| seed: 46762398 | |
| ckpt_path: null | |
| train: true | |
| test: false | |
| path: | |
| exp_root: exp | |
| figures: figures | |
| version_base: null | |
| datamodule: | |
| _target_: open_universe.datasets.DataModule | |
| train: | |
| dataset: vb-train-16k | |
| dl_opts: | |
| pin_memory: true | |
| num_workers: 6 | |
| shuffle: true | |
| batch_size: 10 | |
| val: | |
| dataset: vb-val-16k | |
| dl_opts: | |
| pin_memory: true | |
| num_workers: 4 | |
| shuffle: false | |
| batch_size: 1 | |
| test: | |
| dataset: vb-test-16k | |
| dl_opts: | |
| pin_memory: true | |
| num_workers: 4 | |
| shuffle: false | |
| batch_size: 1 | |
| datasets: | |
| vb-train-16k: | |
| _target_: open_universe.datasets.NoisyDataset | |
| audio_path: data/voicebank_demand/16k | |
| fs: 16000 | |
| split: train | |
| audio_len: 2.0 | |
| vb-val-16k: | |
| _target_: open_universe.datasets.NoisyDataset | |
| audio_path: ${..vb-train-16k.audio_path} | |
| fs: ${..vb-train-16k.fs} | |
| split: val | |
| audio_len: null | |
| vb-test-16k: | |
| _target_: open_universe.datasets.NoisyDataset | |
| audio_path: ${..vb-train-16k.audio_path} | |
| fs: ${..vb-train-16k.fs} | |
| split: test | |
| audio_len: null | |
| vb-train-24k: | |
| _target_: open_universe.datasets.NoisyDataset | |
| audio_path: data/voicebank_demand/24k | |
| fs: 24000 | |
| split: train | |
| audio_len: 2.0 | |
| vb-val-24k: | |
| _target_: open_universe.datasets.NoisyDataset | |
| audio_path: ${..vb-train-24k.audio_path} | |
| fs: ${..vb-train-24k.fs} | |
| split: val | |
| audio_len: null | |
| vb-test-24k: | |
| _target_: open_universe.datasets.NoisyDataset | |
| audio_path: ${..vb-train-24k.audio_path} | |
| fs: ${..vb-train-24k.fs} | |
| split: test | |
| audio_len: null | |
| model: | |
| _target_: open_universe.networks.universe.UniverseGAN | |
| fs: 16000 | |
| normalization_norm: 2 | |
| normalization_kwargs: | |
| ref: both | |
| level_db: -26.0 | |
| edm: | |
| noise: 0.25 | |
| score_model: | |
| _target_: open_universe.networks.universe.ScoreNetwork | |
| fb_kernel_size: 3 | |
| rate_factors: | |
| - 2 | |
| - 4 | |
| - 4 | |
| - 5 | |
| n_channels: 32 | |
| n_rff: 32 | |
| noise_cond_dim: 512 | |
| encoder_gru_conv_sandwich: false | |
| extra_conv_block: true | |
| decoder_act_type: prelu | |
| use_weight_norm: true | |
| use_antialiasing: true | |
| time_embedding: simple | |
| condition_model: | |
| _target_: open_universe.networks.universe.ConditionerNetwork | |
| fb_kernel_size: ${model.score_model.fb_kernel_size} | |
| rate_factors: ${model.score_model.rate_factors} | |
| n_channels: ${model.score_model.n_channels} | |
| n_mels: 80 | |
| n_mel_oversample: 4 | |
| encoder_gru_residual: true | |
| extra_conv_block: ${model.score_model.extra_conv_block} | |
| decoder_act_type: prelu | |
| use_weight_norm: ${model.score_model.use_weight_norm} | |
| use_antialiasing: false | |
| diffusion: | |
| schedule: geometric | |
| sigma_min: 0.0005 | |
| sigma_max: 5.0 | |
| n_steps: 8 | |
| epsilon: 1.3 | |
| losses: | |
| multi_period_discriminator: | |
| mpd_reshapes: | |
| - 2 | |
| - 3 | |
| - 5 | |
| - 7 | |
| - 11 | |
| use_spectral_norm: false | |
| discriminator_channel_mult: 1 | |
| multi_resolution_discriminator: | |
| resolutions: | |
| - - 1024 | |
| - 120 | |
| - 600 | |
| - - 2048 | |
| - 240 | |
| - 1200 | |
| - - 512 | |
| - 50 | |
| - 240 | |
| use_spectral_norm: false | |
| discriminator_channel_mult: 1 | |
| disc_freeze_step: 0 | |
| weights: | |
| mel_l1: 45.0 | |
| score: 1.0 | |
| use_signal_decoupling: true | |
| signal_decoupling_act: snake | |
| score_loss: | |
| _target_: torch.nn.MSELoss | |
| training: | |
| audio_len: ${datamodule.datasets.vb-train-16k.audio_len} | |
| time_sampling: time_normal_0.95 | |
| dynamic_mixing: false | |
| ema_decay: 0.999 | |
| validation: | |
| main_loss: val/pesq | |
| main_loss_mode: max | |
| n_bins: 5 | |
| max_enh_batches: 4 | |
| enh_losses: | |
| val/: | |
| _target_: open_universe.metrics.EvalMetrics | |
| audio_fs: ${model.fs} | |
| optimizer: | |
| accumulate_grad_batches: 1 | |
| generator: | |
| _target_: torch.optim.AdamW | |
| lr: 0.0002 | |
| weight_decay: 0.01 | |
| betas: | |
| - 0.8 | |
| - 0.99 | |
| weight_decay_exclude: | |
| - prelu | |
| - bias | |
| discriminator: | |
| _target_: torch.optim.AdamW | |
| lr: 0.0002 | |
| betas: | |
| - 0.8 | |
| - 0.99 | |
| grad_clip_vals: | |
| mrd: 1000.0 | |
| mpd: 1000.0 | |
| score: 1000.0 | |
| cond: 1000.0 | |
| scheduler: | |
| generator: | |
| scheduler: | |
| _target_: open_universe.utils.schedulers.LinearWarmupCosineAnnealingLR | |
| T_warmup: 20000 | |
| T_cosine: 400000 | |
| eta_min: 1.6e-06 | |
| T_max: ${trainer.max_steps} | |
| interval: step | |
| frequency: 1 | |
| discriminator: | |
| scheduler: | |
| _target_: open_universe.utils.schedulers.LinearWarmupCosineAnnealingLR | |
| T_warmup: 20000 | |
| T_cosine: 400000 | |
| eta_min: 1.6e-06 | |
| T_max: ${trainer.max_steps} | |
| interval: step | |
| frequency: 1 | |
| grad_clipper: | |
| _target_: open_universe.utils.FixedClipper | |
| max_norm: 1000.0 | |
| trainer: | |
| _target_: pytorch_lightning.Trainer | |
| accumulate_grad_batches: 1 | |
| min_epochs: 1 | |
| max_epochs: -1 | |
| max_steps: 600000 | |
| deterministic: warn | |
| accelerator: gpu | |
| devices: -1 | |
| strategy: ddp_find_unused_parameters_true | |
| check_val_every_n_epoch: null | |
| val_check_interval: 5000 | |
| default_root_dir: . | |
| profiler: false | |