train_config.yaml

# Generated 2023-05-14 from:
# /home/agorin/cryceleb2023/hparams/ecapa_voxceleb_basic.yaml
# yamllint disable
# ################################
# Model: Speaker identification with ECAPA for CryCeleb
# Authors: David Budaghyan
# ################################

ckpt_interval_minutes: 15 # save checkpoint every N min

##### SEED
seed: 3011
__set_seed: !apply:crybrain_config_utils.set_seed [3011]

# DataLoader
bs: 32
train_dataloader_options:
  batch_size: 32
  shuffle: true
val_dataloader_options:
  batch_size: 2
  shuffle: false

##### ESTIMATOR COMPONENTS
# Fbank (feature extractor)
n_mels: 80
left_frames: 0
right_frames: 0
deltas: false
compute_features: &id002 !new:speechbrain.lobes.features.Fbank
  n_mels: 80
  left_frames: 0
  right_frames: 0
  deltas: false

# ECAPA
emb_dim: 192
embedding_model: &id001 !new:speechbrain.lobes.models.ECAPA_TDNN.ECAPA_TDNN
  input_size: 80
  channels: [1024, 1024, 1024, 1024, 3072]
  kernel_sizes: [5, 3, 3, 3, 1]
  dilations: [1, 2, 3, 4, 1]
  groups: [1, 1, 1, 1, 1]
  attention_channels: 128
  lin_neurons: 192

# If you do not want to use the pretrained encoder you can simply delete pretrained_encoder field.
pretrained_model_name: spkrec-ecapa-voxceleb
pretrained_embedding_model_path: speechbrain/spkrec-ecapa-voxceleb/embedding_model.ckpt
pretrained_embedding_model: !new:speechbrain.utils.parameter_transfer.Pretrainer
  collect_in: ./experiments/ecapa_voxceleb_ft_basic/ckpts
  loadables:
    model: *id001
  paths:
    model: speechbrain/spkrec-ecapa-voxceleb/embedding_model.ckpt

# CLASSIFIER
n_classes: 348
                        # check-yaml disable


classifier: &id003 !new:speechbrain.lobes.models.ECAPA_TDNN.Classifier
  input_size: 192
  out_neurons: 348

##### EPOCH COUNTER
n_epochs: 1000
epoch_counter: &id005 !new:speechbrain.utils.epoch_loop.EpochCounter
  limit: 1000

##### OPTIMIZER
start_lr: 0.0001
opt_class: !name:torch.optim.Adam
  lr: 0.0001
  weight_decay: 0.000002

#####  LEARNING RATE SCHEDULERS
lrsched_name: cyclic
# one of:
#   onplateau
#   cyclic
lr_min: 0.0000000001
lr_scheduler: &id006 !apply:crybrain_config_utils.choose_lrsched
  lrsched_name: cyclic
  #below are kwargs, only the ones relevant to the type of scheduler will be
  #used for initialization in `choose_lrsched`

  #onplateau (ReduceLROnPlateau)
  lr_min: 0.0000000001
  factor: 0.4
  patience: 10
  dont_halve_until_epoch: 35
  #cyclic (CyclicLRScheduler)
  base_lr: 0.00000001
  max_lr: 0.0001
  step_size: 100
  mode: triangular
  gamma: 1.0
  scale_fn:
  scale_mode: cycle

sample_rate: 16000
mean_var_norm: &id004 !new:speechbrain.processing.features.InputNormalization

  norm_type: sentence
  std_norm: false

modules:
  compute_features: *id002
  embedding_model: *id001
  classifier: *id003
  mean_var_norm: *id004
compute_cost: !new:speechbrain.nnet.losses.LogSoftmaxWrapper
  loss_fn: !new:speechbrain.nnet.losses.AdditiveAngularMargin
    margin: 0.2
    scale: 30

classification_stats: !name:speechbrain.utils.metric_stats.ClassificationStats
  ###################################################################
  ### OUTPUT PATHS ###


experiment_name: ecapa_voxceleb_ft_basic
               # must run from the directory which contains "experiments"


experiment_dir: ./experiments/ecapa_voxceleb_ft_basic
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: ./experiments/ecapa_voxceleb_ft_basic/train_log.txt

checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: ./experiments/ecapa_voxceleb_ft_basic/ckpts
  recoverables:
    embedding_model: *id001
    classifier: *id003
    normalizer: *id004
    counter: *id005
    lr_scheduler: *id006