padding_mode defaults to "reflect" causes Exception when input's time_length is smaller than half of kernel size.

[chenjiasheng]

Describe the bug

A "reflect" padding mode tries to copy half of kernel size numbers of frames from the input,
It fails when input's time_length is smaller than half of kernel size.
It happens when I create a speechbrain.lobes.models.transformer.Transformer.TransformerEncoder with ffn_type=1dcnn with ffn_cnn_kernel_size_list=[5,5]. If the input has only one time frame, the error happens. By the way, the input for me is phoneme sequence, it is definitely a legal input to having only one phoneme.

Sorry that I can't provide a detailed stack traceback be cause I can't access company network now, and I hope the discription is clear enough.

Maybe it's safer to make Conv1D having padding_mode='zeros'?
Or, perhaps in this special case, fallback to replicate mode, and leaving legacy codes/models unaffected?

Expected behaviour

CNN1D works out-of-box for the special case when input's time_length is smaller than half of kernel size.

To Reproduce

No response

Environment Details

No response

Relevant Log Output

No response

Additional Context

No response

[chenjiasheng]

@Adel-Moumen @BenoitWang

[asumagic]

AFAIK this is indeed an expected consequence of using reflect padding. I don't know if there's a "correct" solution other than changing the padding type for that specific usecase indeed.