[runwangdl]

What

Three of the five MLperf Tiny benchmark exporters did not build the reference model. Checked against the int8 reference graphs already shipped in TrainDeeploy (Tests/Models/MLPerf/):

Benchmark	Reference	Before	After
KWS	DS-CNN, 1x49x13→12, 22,160 par	1x25x10, 2,732 par	1x49x10, 23,180 par, 2.657M MACs
AD	Deep AE, 640→640, 10 Gemm, 265,106 par	128→128, 29,120 par	640→640, 267,928 par, 0.264M MACs
VWW	MobileNetV1-0.25, 3x96x96→2, 209,126 par	no exporter at all	213,586 par, 7.490M MACs
IC	ResNet-8, 77,712 par	78,042 par	unchanged ✅
—	CCT-2/3x2	283,979 par	unchanged ✅

Details

DS-CNN. The KWS export defaulted to the xs PULP variant. The registry already had a DSCNN-S entry, but its stem used valid padding and emitted a 20x4 map against the reference's 25x5 — the reference stem is Conv2D(64,(10,4),strides=2,padding='same'). Adds stem_padding="same", computed as the smallest symmetric padding reproducing Keras' output size so no separate Pad node appears. MACs go 1.701M → 2.657M, the published figure. The legacy xs default is untouched.

Autoencoder. The AD export was a 128-dim toy; the pre-existing mlperf variant (128→[128,128,128]→128) was not the reference either. Adds MLPerfADAutoencoder behind variant="mlperf_ad" and an Autoencoder-AD registry entry.

BatchNorm1d over an (N,C) tensor is undefined at N=1 — zero variance, and PyTorch refuses outright — but on-device training runs at batch 1. FrozenStatsBatchNorm1d keeps the layers and their learnable affine parameters and normalizes with randomized running statistics instead, so the training graph carries the reference's 9 BatchNormInternal + 9 BatchNormalizationGrad nodes and its parameter count.

MobileNetV1. The MLperf Tiny VWW reference is MobileNetV1-0.25, not the MobileNetV2-0.35 the registry described as such. There was no V1 exporter, so the mobilenetv1_train fixture deployed in TrainDeeploy could not be regenerated from this repo — even though README.md:185 and scripts/make_single_step.py:42 both use --model MobileNetV1 as an example. Module names match that fixture's parameters (stem_*, blocks_N_{dw,bn_dw,pw,bn_pw}_*, fc_*) so regenerated graphs stay drop-in.

Registry descriptions. DSCNN-S claimed ~270K params (it is ~23K); MobileNetV2-VWW claimed to be the MLperf VWW reference.

Testing

• pytest tests/ → 154 passed, 4 skipped
• Regenerating ResNet-8 reproduces the committed TrainDeeploy fixture node for node (114 nodes, identical op histogram, 78,042 params) — i.e. the toolchain is unchanged for models that were already correct.
• All numbers above measured from the generated ONNX with onnxruntime-training==1.19.2 under the pinned versions.
• black / isort clean; no new flake8 findings.

Follow-up (not in this PR)

The regenerated DS-CNN and Autoencoder fixtures still have to land in TrainDeeploy, where the DS-CNN graph is ~8x and the AE ~9x their previous size — test_gap9_tiled_config.py L1/L2 budgets will need re-tuning.

🤖 Generated with Claude Code

https://claude.ai/code/session_012vb2GVs3qbiyLV1KkV2baN