Skip to content

ADR-0364: saliency_student_v2 — Resize-decoder ablation on the v1 recipe

  • Status: Accepted (gate passed: v2 IoU 0.7105 ≥ v1 0.6558)
  • Date: 2026-05-09
  • Deciders: Lusoris, Claude (Anthropic)
  • Tags: ai, dnn, mobilesal, saliency, training, fork-local, docs

Context

ADR-0286 shipped saliency_student_v1 (2026-05-03) — a fork-trained tiny U-Net (~113 K params, ConvTranspose stride-2 upsample) trained on DUTS-TR as the production weights for the C-side mobilesal extractor. v1 deliberately avoided Resize because the ONNX op-allowlist (core/src/dnn/op_allowlist.c) did not yet admit it; v1's "Alternatives considered" table records the trade-off explicitly:

Use Resize for upsampling instead of ConvTranspose … Resize is not on core/src/dnn/op_allowlist.c at the time of this PR; adding it widens the PR scope into a new C-side audit + scanner change in the same diff as the training run. Rejected — ConvTranspose is on the allowlist already and produces a numerically equivalent stride-2 upsample for our purposes.

ADR-0258 (accepted 2026-05-03) closed that gap by widening the allowlist to admit Resize. The "Resize + Conv" pattern — bilinear F.interpolate followed by a 3×3 Conv2d — is the de-facto standard U-Net upsampling shape in the broader image-segmentation literature; v2 is the natural ablation that swaps in this pattern while holding everything else (encoder backbone, channel widths, skip connections, loss, optimizer, schedule, augmentation, seed) identical to v1, so the ablation is clean.

The training corpus, ship-gate (≥ v1's 0.6558 held-out IoU per ADR-0303-shape gate, no test-weakening), and I/O contract stay unchanged.

Decision

We will train and ship saliency_student_v2 — a fork-trained tiny U-Net (~124 K parameters, ONNX opset 17, BSD-3-Clause-Plus-Patent) as a new model/tiny/saliency_student_v2.onnx registry entry. The model is trained from scratch on DUTS-TR with ai/scripts/train_saliency_student_v2.py (forked from v1's trainer; only the decoder upsampling path changed) and exposes the same input / saliency_map tensor contract as v1 and mobilesal_placeholder_v0.

saliency_student_v1 stays as the production weights for the mobilesal extractor; v2 ships as a parallel artefact under model/tiny/. Promotion to production is a future PR after empirical validation in real ROI encodes — this PR establishes the pattern and the artefact, not the production flip.

The held-out IoU gate from ADR-0303-shape is non-negotiable per feedback_no_test_weakening: v2's 5 % DUTS-TR validation-fold IoU must match or beat v1's 0.6558. If it does not, the model ships with Status: Proposed (gate-failed: v2 IoU [%.4f] < v1 [0.6558]) and the production-flip PR is blocked.

The op-allowlist scope: ADR-0258 admits Resize op-type-only and explicitly does not enforce attributes. v2 exports with mode='linear', coordinate_transformation_mode='half_pixel', antialias=0 (the bilinear F.interpolate(align_corners=False) default produced by the legacy TorchScript exporter at opset 17).

Alternatives considered

Option Pros Cons Why not chosen
Resize + 3×3 Conv decoder (this ADR) Standard U-Net upsampling pattern from segmentation literature; clean ablation against v1 (one architectural change); exercises the new ADR-0258 allowlist affordance; learnable filter after a fixed-kernel resample is the strictly more expressive shape ~+11 K params vs v1 (3×3 vs 2×2 kernel); marginally slower forward (one extra resample-then-conv vs a fused transposed-conv) Chosen — establishes the pattern post-ADR-0258 and is the directly comparable ablation
Resize + 1×1 Conv decoder Strict parameter reduction vs v1 (1×1 has fewer weights than ConvT 2×2) 1×1 cannot mix neighbouring spatial info after the resample; reduces decoder capacity below v1; not a faithful "upgrade v1" ablation Rejected — the brief asked for the standard pattern, not a parameter-minimised variant
Bigger student + multi-dataset corpus Higher absolute IoU; closer to upstream u2netp territory Conflates two changes (architecture + corpus); v2 stops being a clean ablation; deferred per ADR-0286's saliency_student_v2 follow-up plan Rejected here — saved for saliency_student_v3
Stay on ConvTranspose (no v2) Zero engineering ADR-0258 is an unused affordance; the segmentation-literature pattern stays unverified on this fork's corpus Rejected — the affordance was opened specifically to unblock this kind of pattern
Replace v1 in the registry instead of shipping v2 in parallel Single canonical artefact Untested empirically in real ROI encodes; user direction is "v2 ships as a parallel artifact; promotion to production is a future PR after empirical validation" Rejected — production flip is a separate PR

Consequences

  • Positive:
  • Exercises the ADR-0258 Resize affordance with a real, in-tree consumer — proves the allowlist change end-to-end (ORT loads, parity check passes, op-scanner accepts).
  • Establishes a clean, single-variable ablation on the v1 recipe; future architectural choices (kernel size, multi-dataset corpus) can stack on this baseline.
  • Adds Resize to the in-tree fork-trained ONNX surface so any future tiny-AI model that wants the pattern has a precedent (docs/ai/models/saliency_student_v2.md).
  • Negative:
  • Adds a second saliency student to the model directory and the registry — operators must decide which to consume. Mitigation: saliency_student_v1 stays the documented production weights until v2 passes ROI-encode validation in a follow-up PR.
  • DUTS-TR is still 271 MB and not redistributed in-tree; reproducing v2 has the same dataset-fetch step as v1.
  • Neutral / follow-ups:
  • Production flip: a follow-up PR will A/B v1 vs v2 in real ROI-encode pipelines and either promote v2 or rule it not-affecting-quality. Filed as a backlog row.
  • saliency_student_v3: the multi-dataset (DUTS-TR ∪ MSRA10K ∪ HKU-IS) and ~300 K-param backlog item from ADR-0286 stays open and is independent of v2.
  • External evaluation: the DUTS-TE / ECSSD held-out evaluation follow-up from ADR-0286 should run against both v1 and v2 once the harness lands.

References

  • ADR-0258 — admits Resize op-type to the allowlist; this ADR is the first fork-local consumer.
  • ADR-0286 — v1 decision record; this ADR holds v2's recipe identical to v1's except for the upsampling path.
  • ADR-0303 — the no-test-weakening production-flip gate shape this ADR follows.
  • ADR-0042 — tiny-AI doc-substance rule this PR satisfies (ships docs/ai/models/saliency_student_v2.md).
  • ADR-0108 — fork-local PR deep-dive deliverables checklist.
  • DUTS-TR dataset: Wang, Lu, Wang, Feng, Wang, Yin, Ruan, "Learning to Detect Salient Objects with Image-Level Supervision", CVPR 2017. Project page: http://saliencydetection.net/duts/. Direct URL: https://saliencydetection.net/duts/download/DUTS-TR.zip. Distribution: free for academic and research purposes (per the project page). Same provenance as v1.
  • Source: paraphrased — task brief directive "train saliency_student_v2.onnx using the Resize op now allowed in the op-allowlist (per ADR-0258), v2 uses Resize (linear, half_pixel) for the upsampling path."