Some encoders (e.g. satlasnet_mi) already merge information on the temporal dimension. Then, no multi_temporal_strategy is needed; if one is included in the configs, it will be skipped.
When using the LTAE, the decoder input channels are changed to match LTAE output channels (LTAE only works with all inputs having the same number of channels)
The bug was that even if the LTAE was skipped, the decoder's input channels were modified, causing shape errors.
Fix: change decoder input channels if LTAE is used AND the encoder has a multi-temporal output.
(for both SegMTUPerNet and RegMTUPerNet)
Error:
Fix: change decoder input channels if LTAE is used AND the encoder has a multi-temporal output. (for both SegMTUPerNet and RegMTUPerNet)