You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
It's only really relevant for small encoders, e.g. with 2 layers, which is usually when it is used as a frontend, e.g. for Conformer. Although this also might happen during pretraining of a larger BLSTM encoder.
In that case, having one pooling of size 6 is usually worse than having two pools of size 2 and 3. Or maybe it depends what comes afterwards, but this is the case when this is used as a frontend for a Conformer. So in this use case, the current default is suboptimal.
I'm not really sure what's expected.
The text was updated successfully, but these errors were encountered:
It's only really relevant for small encoders, e.g. with 2 layers, which is usually when it is used as a frontend, e.g. for Conformer. Although this also might happen during pretraining of a larger BLSTM encoder.
In that case, having one pooling of size 6 is usually worse than having two pools of size 2 and 3. Or maybe it depends what comes afterwards, but this is the case when this is used as a frontend for a Conformer. So in this use case, the current default is suboptimal.
I'm not really sure what's expected.
The text was updated successfully, but these errors were encountered: