checkpoints/AV_MossFormer2_TSE_16K/config.yaml

## Config file

# Log 
seed: 777
use_cuda: 1           # 1 for True, 0 for False

# dataset
speaker_no: 2
mix_lst_path: ./data/allData/voxceleb2/mixture_data_list_2mix_pretrain.csv
audio_direc: /mnt/nas_sg/wulanchabu/zexu.pan/datasets/voxceleb2/audio_clean
reference_direc: /mnt/nas_sg/wulanchabu/zexu.pan/datasets/ # not used
audio_sr: 16000
ref_sr: 25

# dataloader
num_workers: 4
batch_size: 2         # 4-GPU training with a total effective batch size of 8
accu_grad: 0
effec_batch_size: 4   # per GPU, only used if accu_grad is set to 1, must be multiple times of batch size
max_length: 5         # truncate the utterances in dataloader, in seconds 

# network settings
init_from: checkpoints/log_2024-09-30(09:49:14)       # 'None' or a log name 'log_2024-07-22(18:12:13)'
causal: 0             # 1 for True, 0 for False
network_reference:
  cue: lip            # lip or speech or gesture or EEG
  backbone: resnet18  # resnet18 or shufflenetV2 or blazenet64
  emb_size: 256       # resnet18:256
network_audio:
  backbone: mossformer2
  encoder_kernel_size: 16
  encoder_out_nchannels: 512
  encoder_in_nchannels: 1
  
  masknet_numspks: 1
  masknet_chunksize: 250
  masknet_numlayers: 1
  masknet_norm: "ln"
  masknet_useextralinearlayer: False
  masknet_extraskipconnection: True

  intra_numlayers: 24
  intra_nhead: 8
  intra_dffn: 1024
  intra_dropout: 0
  intra_use_positional: True
  intra_norm_before: True
        

# optimizer
loss_type: hybrid      # "snr", "sisdr", "hybrid"
init_learning_rate: 0.00015
max_epoch: 150
clip_grad_norm: 5