espnet2.layers.mask_along_axis.MaskAlongAxis

source

class espnet2.layers.mask_along_axis.MaskAlongAxis(mask_width_range: int | Sequence[int] = (0, 30), num_mask: int = 2, dim: int | str = 'time', replace_with_zero: bool = True)

Bases: Module

Mask input tensor along a specified axis with random masking.

This class provides functionality to apply random masking to input tensors, allowing for more robust training of models by simulating missing data.

mask_width_range

The range of widths for the masks. Can be a single integer or a tuple defining the minimum and maximum width.

• Type: Union[int, Sequence[int]]

num_mask

The number of masks to apply to the input tensor.

• Type: int

dim

The dimension along which to apply the mask. Can be specified as an integer (1 for time, 2 for frequency) or as a string.

• Type: Union[int, str]

replace_with_zero

Whether to replace the masked values with zeros or with the mean of the tensor.

• Type: bool

• Parameters:

  • mask_width_range (Union *[*int , Sequence *[*int ] ]) – The range of widths for the masks.
  • num_mask (int) – The number of masks to apply.
  • dim (Union *[*int , str ]) – The dimension along which to mask (‘time’ or ‘freq’).
  • replace_with_zero (bool) – Flag to determine the value used to replace masked elements.

####### Examples

>>> mask_layer = MaskAlongAxis(mask_width_range=(0, 20), num_mask=3)
>>> masked_spec, lengths = mask_layer(spec_tensor, spec_lengths)

• Raises:
  • TypeError – If mask_width_range is not a tuple of two integers.
  • ValueError – If dim is not an integer or one of the specified strings.

NOTE

The masking is performed in-place for tensors that do not require gradients.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

extra_repr()

Returns a string representation of the MaskAlongAxis module’s attributes.

The string representation includes the mask width range, the number of masks, and the masking axis used in the masking process.

mask_width_range

The range of mask widths.

• Type: Union[int, Sequence[int]]

num_mask

The number of masks to apply.

• Type: int

mask_axis

The axis along which the masking is applied, either “time” or “freq”.

• Type: str

• Returns: A formatted string representation of the module’s parameters.

• Return type: str

####### Examples

>>> mask_layer = MaskAlongAxis(mask_width_range=(5, 10), num_mask=3)
>>> print(mask_layer.extra_repr())
mask_width_range=(5, 10), num_mask=3, axis=time

forward(spec: Tensor, spec_lengths: Tensor | None = None)

Forward function.

• Parameters:spec – (Batch, Length, Freq)