| allow_cache: true | |
| batch_max_steps: 8192 | |
| batch_max_steps_valid: 81920 | |
| batch_size: 16 | |
| config: egs/ljspeech/conf/multi_stft_melgan.yaml | |
| dev_dir: ./egs/ljspeech/dump/valid | |
| discriminator_mixed_precision: false | |
| discriminator_optimizer_params: | |
| beta_1: 0.5 | |
| beta_2: 0.9 | |
| decay_steps: 4000000 | |
| end_learning_rate: 0.0001 | |
| initial_learning_rate: 0.0001 | |
| warmup_steps: 0 | |
| weight_decay: 0.0 | |
| discriminator_params: | |
| downsample_pooling: AveragePooling1D | |
| downsample_pooling_params: | |
| pool_size: 4 | |
| strides: 2 | |
| downsample_scales: | |
| - 4 | |
| - 4 | |
| - 4 | |
| - 4 | |
| filters: 16 | |
| is_weight_norm: false | |
| kernel_sizes: | |
| - 5 | |
| - 3 | |
| max_downsample_filters: 1024 | |
| nonlinear_activation: LeakyReLU | |
| nonlinear_activation_params: | |
| alpha: 0.2 | |
| out_channels: 1 | |
| scales: 3 | |
| discriminator_train_start_steps: 100000 | |
| eval_interval_steps: 5000 | |
| format: npy | |
| generator_mixed_precision: false | |
| generator_optimizer_params: | |
| beta_1: 0.5 | |
| beta_2: 0.9 | |
| decay_steps: 100000 | |
| end_learning_rate: 0.0001 | |
| initial_learning_rate: 0.001 | |
| warmup_steps: 0 | |
| weight_decay: 0.0 | |
| generator_params: | |
| filters: 512 | |
| is_weight_norm: false | |
| kernel_size: 7 | |
| out_channels: 1 | |
| stack_kernel_size: 3 | |
| stacks: 3 | |
| upsample_scales: | |
| - 8 | |
| - 8 | |
| - 2 | |
| - 2 | |
| hop_size: 256 | |
| is_shuffle: true | |
| lambda_adv: 4.0 | |
| lambda_feat_match: 10.0 | |
| log_interval_steps: 200 | |
| num_save_intermediate_results: 1 | |
| outdir: ./egs/ljspeech/exp/melgan.stft.v2/ | |
| remove_short_samples: true | |
| resume: egs/ljspeech/exp/melgan.stft.v2/checkpoints/ckpt-100000 | |
| sampling_rate: 22050 | |
| save_interval_steps: 20000 | |
| stft_loss_params: | |
| fft_lengths: | |
| - 1024 | |
| - 2048 | |
| - 512 | |
| frame_lengths: | |
| - 600 | |
| - 1200 | |
| - 240 | |
| frame_steps: | |
| - 120 | |
| - 240 | |
| - 50 | |
| train_dir: ./egs/ljspeech/dump/train/ | |
| train_max_steps: 4000000 | |
| use_norm: true | |
| verbose: 1 | |
| version: 0.3.4 | |