Vox2-IIANet.yml

# Network config
audionet:
  audionet_name: IIANet
  audionet_config:
    out_channels: 128
    in_channels: 512
    vpre_channels: 512
    vin_channels: 64
    vout_channels: 64
    num_blocks: 16
    upsampling_depth: 5
    enc_kernel_size: 1 # ms
    num_sources: 1

videonet:
  videonet_name: ResNetVideoModel
  videonet_config:
    pretrain: pretrain_zoo/lrw_resnet18_mstcn_adamw_s3.pth.tar

# Loss config
loss:
  train:
    loss_func: PITLossWrapper
    sdr_type: pairwise_neg_snr
    config:
      pit_from: pw_mtx
      threshold_byloss: false
  val:
    loss_func: PITLossWrapper
    sdr_type: pairwise_neg_sisdr
    config:
      pit_from: pw_mtx
      threshold_byloss: false

# Training config
training:
  system: AudioVisualLightningModule
  gpus: [0,1,2,3,4,5,6,7]
  parallel: ddp
  epochs: 500
  early_stop:
    monitor: val_loss/dataloader_idx_0
    mode: min
    patience: 20
    verbose: true
  
# Optim config
optimizer:
  optim_name: adam
  lr: 0.001
  weight_decay: 0.

# Sche config
scheduler: 
  sche_name: ReduceLROnPlateau
  sche_config:
    patience: 10
    factor: 0.5

# Data config
datamodule:
  data_name: AVSpeechDyanmicDataModule
  data_config:
    train_dir: DataPreProcess/Vox2/tr
    valid_dir: DataPreProcess/Vox2/cv
    test_dir: DataPreProcess/Vox2/tt
    n_src: 1
    sample_rate: 16000
    segment: 2.0
    normalize_audio: false
    batch_size: 3
    num_workers: 24
    pin_memory: true
    persistent_workers: false
  
exp:
  exp_name: Vox2-IIANet