audiosep_base.yaml

---
task_name: AudioSep

data:
    datafiles:
        - 'datafiles/template.json'

    sampling_rate: 32000
    segment_seconds: 5
    loudness_norm:
        lower_db: -10
        higher_db: 10
    max_mix_num: 2

model:
    query_net: CLAP
    condition_size: 512
    model_type: ResUNet30
    input_channels: 1
    output_channels: 1
    resume_checkpoint: ""
    use_text_ratio: 1.0

train:
    optimizer:
        optimizer_type: AdamW
        learning_rate: 1e-3
        warm_up_steps: 10000
        reduce_lr_steps: 1000000
        lr_lambda_type: constant_warm_up
    num_nodes: 1
    num_workers: 6
    loss_type: l1_wav
    sync_batchnorm: True
    batch_size_per_device: 12
    steps_per_epoch: 10000  # Every 10000 steps is called an `epoch`.
    evaluate_step_frequency: 10000     # Evaluate every #evaluate_step_frequency steps.
    save_step_frequency: 20000  # Save every #save_step_frequency steps.
    early_stop_steps: 10000001
    random_seed: 1234