Spaces:
Running
Running
| # For De-limit task, Conv-TasNet. | |
| # si_sdr loss | |
| # | |
| # ozone_train_fixed is about 6.36 hours | |
| # 300,000 segments is about 333.33 hours | |
| # ratio should be about 0.019 | |
| wandb_params: | |
| use_wandb: true | |
| entity: null # your wandb id | |
| project: delimit # your wandb project | |
| rerun_id: null # use when you rerun wandb. | |
| sweep: false | |
| sys_params: | |
| nb_workers: 4 | |
| seed: 777 | |
| n_nodes: 1 | |
| port: null | |
| rank: 0 | |
| task_params: | |
| target: all # choices=["all"] | |
| train: true | |
| dataset: delimit # choices=["musdb", "delimit"] | |
| dir_params: | |
| root: /path/to/musdb18hq | |
| output_directory: /path/to/results | |
| exp_name: convtasnet_6_s # you MUST specify this | |
| resume: null # "path of checkpoint folder" | |
| continual_train: false # when we want to use a pre-trained model but not want to use lr_scheduler history. | |
| delimit_valid_root: null | |
| delimit_valid_L_root: null | |
| ozone_root: /path/to/musdb-XL-train # you have to specify data_params.use_fixed | |
| hyperparams: | |
| batch_size: 8 # with 1 gpus (we used 2080ti 11GB) | |
| epochs: 200 | |
| optimizer: adamw | |
| weight_decay: 0.01 | |
| lr: 0.00003 | |
| lr_decay_gamma: 0.5 | |
| lr_decay_patience: 15 | |
| patience: 50 | |
| lr_scheduler: step_lr | |
| gradient_clip: 5.0 | |
| ema: false | |
| data_params: | |
| nfft: 4096 | |
| nhop: 1024 | |
| nb_channels: 2 | |
| sample_rate: 44100 | |
| seq_dur: 4.0 | |
| singleset_num_frames: null | |
| samples_per_track: 128 # "Number of samples per track to use for training." | |
| limitaug_method: ozone | |
| limitaug_mode: null | |
| limitaug_custom_target_lufs: null | |
| limitaug_custom_target_lufs_std: null | |
| target_loudnorm_lufs: -14.0 | |
| random_mix: true | |
| target_limitaug_mode: null | |
| target_limitaug_custom_target_lufs: null | |
| target_limitaug_custom_target_lufs_std: null | |
| custom_limiter_attack_range: null | |
| custom_limiter_release_range: null | |
| use_fixed: 0.019 # range 0.0 ~ 1.0 => 1.0 will use fixed Ozoned_mixture training examples only. | |
| model_loss_params: | |
| architecture: conv_tasnet_mask_on_output # Sample-wise Gain Inversion (SGI) | |
| train_loss_func: [si_sdr] | |
| train_loss_scales: [1.] | |
| valid_loss_func: [si_sdr] | |
| valid_loss_scales: [1.] | |
| conv_tasnet_params: | |
| encoder_activation: relu | |
| n_filters: 512 | |
| kernel_size: 128 # about 3ms in 44100Hz | |
| stride: 64 | |
| n_blocks: 5 | |
| n_repeats: 2 | |
| bn_chan: 128 | |
| hid_chan: 512 | |
| skip_chan: 128 | |
| # conv_kernel_size: | |
| # norm_type: | |
| mask_act: relu | |
| # causal: | |
| decoder_activation: sigmoid | |