aishell example

游雁

2024-02-19 1448e021accfdb03a381651cb5a8be6d1a6e8adf

 examples/aishell/conf/train_asr_paraformer_conformer_12e_6d_2048_256.yaml

@@ -1,6 +1,6 @@

# network architecture
model: funasr.cli.models.paraformer:Paraformer
model: Paraformer
model_conf:
    ctc_weight: 0.3
    lsm_weight: 0.1
@@ -9,9 +9,8 @@
    sampling_ratio: 0.4
    use_1st_decoder_loss: true


# encoder related
encoder: conformer
# encoder
encoder: ConformerEncoder
encoder_conf:
    output_size: 256    # dimension of attention
    attention_heads: 4
@@ -29,8 +28,8 @@
    use_cnn_module: true
    cnn_module_kernel: 15

# decoder related
decoder: paraformer_decoder_san
# decoder
decoder: ParaformerSANDecoder
decoder_conf:
    attention_heads: 4
    linear_units: 2048
@@ -40,8 +39,17 @@
    self_attention_dropout_rate: 0.0
    src_attention_dropout_rate: 0.0

# predictor
predictor: CifPredictor
predictor_conf:
    idim: 256
    threshold: 1.0
    l_order: 1
    r_order: 1
    tail_threshold: 0.45

# frontend related
frontend: wav_frontend
frontend: WavFrontend
frontend_conf:
    fs: 16000
    window: hamming
@@ -51,29 +59,7 @@
    lfr_m: 1
    lfr_n: 1


train_conf:
  accum_grad: 1
  grad_clip: 5
  max_epoch: 150
  val_scheduler_criterion:
      - valid
      - acc
  best_model_criterion:
  -   - valid
      - acc
      - max
  keep_nbest_models: 10
  log_interval: 50

optim: adam
optim_conf:
   lr: 0.0005
scheduler: warmuplr
scheduler_conf:
   warmup_steps: 30000

specaug: specaug
specaug: SpecAug
specaug_conf:
    apply_time_warp: true
    time_warp_window: 5
@@ -89,25 +75,43 @@
    - 40
    num_time_mask: 2

predictor: cif_predictor
predictor_conf:
    idim: 256
    threshold: 1.0
    l_order: 1
    r_order: 1
    tail_threshold: 0.45
train_conf:
  accum_grad: 1
  grad_clip: 5
  max_epoch: 150
  keep_nbest_models: 10
  avg_nbest_model: 5
  log_interval: 50

optim: adam
optim_conf:
   lr: 0.0005
scheduler: warmuplr
scheduler_conf:
   warmup_steps: 30000

dataset: AudioDataset
dataset_conf:
    data_names: speech,text
    data_types: sound,text
    index_ds: IndexDSJsonl
    batch_sampler: RankFullLocalShuffleBatchSampler
    batch_type: example # example or length
    batch_size: 32 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
    max_token_length: 2048 # filter samples if source_token_len+target_token_len > max_token_length,
    buffer_size: 1024
    shuffle: True
    shuffle_conf:
        shuffle_size: 2048
        sort_size: 500
    batch_conf:
        batch_type: example
        batch_size: 2
    num_workers: 8
    num_workers: 0

tokenizer: CharTokenizer
tokenizer_conf:
  unk_symbol: <unk>
  split_with_space: true


normalize: null
ctc_conf:
    dropout_rate: 0.0
    ctc_type: builtin
    reduce: true
    ignore_nan_grad: true
normalize: null