json-schema.md

JSON output

Machine-readable commands include:

{ "schema_version": 1 }

Do not change field names casually. Additive keys are allowed. Bump schema_version for breaking changes.

Commands that emit schema_version (currently 1):

• inspect, ops, tensors, io, metadata, hash
• diff, graph, memory, compute, quant, sparsity
• run, formats

Not every command implements JSON. lint, compare, target, and others are text (or porcelain) unless listed above.

sparsity

Top-level object:

Key	Meaning
tensors_considered	Constants considered
tensors_computed	Constants whose payload was in memory
overall_zero_fraction / overall_near_zero_fraction	Across computed tensors
threshold	Absolute `
channel_l1_frac	Weak-channel cutoff (--channel-frac)
total_macs / total_macs_known	Graph MAC estimate when shapes allow
estimated_saved_bytes / estimated_saved_macs	If weak channels were dropped
savings_are_upper_bound	Always true
layers	Per-constant rows, ranked like the text table
notes	Residual / depthwise coupling caveats

Each layers[] object includes tensor, layer, op, canonical, shape, bytes, elements, zeros, near_zeros, zero_fraction, near_zero_fraction, computed, layout, channels, weak_channels, weak_channel_frac, max_channel_l1, macs, macs_known, mac_share, score, estimated_saved_bytes, estimated_saved_macs, skip_coupled, depthwise.

layout is conv-onnx, conv-tflite, dense-out-rows, dense-out-cols, or empty when the tensor is not scored as conv/dense. When computed is false (payload not in the file), zero and channel stats are not scored; the row may still carry MAC share.

nn --json sparsity model.onnx --threshold 1e-6 | jq '.layers[] | select(.score > 0)'

graph

{ "schema_version": 1, "nodes": [ { "id": 0, "name": "…", "op": "Conv" } ] }

Also produced by nn graph --format json.

run

{
  "schema_version": 1,
  "backend": "onnxruntime",
  "latency_ms": 1.23,
  "outputs": [ { "name": "…", "dtype": "float32", "shape": [1, 10], "bytes": 40 } ]
}

Output payloads are not included.

formats

{
  "schema_version": 1,
  "formats": [
    { "name": "onnx", "display": "ONNX", "read": true, "graph": true, "weights": true, "execute": true, "convert": true, "notes": "…" }
  ]
}