AllMLPDecoder

All-MLP decoder based on SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers.

We provide the AllMLP Decoder, the head of SegFormer, as a freely usable head module. AllMLP Decoder takes intermediate features from previous backbone or neck module and outputs a segmentation map of the target size.

Compatibility matrix

Supporting backbones	Supporting necks	torch.fx	NetsPresso
ResNet MobileNetV3 MixNet CSPDarkNet MobileViT MixTransformer EfficientFormer	FPN YOLOPAFPN	Supported	Supported

Field list

Field	Description
name	(str) Name must be "all_mlp_decoder" to use AllMLPDecoder head.
params.intermediate_channels	(int) Intermediate feature dimension of the decoder.
params.classifier_dropout_prob	(float) Dropout probability of classifier.

Model configuration example

AllMLP decoder

model:
  architecture:
    head:
      name: all_mlp_decoder
      params:
        intermediate_channels: 256
        classifier_dropout_prob: 0.

Related links