css-models/notsofar/conformer1.0/mc/config.yaml

# Note that when run on AzureML, train_dir and val_dir are relative to the input Azure Blob Storage container, and
# out_dir is relative to the output container. When run locally, all paths are relative to the project root.
train_dir: ./v1.5/Clean_SCN_P_3Spk_utts_100/train
val_dir: ./v1.5/Clean_SCN_P_3Spk_utts_100/val
out_dir: ./

train_set_cfg:
  sample_frac: 1.0
  max_urls: null  # null means no limit
val_set_cfg:
  sample_frac: 1.0
  max_urls: null  # null means no limit

calc_side_info: True  # New!
log_params_mlflow: True
log_metrics_mlflow: True

scheduler_step_every: [1, iterations]
scheduler_name: step_lr
scheduler_step_lr_cfg:
  # Fixed LR
  step_size: 1
  gamma: 1.0

stop_after: [520000, iterations]
eval_every: [1000, iterations]
save_every: [1000, iterations]

loss_name: 'mask'
base_loss_name: 'l1'

global_batch_size: 256
learning_rate: 1e-4
weight_decay: 1e-2  # according to the paper set to 1e-2

# Large model per CSS with Conformer definition
conformer_css_cfg:
  nnet_conf:
    conformer_conf:
      attention_dim: 512  # default 256
      attention_heads: 8  # default 4
      num_blocks: 18  # default 16
      dropout_rate: 0.0  # New! The default was 0.1.