Closed oukohou closed 3 years ago
@oukohou Yes, at the start, we used the boilerplate code from CBAM repository to keep our code consistent. That's why those parameters were there when we trained the initial models. However, those unused parameters don't cause any errors, but you can use the updated Truplet Attention code from source. The old one within the MODELS
directory is present to load the weights correctly.
in
triplet_attention.py
:seems like
gate_channels
,reduction_ratio
,pool_types
are never used?