---

name: encoders-preprocessing description: "Choose segmentation_models_pytorch encoders/backbones, configure weights, and retrieve preprocessing parameters without relying on the source repository." disable-model-invocation: true metadata: disco-role: operating license: MIT

Encoders and Preprocessing

Use this sub-skill when a task asks how to choose an SMP encoder/backbone, list supported encoders, use tu- timm universal names, migrate deprecated timm- requests, choose a DPT-compatible encoder, or configure get_preprocessing_params / get_preprocessing_fn.

Start Here

• For native and ported encoders, use segmentation_models_pytorch.encoders.get_encoder_names() to list registry names and get_encoder(name, in_channels=3, depth=5, weights=None, output_stride=32, **kwargs) to instantiate the encoder directly.
• For model constructors, pass the same backbone as encoder_name=... and weights as encoder_weights=...; see ../model-building/SKILL.md for complete model recipes.
• For preprocessing, call get_preprocessing_params(encoder_name, pretrained="imagenet") when you need mean/std/input range as data, or get_preprocessing_fn(encoder_name, pretrained="imagenet") when you need a NumPy preprocessing callable.
• For timm universal encoders, prefer the tu- prefix. The pretrained variant belongs in the encoder name, and encoder_weights / weights should be True to request pretrained timm weights or None for random/offline initialization.
• For DPT, use tu- ViT-like encoder names supported by the DPT architecture and route full model construction details to ../model-building/SKILL.md.

Decision Flow

1. Choose a family from encoder-reference.md: native registry for stable ImageNet-style names, tu- timm universal for broad timm coverage, or DPT-compatible tu- ViT-style names for smp.DPT.
2. Decide whether weights may download. Use weights=None / encoder_weights=None for offline random initialization; use a native string such as "imagenet" only when downloads or cached configs are acceptable.
3. Align preprocessing with the weights choice. Preprocessing metadata is meaningful for pretrained weights; if the run is offline, collect params from cached/local config or hard-code a known policy deliberately.
4. Set depth and output_stride only when the model architecture expects them. Native encoders commonly produce depth + 1 feature maps including the input; timm universal supports depth up to 5 and may reject unsupported output strides.
5. Validate an encoder choice with the bundled script before writing larger model code:

python sub-skills/encoders-preprocessing/scripts/check_encoder.py resnet34 --preprocessing-only
python sub-skills/encoders-preprocessing/scripts/check_encoder.py tu-resnet18 --weights none --depth 5

References

• Encoder reference: native families, tu- naming, timm- migration, DPT caveats, depth, and output stride.
• Preprocessing: get_preprocessing_params, get_preprocessing_fn, offline behavior, and data-shape caveats.
• Troubleshooting: wrong encoder names, wrong weights, Hugging Face/timm downloads, DPT fixed-size inputs, and unavailable preprocessing params.
• check_encoder.py: inspect encoder/preprocessing metadata as JSON while avoiding downloads by default.

Boundaries

• Model constructor recipes, decoder-specific parameters, and architecture selection belong in ../model-building/SKILL.md.
• Losses, metrics, training loops, and evaluation belong in ../training-evaluation/SKILL.md.
• Export, tracing, deployment, and runtime packaging belong in ../model-export/SKILL.md.