funasr2

游雁

2023-12-19 00ea1186f96e6732e2edb4fab6c0ed6896e3b352

funasr2

 funasr/bin/inference.py

@@ -81,7 +81,7 @@

class AutoModel:
    def __init__(self, **kwargs):
        registry_tables.print_register_tables()
        registry_tables.print()
        assert "model" in kwargs
        if "model_conf" not in kwargs:
            logging.info("download models from model hub: {}".format(kwargs.get("model_hub", "ms")))
@@ -108,6 +108,7 @@
            frontend_class = registry_tables.frontend_classes.get(frontend.lower())
            frontend = frontend_class(**kwargs["frontend_conf"])
            kwargs["frontend"] = frontend
            kwargs["input_size"] = frontend.output_size()
        
        # build model
        model_class = registry_tables.model_classes.get(kwargs["model"].lower())

 funasr/bin/train.py

@@ -39,7 +39,7 @@
    # preprocess_config(kwargs)
    # import pdb; pdb.set_trace()
    # set random seed
    registry_tables.print_register_tables()
    registry_tables.print()
    set_all_random_seed(kwargs.get("seed", 0))
    torch.backends.cudnn.enabled = kwargs.get("cudnn_enabled", torch.backends.cudnn.enabled)
    torch.backends.cudnn.benchmark = kwargs.get("cudnn_benchmark", torch.backends.cudnn.benchmark)
@@ -72,6 +72,7 @@
        frontend_class = registry_tables.frontend_classes.get(frontend.lower())
        frontend = frontend_class(**kwargs["frontend_conf"])
        kwargs["frontend"] = frontend
        kwargs["input_size"] = frontend.output_size()
    
    # import pdb;
    # pdb.set_trace()

 funasr/models/conformer/template.yaml

New file
@@ -0,0 +1,117 @@
# This is an example that demonstrates how to configure a model file.
# 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()

# network architecture
#model: funasr.models.paraformer.model:Paraformer
model: Transformer
model_conf:
    ctc_weight: 0.3
    lsm_weight: 0.1     # label smoothing option
    length_normalized_loss: false

# encoder
encoder: ConformerEncoder
encoder_conf:
    output_size: 256    # dimension of attention
    attention_heads: 4
    linear_units: 2048  # the number of units of position-wise feed forward
    num_blocks: 12      # the number of encoder blocks
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    attention_dropout_rate: 0.0
    input_layer: conv2d # encoder architecture type
    normalize_before: true
    pos_enc_layer_type: rel_pos
    selfattention_layer_type: rel_selfattn
    activation_type: swish
    macaron_style: true
    use_cnn_module: true
    cnn_module_kernel: 15

# decoder
decoder: TransformerDecoder
decoder_conf:
    attention_heads: 4
    linear_units: 2048
    num_blocks: 6
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    self_attention_dropout_rate: 0.0
    src_attention_dropout_rate: 0.0


# frontend related
frontend: WavFrontend
frontend_conf:
    fs: 16000
    window: hamming
    n_mels: 80
    frame_length: 25
    frame_shift: 10
    lfr_m: 1
    lfr_n: 1

specaug: SpecAug
specaug_conf:
    apply_time_warp: true
    time_warp_window: 5
    time_warp_mode: bicubic
    apply_freq_mask: true
    freq_mask_width_range:
    - 0
    - 30
    num_freq_mask: 2
    apply_time_mask: true
    time_mask_width_range:
    - 0
    - 40
    num_time_mask: 2

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

dataset: AudioDataset
dataset_conf:
    index_ds: IndexDSJsonl
    batch_sampler: DynamicBatchLocalShuffleSampler
    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,
    buffer_size: 500
    shuffle: True
    num_workers: 0

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


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

 funasr/models/neat_contextual_paraformer/template.yaml

File was renamed from examples/industrial_data_pretraining/paraformer-large/conf/finetune.yaml
@@ -1,6 +1,12 @@
# This is an example that demonstrates how to configure a model file.
# 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()

# network architecture
model: funasr.cli.models.paraformer:Paraformer
model: NeatContextualParaformer
model_conf:
    ctc_weight: 0.0
    lsm_weight: 0.1
@@ -8,9 +14,10 @@
    predictor_weight: 1.0
    predictor_bias: 1
    sampling_ratio: 0.75
    inner_dim: 512

# encoder
encoder: sanm
encoder: SANMEncoder
encoder_conf:
    output_size: 512
    attention_heads: 4
@@ -26,8 +33,9 @@
    sanm_shfit: 0
    selfattention_layer_type: sanm


