inference-and-deployment

DLTK inference and deployment

Use this route when the task is to run prediction with an exported DLTK model, assemble patch predictions into a full volume, average stochastic crops, calculate segmentation metrics, or write a NIfTI result. This is an operating guide for DLTK 0.2.1's TensorFlow 1.x graph-execution deployment path.

Compatibility gate

• Expect Python 3.7 and TensorFlow 1.15-era graph execution. The documented deployment path is tensorflow.contrib.predictor.from_saved_model; do not claim that a modern TF2 installation provides this predictor or that these recipes are TF2-compatible.
• Confirm the model input and fetched output static shapes before running a volume. The sliding-window utility expects rank-compatible static shapes and a channels-last tensor. The input sample normally has [batch, spatial..., channels]; a full-volume application adds a dummy batch dimension.
• Keep model construction and output-key details in model-building, input/file/reader assumptions in data-pipelines, and export/training setup in training-and-estimators.

Choose the workflow

1. From the directory containing this file (inference-and-deployment/), run python scripts/sliding_window_plan.py --help, then provide spatial --input-shape, --window-shape, and --output-shape. The helper performs validation without importing TensorFlow, reading data, or contacting a network.
2. From that same sub-skill directory, run python scripts/sliding_window_smoke.py. It exercises the native test shape plus output-smaller, overlapping, and batched assembly using constant synthetic predictions; it does not load a model or dataset.
3. For a real export, read deployment-workflows.md and api-reference.md first. Discover a numeric export directory, load the TF1 predictor, inspect its tensors, assemble a volume, convert probabilities/logits to the required result, calculate metrics if labels exist, and write output only to an explicitly chosen directory.

Inference checklist

• Validate all spatial dimensions, the predictor's input/output keys, channel count, batch_size, and explicit stride before session execution. A malformed output/window combination can fail in padding or produce empty slices and a zero division counter.
• Treat sliding_window_segmentation_inference as an averaging assembler: it runs each patch, adds each output to the corresponding output slice, increments a separate coverage counter, and returns sum / counter for each requested op. It returns a list in ops_list order, not class labels.
• For segmentation, convert a [batch, spatial..., classes] probability or logit result with np.argmax(result, axis=-1). For probabilities, averaging overlapping patches before argmax is the DLTK MRBrainS pattern. For logits, a numerically stable softmax is useful when probabilities are needed, but argmax is unchanged.
• Exclude or report background explicitly when aggregating Dice. DLTK's dice returns NaN for a class with no predicted and no reference voxels; np.nanmean(dice[1:]) is an application policy, not a universal score.
• For SimpleITK output, create the array image with the correct spatial order, call CopyInformation from the source image, and only then write the requested .nii/.nii.gz. A shape or direction/spacing mismatch must be diagnosed rather than silently discarded.
• Random-crop regression/classification deployment is stochastic. Record the crop count, shape, seed policy (if any), and aggregation axis; the original IXI examples average four random crops and explicitly warn that results vary between runs.

See troubleshooting.md before changing geometry, predictor access, export discovery, or metric handling.