---

name: model-architecture description: "Author, inspect, and troubleshoot DeepMedic modelConfig architecture definitions for normal, subsampled, and fully connected pathways." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

DeepMedic model architecture

Use this skill when a researcher must create or resize a DeepMedic model, or explain why a model configuration cannot build. Keep this skill focused on the architecture contract: model fields, pathway geometry, feature-map/channel flow, and architecture-side memory trade-offs. Use model-config.md for the field reference and troubleshooting.md for failure diagnosis.

Operating procedure

1. Start with a trusted Python-syntax model config. The native loader is ModelConfig(abs_path_to_cfg) and executes the file as Python; therefore do not treat an untrusted config as harmless data. For a no-TensorFlow static check, use the bundled scripts/inspect_model_config.py.
2. Normalize the required task dimensions: numberOfOutputClasses includes background, and numberOfInputChannels must equal the number of channels supplied by every input-channel manifest. Data-list validation belongs to data preparation, not this skill.
3. Define the normal pathway with one positive feature-map count and one 3-integer kernel per layer. Its stride-1 receptive field is 1 + sum(kernel[d] - 1) independently for each spatial dimension. Every configured train/validation/inference segment must be at least this large; account separately for any VALID FC kernels when checking actual output size.
4. Add a subsampled pathway only when needed. Normalize a single factor such as [3,3,3] to one pathway, or use [[3,3,3], [5,5,5]] for two. The source requires each active subsampled pathway to have exactly the same receptive field as the normal pathway. Prefer mirroring normal FMs and kernels first; then change FMs, not layer count, unless kernels are redesigned to preserve the field. Factors should be positive odd 3-vectors.
5. Treat numberFMsPerLayerFC as hidden FC layers only; the final classifier is appended automatically with numberOfOutputClasses feature maps. Supply exactly len(numberFMsPerLayerFC) + 1 entries in kernelDimPerLayerFC and padTypePerLayerFC. The FC input has the final feature maps from the normal path plus every active subsampled path.
6. Use one-based layer numbers in residual and lower-rank config lists. Layer 1 is invalid for residuals. Residual addition crops the earlier tensor at the center and pads or truncates channels; it is not a learned projection. Keep residual points on compatible, shrinking or same-sized geometry.
7. Re-run the inspector after each architecture edit. Before training or inference, ensure the checkpoint was made with the same classes, channels, pathway count/factors, layer widths, kernels, padding, activations, and FC layout. Training optimizer/session behavior is covered by training; inference is relevant here only for architecture/checkpoint compatibility.

Safe small-model recipe

For a minimal two-channel model, use numberOfInputChannels = 2, a small normal list such as [4, 5, 6], three [3,3,3] kernels, and segmentsDimTrain = [7,7,7] (normal receptive field 7). Set useSubsampledPathway = False, leave numberFMsPerLayerFC = [], and use the single default classifier kernel [[1,1,1]]. Choose a validation segment at least [7,7,7]; inference can be larger if memory permits. Set the class count to the actual number of labels plus background, not the number of foreground labels alone.

Memory-safe resizing

Reduce inference/training segment dimensions only down to the required receptive field; shrinking below it is invalid. To reduce model memory while preserving geometry, first lower FMs in later normal/subsampled layers, remove unused subsampled paths, or remove FC hidden layers. Keep all list lengths aligned, preserve equal normal/subsampled receptive fields, and check residual points after removing layers. Larger segments increase activation memory and can be more expensive than modestly reducing FMs. A lower-rank layer can reduce parameter cost, but this release's low-rank implementation has edge cases; see troubleshooting.md before applying it.