Zhanzhao (Deo) Liang
2024-12-25 8c7b7e5feb68fda1fc4ddd627bad0f915358149e
funasr/models/seaco_paraformer/template.yaml
@@ -2,8 +2,8 @@
# You can modify the configuration according to your own requirements.
# to print the register_table:
# from funasr.utils.register import registry_tables
# registry_tables.print()
# from funasr.register import tables
# tables.print()
# network architecture
model: SeacoParaformer
@@ -36,7 +36,7 @@
    pos_enc_class: SinusoidalPositionEncoder
    normalize_before: true
    kernel_size: 11
    sanm_shfit: 0
    sanm_shift: 0
    selfattention_layer_type: sanm
# decoder
@@ -51,7 +51,7 @@
    src_attention_dropout_rate: 0.1
    att_layer_num: 16
    kernel_size: 11
    sanm_shfit: 0
    sanm_shift: 0
# seaco decoder
seaco_decoder: ParaformerSANMDecoder
@@ -64,17 +64,22 @@
    self_attention_dropout_rate: 0.1
    src_attention_dropout_rate: 0.1
    kernel_size: 21
    sanm_shfit: 0
    sanm_shift: 0
    use_output_layer: false
    wo_input_layer: true
predictor: CifPredictorV2
predictor: CifPredictorV3
predictor_conf:
    idim: 512
    threshold: 1.0
    l_order: 1
    r_order: 1
    tail_threshold: 0.45
    smooth_factor2: 0.25
    noise_threshold2: 0.01
    upsample_times: 3
    use_cif1_cnn: false
    upsample_type: cnn_blstm
# frontend related
frontend: WavFrontend
@@ -129,7 +134,7 @@
dataset: AudioDataset
dataset_conf:
    index_ds: IndexDSJsonl
    batch_sampler: DynamicBatchLocalShuffleSampler
    batch_sampler: BatchSampler
    batch_type: example # example or length
    batch_size: 1 # 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,