configs/autoencoder1d.yaml

model:
  base_learning_rate: 4.5e-06
  target: ldm.models.autoencoder1d.AutoencoderKL
  params:
    embed_dim: 20
    monitor: val/rec_loss
    ddconfig:
      double_z: true
      in_channels: 80
      out_ch: 80
      z_channels: 20
      kernel_size: 5
      ch: 384
      ch_mult:
      - 1
      - 2
      - 4
      num_res_blocks: 2
      attn_layers:
      - 3
      down_layers:
      - 0
      dropout: 0.0
    lossconfig:
      target: ldm.modules.losses_audio.contperceptual.LPAPSWithDiscriminator
      params:
        disc_start: 80001
        perceptual_weight: 0.0
        kl_weight: 1.0e-06
        disc_weight: 0.5
        disc_in_channels: 1
        disc_loss: mse
        disc_factor: 2
        disc_conditional: false
        r1_reg_weight: 3

lightning:
  callbacks:
    image_logger:
      target: main.AudioLogger
      params:
        for_specs: true
        increase_log_steps: false
        batch_frequency: 5000
        max_images: 8
        rescale: false
        melvmin: -5
        melvmax: 1.5
        vocoder_cfg:
          target: vocoder.bigvgan.models.VocoderBigVGAN
          params:
            ckpt_vocoder: vocoder/logs/bigvnat16k93.5w
  trainer:
    sync_batchnorm: false # not working with r1_regularization
    strategy: ddp


data:
  target: main.SpectrogramDataModuleFromConfig
  params:
    batch_size: 4
    num_workers: 16
    spec_dir_path: ldm/data/tsv_dirs/full_data/V1_new
    mel_num: 80
    spec_len: 624
    spec_crop_len: 624
    train:
      target: ldm.data.joinaudiodataset_624.JoinSpecsTrain
      params:
        specs_dataset_cfg: null
    validation:
      target: ldm.data.joinaudiodataset_624.JoinSpecsValidation
      params:
        specs_dataset_cfg: null