`smoke_v0_symbolic_batch` — CI symbolic batch dimension probe¶

Status — CI internal. This is not a quality model and must not be used for video quality assessment. It exists solely as a shape-handling probe for the libvmaf DNN integration CI gate.

smoke_v0_symbolic_batch is a minimal ONNX graph (opset 17) whose input shape declares a symbolic batch dimension (dim_param="batch", surfaced by ORT as -1 or 0). It was introduced under ADR-0524 to verify that dnn_attach_api.c and dnn_api.c accept symbolic batch inputs and fold them to batch size 1 rather than rejecting them with -ENOTSUP.

Checkpoint facts¶

Field	Value
Model id	`smoke_v0_symbolic_batch`
Location	`model/tiny/smoke_v0_symbolic_batch.onnx`
Architecture	Identity mapping with dynamic batch axis
Trainable parameters	0
Training	None: an Identity graph with no weights
Input	`frame` — float32 `["batch", 1, 4, 4]`
Output	`score` — float32 `["batch", 1, 4, 4]`
ONNX opset	17
License	BSD-2-Clause-Patent
Registry entry	`smoke_v0_symbolic_batch` in `model/tiny/registry.json` (`"smoke": true`)
SHA-256	`cc1a75a7518b27cdfc9df481bb1282950cc51fe7ef6f62c5dbccb106b73cc509`

Purpose¶

All shipped No-Reference tiny checkpoints (such as nr_metric_v1.onnx) declare their input tensor with dim_param="batch". Historically, the libvmaf shape gate rejected in_shape[0] != 1. smoke_v0_symbolic_batch provides a lightweight, fast CI probe verifying:

in_shape[0] ∈ {1, -1, 0} is accepted at attach time.
Symbolic H/W or fixed batch > 1 remains rejected with clear diagnostics.
Exercised in core/test/dnn/test_vmaf_use_tiny_model.c via test_attach_accepts_symbolic_batch_rank4.

Output interpretation¶

Outputs are synthetic test signals. Values reflect fixture identities, not perceptual quality. PLCC / SROCC / RMSE are not applicable.

Runnable usage example¶

# Verify via the C unit test suite:
python3 "$(git rev-parse --show-toplevel)/scripts/ci/run_meson_test.py" -- \
  -C build --suite=dnn test_vmaf_use_tiny_model

# Or inspect the symbolic dimension via Python:
python3 -c 'import onnxruntime as ort; sess = ort.InferenceSession("model/tiny/smoke_v0_symbolic_batch.onnx"); print("Input shape:", sess.get_inputs()[0].shape)'

Known limitations¶

CI internal only: do not use for perceptual quality assessment.
Batch folding: symbolic batch is folded to 1; parallel multi-batch scheduling is not implemented.
CPU only: designed for CI execution.

smoke_v0_symbolic_batch — CI symbolic batch dimension probe¶