espnet2.asr.encoder.beats_encoder.BeatsConfig

source

class espnet2.asr.encoder.beats_encoder.BeatsConfig(cfg=None)

Bases: object

Configuration class for the BEATs encoder model.

This class defines the various hyperparameters and configuration options for the BEATs model used in audio pre-training with acoustic tokenizers. The default values are set in the constructor, but they can be updated using the update method.

input_patch_size

Patch size for patch embedding.

• Type: int

embed_dim

Dimension of patch embedding.

• Type: int

conv_bias

Whether to include bias in the convolutional encoder.

• Type: bool

encoder_layers

Number of encoder layers in the transformer.

• Type: int

encoder_embed_dim

Encoder embedding dimension.

• Type: int

encoder_ffn_embed_dim

Feed-forward network embedding dimension.

• Type: int

encoder_attention_heads

Number of attention heads in the encoder.

• Type: int

activation_fn

Activation function used in the model.

• Type: str

layer_wise_gradient_decay_ratio

Ratio for layer-wise gradient decay.

• Type: float

layer_norm_first

Whether to apply layer normalization first.

• Type: bool

deep_norm

Whether to apply deep normalization first.

• Type: bool

dropout

Dropout probability for the transformer.

• Type: float

attention_dropout

Dropout probability for attention weights.

• Type: float

activation_dropout

Dropout probability after activation in FFN.

• Type: float

encoder_layerdrop

Probability of dropping a transformer layer.

• Type: float

dropout

_input

Dropout to apply to the input after feature extraction.

• Type: float

conv_pos

Number of filters for convolutional positional embeddings.

• Type: int

conv_pos

_groups

Number of groups for convolutional positional embedding.

• Type: int

relative_position_embedding

Whether to apply relative position embedding.

• Type: bool

num_buckets

Number of buckets for relative position embedding.

• Type: int

max_distance

Maximum distance for relative position embedding.

• Type: int

gru_rel_pos

Whether to apply gated relative position embedding.

• Type: bool

finetuned_model

Indicates if the model is fine-tuned.

• Type: bool

predictor_dropout

Dropout probability for the predictor.

• Type: float

predictor_class

Target class number for the predictor.

• Type: int

• Parameters:cfg (dict , optional) – Configuration dictionary to update the default values.

####### Examples

Create a default configuration

config = BeatsConfig()

Create a configuration with custom settings

custom_config = BeatsConfig(cfg={

‘input_patch_size’: 32, ‘dropout’: 0.2, ‘finetuned_model’: True

})

update(cfg: dict)

Update the configuration of the BeatsConfig instance.

This method updates the attributes of the BeatsConfig instance with values from the provided configuration dictionary. It modifies the instance’s internal state directly by updating its __dict__ attribute.

• Parameters:cfg (dict) – A dictionary containing configuration parameters where keys correspond to attribute names and values are the new values to set.

####### Examples

>>> config = BeatsConfig()
>>> new_cfg = {
...     'input_patch_size': 32,
...     'dropout': 0.2,
... }
>>> config.update(new_cfg)
>>> print(config.input_patch_size)
32
>>> print(config.dropout)
0.2

NOTE

Ensure that the keys in the cfg dictionary match the attribute names of the BeatsConfig class to avoid any unexpected behavior.