Double Checking implementation detail in SegFormerHead

[Rahul Singhal]

The paper on SegFormer suggests an All MLP decoder.

The SegformerHead.py shows the use of a Conv2D for the final layer.

Can you help me understand if this is a deviation from the paper or mentioned in a followup paper somewhere? I apologize in advance if there is an obvious answer.

[sithu3]

1x1 conv is basically the same as linear operation. In segmentation, 1x1 conv is usually used as the last layer. You can also check the official segformer implementation.

[Rahul Singhal]

Thank you!