# decoder
decoder: paraformer_decoder_sanm
decoder: ContextualParaformerDecoder
decoder_conf:
    attention_heads: 4
    linear_units: 2048
@@ -40,7 +48,7 @@
    kernel_size: 11
    sanm_shfit: 0

predictor: cif_predictor_v2
predictor: CifPredictorV2
predictor_conf:
    idim: 512
    threshold: 1.0
@@ -49,7 +57,7 @@
    tail_threshold: 0.45

# frontend related
frontend: wav_frontend
frontend: WavFrontend
frontend_conf:
    fs: 16000
    window: hamming
@@ -59,7 +67,7 @@
    lfr_m: 7
    lfr_n: 6

specaug: specaug_lfr
specaug: SpecAugLFR
specaug_conf:
    apply_time_warp: false
    time_warp_window: 5
@@ -97,21 +105,22 @@
scheduler_conf:
   warmup_steps: 30000


dataset: AudioDataset
dataset_conf:
    data_names: speech,text
    data_types: sound,text
    index_ds: IndexDSJsonl
    batch_sampler: DynamicBatchLocalShuffleSampler
    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,
    buffer_size: 500
    shuffle: True
    shuffle_conf:
        shuffle_size: 2048
        sort_size: 500
    batch_conf:
        batch_type: example
        batch_size: 2
    num_workers: 8
    num_workers: 0

split_with_space: true
input_size: 560
tokenizer: CharTokenizer
tokenizer_conf:
  unk_symbol: <unk>
  split_with_space: true

ctc_conf:
    dropout_rate: 0.0
    ctc_type: builtin

 funasr/models/paraformer/model.py

@@ -39,8 +39,6 @@
    def __init__(
        self,
        # token_list: Union[Tuple[str, ...], List[str]],
        frontend: Optional[str] = None,
        frontend_conf: Optional[Dict] = None,
        specaug: Optional[str] = None,
        specaug_conf: Optional[Dict] = None,
        normalize: str = None,

 funasr/models/paraformer/template.yaml

@@ -1,6 +1,10 @@
# This is an example that demonstrates how to configure a model file.
# 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()

# network architecture
#model: funasr.models.paraformer.model:Paraformer
model: Paraformer
@@ -117,7 +121,6 @@
  split_with_space: true


input_size: 560
ctc_conf:
    dropout_rate: 0.0
    ctc_type: builtin

 funasr/models/transformer/template.yaml

New file
@@ -0,0 +1,111 @@
# This is an example that demonstrates how to configure a model file.
# 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()

# network architecture
#model: funasr.models.paraformer.model:Paraformer
model: Transformer
model_conf:
    ctc_weight: 0.3
    lsm_weight: 0.1     # label smoothing option
    length_normalized_loss: false

# encoder
encoder: TransformerEncoder
encoder_conf:
    output_size: 256    # dimension of attention
    attention_heads: 4
    linear_units: 2048  # the number of units of position-wise feed forward
    num_blocks: 12      # the number of encoder blocks
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    attention_dropout_rate: 0.0
    input_layer: conv2d # encoder architecture type
    normalize_before: true

# decoder
decoder: TransformerDecoder
decoder_conf:
    attention_heads: 4
    linear_units: 2048
    num_blocks: 6
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    self_attention_dropout_rate: 0.0
    src_attention_dropout_rate: 0.0


# frontend related
frontend: WavFrontend
frontend_conf:
    fs: 16000
    window: hamming
    n_mels: 80
    frame_length: 25
    frame_shift: 10
    lfr_m: 1
    lfr_n: 1

specaug: SpecAug
specaug_conf:
    apply_time_warp: true
    time_warp_window: 5
    time_warp_mode: bicubic
    apply_freq_mask: true
    freq_mask_width_range:
    - 0
    - 30
    num_freq_mask: 2
    apply_time_mask: true
    time_mask_width_range:
    - 0
    - 40
    num_time_mask: 2

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: 0.002
scheduler: warmuplr
scheduler_conf:
   warmup_steps: 30000

dataset: AudioDataset
dataset_conf:
    index_ds: IndexDSJsonl
    batch_sampler: DynamicBatchLocalShuffleSampler
    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,
    buffer_size: 500
    shuffle: True
    num_workers: 0

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


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

 funasr/utils/register.py

@@ -1,6 +1,6 @@
import logging
import inspect
from dataclasses import dataclass, fields
from dataclasses import dataclass


@dataclass
@@ -19,7 +19,7 @@
    dataset_classes = {}
    index_ds_classes = {}

    def print_register_tables(self,):
    def print(self,):
        print("\nregister_tables: \n")
        fields = vars(self)
        for classes_key, classes_dict in fields.items():