嘉渊
2023-05-15 8f21baf63482020397be16db846a533ad8a8731a
egs/aishell2/data2vec_pretrain/conf/train_pretrain_transformer.yaml
@@ -30,19 +30,24 @@
  require_same_masks: true
  mask_dropout: 0
log_interval: 50
normalize: None
# frontend related
frontend: wav_frontend
frontend_conf:
    fs: 16000
    window: hamming
    n_mels: 80
    frame_length: 25
    frame_shift: 10
    lfr_m: 1
    lfr_n: 1
# minibatch related
batch_type: length
batch_bins: 64000
num_workers: 16
model: data2vec
# optimization related
accum_grad: 1
grad_clip: 5
patience: none
max_epoch: 600
max_epoch: 1800
val_scheduler_criterion:
    - valid
    - acc
@@ -68,7 +73,7 @@
dataset_conf:
    batch_mode: clipping
    data_names: speech,none
    data_types: kaldi_ark,none
    data_types: sound,none
    shuffle: true
    shuffle_conf:
        shuffle_size: 12800
@@ -77,3 +82,6 @@
        batch_type: token
        batch_size: 64000
    num_workers: 8
log_interval: 50
normalize: None