|
| 1 | + |
| 2 | +# This is the hyperparameter configuration file for Multi-Band MelGAN. |
| 3 | +# Please make sure this is adjusted for the LJSpeech dataset. If you want to |
| 4 | +# apply to the other dataset, you might need to carefully change some parameters. |
| 5 | +# This configuration performs 1000k iters. |
| 6 | + |
| 7 | +########################################################### |
| 8 | +# FEATURE EXTRACTION SETTING # |
| 9 | +########################################################### |
| 10 | +sampling_rate: 22050 |
| 11 | +hop_size: 256 # Hop size. |
| 12 | +format: "npy" |
| 13 | + |
| 14 | + |
| 15 | +########################################################### |
| 16 | +# GENERATOR NETWORK ARCHITECTURE SETTING # |
| 17 | +########################################################### |
| 18 | +model_type: "multiband_melgan_generator" |
| 19 | + |
| 20 | +multiband_melgan_generator_params: |
| 21 | + out_channels: 4 # Number of output channels (number of subbands). |
| 22 | + kernel_size: 7 # Kernel size of initial and final conv layers. |
| 23 | + filters: 384 # Initial number of channels for conv layers. |
| 24 | + upsample_scales: [8, 4, 2] # List of Upsampling scales. |
| 25 | + stack_kernel_size: 3 # Kernel size of dilated conv layers in residual stack. |
| 26 | + stacks: 4 # Number of stacks in a single residual stack module. |
| 27 | + is_weight_norm: false # Use weight-norm or not. |
| 28 | + |
| 29 | +########################################################### |
| 30 | +# DISCRIMINATOR NETWORK ARCHITECTURE SETTING # |
| 31 | +########################################################### |
| 32 | +multiband_melgan_discriminator_params: |
| 33 | + out_channels: 1 # Number of output channels. |
| 34 | + scales: 3 # Number of multi-scales. |
| 35 | + downsample_pooling: "AveragePooling1D" # Pooling type for the input downsampling. |
| 36 | + downsample_pooling_params: # Parameters of the above pooling function. |
| 37 | + pool_size: 4 |
| 38 | + strides: 2 |
| 39 | + kernel_sizes: [5, 3] # List of kernel size. |
| 40 | + filters: 16 # Number of channels of the initial conv layer. |
| 41 | + max_downsample_filters: 512 # Maximum number of channels of downsampling layers. |
| 42 | + downsample_scales: [4, 4, 4] # List of downsampling scales. |
| 43 | + nonlinear_activation: "LeakyReLU" # Nonlinear activation function. |
| 44 | + nonlinear_activation_params: # Parameters of nonlinear activation function. |
| 45 | + alpha: 0.2 |
| 46 | + is_weight_norm: false # Use weight-norm or not. |
| 47 | + |
| 48 | +########################################################### |
| 49 | +# STFT LOSS SETTING # |
| 50 | +########################################################### |
| 51 | +stft_loss_params: |
| 52 | + fft_lengths: [1024, 2048, 512] # List of FFT size for STFT-based loss. |
| 53 | + frame_steps: [120, 240, 50] # List of hop size for STFT-based loss |
| 54 | + frame_lengths: [600, 1200, 240] # List of window length for STFT-based loss. |
| 55 | + |
| 56 | +subband_stft_loss_params: |
| 57 | + fft_lengths: [384, 683, 171] # List of FFT size for STFT-based loss. |
| 58 | + frame_steps: [30, 60, 10] # List of hop size for STFT-based loss |
| 59 | + frame_lengths: [150, 300, 60] # List of window length for STFT-based loss. |
| 60 | + |
| 61 | +########################################################### |
| 62 | +# ADVERSARIAL LOSS SETTING # |
| 63 | +########################################################### |
| 64 | +lambda_feat_match: 10.0 # Loss balancing coefficient for feature matching loss |
| 65 | +lambda_adv: 2.5 # Loss balancing coefficient for adversarial loss. |
| 66 | + |
| 67 | +########################################################### |
| 68 | +# DATA LOADER SETTING # |
| 69 | +########################################################### |
| 70 | +batch_size: 64 # Batch size for each GPU with assuming that gradient_accumulation_steps == 1. |
| 71 | +batch_max_steps: 8192 # Length of each audio in batch for training. Make sure dividable by hop_size. |
| 72 | +batch_max_steps_valid: 8192 # Length of each audio for validation. Make sure dividable by hope_size. |
| 73 | +remove_short_samples: true # Whether to remove samples the length of which are less than batch_max_steps. |
| 74 | +allow_cache: true # Whether to allow cache in dataset. If true, it requires cpu memory. |
| 75 | +is_shuffle: true # shuffle dataset after each epoch. |
| 76 | + |
| 77 | +########################################################### |
| 78 | +# OPTIMIZER & SCHEDULER SETTING # |
| 79 | +########################################################### |
| 80 | +generator_optimizer_params: |
| 81 | + lr_fn: "PiecewiseConstantDecay" |
| 82 | + lr_params: |
| 83 | + boundaries: [100000, 200000, 300000, 400000, 500000, 600000, 700000] |
| 84 | + values: [0.0005, 0.0005, 0.00025, 0.000125, 0.0000625, 0.00003125, 0.000015625, 0.000001] |
| 85 | + amsgrad: false |
| 86 | + |
| 87 | +discriminator_optimizer_params: |
| 88 | + lr_fn: "PiecewiseConstantDecay" |
| 89 | + lr_params: |
| 90 | + boundaries: [100000, 200000, 300000, 400000, 500000] |
| 91 | + values: [0.00025, 0.000125, 0.0000625, 0.00003125, 0.000015625, 0.000001] |
| 92 | + |
| 93 | + amsgrad: false |
| 94 | + |
| 95 | +gradient_accumulation_steps: 1 |
| 96 | +########################################################### |
| 97 | +# INTERVAL SETTING # |
| 98 | +########################################################### |
| 99 | +discriminator_train_start_steps: 200000 # steps begin training discriminator |
| 100 | +train_max_steps: 4000000 # Number of training steps. |
| 101 | +save_interval_steps: 20000 # Interval steps to save checkpoint. |
| 102 | +eval_interval_steps: 5000 # Interval steps to evaluate the network. |
| 103 | +log_interval_steps: 200 # Interval steps to record the training log. |
| 104 | + |
| 105 | +########################################################### |
| 106 | +# OTHER SETTING # |
| 107 | +########################################################### |
| 108 | +num_save_intermediate_results: 1 # Number of batch to be saved as intermediate results. |
0 commit comments