espnet2.asr.state_spaces.base.TransposedModule

source

espnet2.asr.state_spaces.base.TransposedModule(module)

Transposed module.

This function serves as a decorator that wraps a SequenceModule class to allow it to accept a transposed parameter. When transposed is set to True, the input tensor’s dimensions are transposed before and after the forward pass, enabling compatibility with different input shapes.

espnet2.asr.state_spaces.base.transposed

Indicates whether the input should be transposed.

• Type: bool

• Parameters:module (type) – A subclass of SequenceModule that will be wrapped.

• Returns: A subclass of module with added transposition capabilities.

• Return type: type

Examples

>>> @TransposedModule
>>> class MyModule(SequenceModule):
>>>     def __init__(self, d_model):
>>>         super().__init__()
>>>         self.d_model = d_model
>>>         self.d_output = d_model
>>>
>>> my_module = MyModule(d_model=128, transposed=True)
>>> input_tensor = torch.randn(32, 128, 10)  # Shape (n_batch, d_model, l_sequence)
>>> output, state = my_module(input_tensor)

NOTE

This decorator modifies the behavior of the forward method to handle the transposed state appropriately.