes ~3 days on RTX TITAN. |
| 5 | + |
| 6 | +########################################################### |
| 7 | +# FEATURE EXTRACTION SETTING # |
| 8 | +########################################################### |
| 9 | +fs: 24000 # Sampling rate. |
| 10 | +n_fft: 2048 # FFT size. (in samples) |
| 11 | +n_shift: 300 # Hop size. (in samples) |
| 12 | +win_length: 1200 # Window length. (in samples) |
| 13 | + # If set to null, it will be the same as fft_size. |
| 14 | +window: "hann" # Window function. |
| 15 | +n_mels: 80 # Number of mel basis. |
| 16 | +fmin: 80 # Minimum freq in mel basis calculation. (Hz) |
| 17 | +fmax: 7600 # Maximum frequency in mel basis calculation. (Hz) |
| 18 | + |
| 19 | +########################################################### |
| 20 | +# GENERATOR NETWORK ARCHITECTURE SETTING # |
| 21 | +########################################################### |
| 22 | +generator_params: |
| 23 | + in_channels: 1 # Number of input channels. |
| 24 | + out_channels: 1 # Number of output channels. |
| 25 | + kernel_size: 3 # Kernel size of dilated convolution. |
| 26 | + layers: 30 # Number of residual block layers. |
| 27 | + stacks: 3 # Number of stacks i.e., dilation cycles. |
| 28 | + residual_channels: 64 # Number of channels in residual conv. |
| 29 | + gate_channels: 128 # Number of channels in gated conv. |
| 30 | + skip_channels: 64 # Number of channels in skip conv. |
| 31 | + aux_channels: 80 # Number of channels for auxiliary feature conv. |
| 32 | + # Must be the same as num_mels. |
| 33 | + aux_context_window: 2 # Context window size for auxiliary feature. |
| 34 | + # If set to 2, previous 2 and future 2 frames will be considered. |
| 35 | + dropout: 0.0 # Dropout rate. 0.0 means no dropout applied. |
| 36 | + use_weight_norm: true # Whether to use weight norm. |
| 37 | + # If set to true, it will be applied to all of the conv layers. |
| 38 | + upsample_scales: [4, 5, 3, 5] # Upsampling scales. Prodcut of these must be the same as hop size. |
| 39 | + |
| 40 | +########################################################### |
| 41 | +# DISCRIMINATOR NETWORK ARCHITECTURE SETTING # |
| 42 | +########################################################### |
| 43 | +discriminator_params: |
| 44 | + in_channels: 1 # Number of input channels. |
| 45 | + out_channels: 1 # Number of output channels. |
| 46 | + kernel_size: 3 # Number of output channels. |
| 47 | + layers: 10 # Number of conv layers. |
| 48 | + conv_channels: 64 # Number of chnn layers. |
| 49 | + bias: true # Whether to use bias parameter in conv. |
| 50 | + use_weight_norm: true # Whether to use weight norm. |
| 51 | + # If set to true, it will be applied to all of the conv layers. |
| 52 | + nonlinear_activation: "LeakyReLU" # Nonlinear function after each conv. |
| 53 | + nonlinear_activation_params: # Nonlinear function parameters |
| 54 | + negative_slope: 0.2 # Alpha in LeakyReLU. |
| 55 | + |
| 56 | +########################################################### |
| 57 | +# STFT LOSS SETTING # |
| 58 | +########################################################### |
| 59 | +stft_loss_params: |
| 60 | + fft_sizes: [1024, 2048, 512] # List of FFT size for STFT-based loss. |
| 61 | + hop_sizes: [120, 240, 50] # List of hop size for STFT-based loss |
| 62 | + win_lengths: [600, 1200, 240] # List of window length for STFT-based loss. |
| 63 | + window: "hann" # Window function for STFT-based loss |
| 64 | + |
| 65 | +########################################################### |
| 66 | +# ADVERSARIAL LOSS SETTING # |
| 67 | +########################################################### |
| 68 | +lambda_adv: 4.0 # Loss balancing coefficient. |
| 69 | + |
| 70 | +########################################################### |
| 71 | +# DATA LOADER SETTING # |
| 72 | +########################################################### |
| 73 | +batch_size: 8 # Batch size. |
| 74 | +batch_max_steps: 24000 # Length of each audio in batch. Make sure dividable by hop_size. |
| 75 | +pin_memory: true # Whether to pin memory in Pytorch DataLoader. |
| 76 | +num_workers: 4 # Number of workers in Pytorch DataLoader. |
| 77 | +remove_short_samples: true # Whether to remove samples the length of which are less than batch_max_steps. |
| 78 | +allow_cache: true # Whether to allow cache in dataset. If true, it requires cpu memory. |
| 79 | + |
| 80 | +########################################################### |
| 81 | +# OPTIMIZER & SCHEDULER SETTING # |
| 82 | +########################################################### |
| 83 | +generator_optimizer_params: |
| 84 | + epsilon: 1.0e-6 # Generator's epsilon. |
| 85 | + weight_decay: 0.0 # Generator's weight decay coefficient. |
| 86 | +generator_scheduler_params: |
| 87 | + learning_rate: 0.0001 # Generator's learning rate. |
| 88 | + step_size: 200000 # Generator's scheduler step size. |
| 89 | + gamma: 0.5 # Generator's scheduler gamma. |
| 90 | + # At each step size, lr will be multiplied by this parameter. |
| 91 | +generator_grad_norm: 10 # Generator's gradient norm. |
| 92 | +discriminator_optimizer_params: |
| 93 | + epsilon: 1.0e-6 # Discriminator's epsilon. |
| 94 | + weight_decay: 0.0 # Discriminator's weight decay coefficient. |
| 95 | +discriminator_scheduler_params: |
| 96 | + learning_rate: 0.00005 # Discriminator's learning rate. |
| 97 | + step_size: 200000 # Discriminator's scheduler step size. |
| 98 | + gamma: 0.5 # Discriminator's scheduler gamma. |
| 99 | + # At each step size, lr will be multiplied by this parameter. |
| 100 | +discriminator_grad_norm: 1 # Discriminator's gradient norm. |
| 101 | + |
| 102 | +########################################################### |
| 103 | +# INTERVAL SETTING # |
| 104 | +########################################################### |
| 105 | +discriminator_train_start_steps: 100000 # Number of steps to start to train discriminator. |
| 106 | +train_max_steps: 1000000 # Number of training steps. |
| 107 | +save_interval_steps: 5000 # Interval steps to save checkpoint. |
| 108 | +eval_interval_steps: 1000 # Interval steps to evaluate the network. |
| 109 | + |
| 110 | +########################################################### |
| 111 | +# OTHER SETTING # |
| 112 | +########################################################### |
| 113 | +num_save_intermediate_results: 4 # Number of results to be saved as intermediate results. |
| 114 | +num_snapshots: 10 # max number of snapshots to keep while training |
| 115 | +seed: 42 # random seed for paddle, random, and np.random |
0 commit comments