You can deploy this model and start earning money today!

legojoey17/
batai-mobilenet-onnx

Name: batai-mobilenet-onnx
Availability: InStock

Views

No views yet

Batai MobileNet ONNX

This repository pins the Batai MobileNet model used by OpenBat. The ONNX file is copied unchanged from the Apache-2.0 Batai repository. This is an unofficial mirror, not a Kitware release.

Artifact

File	Size	SHA-256
`model.mobilenet.onnx`	12,135,629 bytes	`e6e36494aa6457f8d93bb882e90fe7bf2100484fdaa2085d15856217ba1fcbf1`

The model comes from assets/model.mobilenet.onnx at Batai revision 4fb905fbec73d3e5f90bb3ada5a62af283612015.

Fetch the upstream model

Upstream already provides this model as ONNX, so there is no conversion step. Fetch the pinned Git LFS object and verify its size and SHA-256 with:

./scripts/fetch-upstream.sh

The script writes reproduced/model.mobilenet.onnx. That directory is ignored. A successful run means the fetched file matches the committed model byte for byte.

Model contract

Input: input, uint8, NHWC [batch, 224, 224, 3].
Channel order: BGR, matching the upstream training and ONNX-validation path.
Output: output, float32, [batch, 35] class probabilities.
Region: North America.
The 35 class labels, including NOISE, are embedded in the ONNX metadata.

The model does not accept raw audio. Callers must reproduce the Batai spectrogram, activity-cropping, rasterization, resize, and channel-order contract.

Preprocessing

preprocessing/README.md links directly to the closest available upstream implementation and includes unchanged copies of the later spectrogram and training image paths. The model revision itself does not contain preprocessing code, so the repository does not claim that these later files produced this exact model. scripts/fetch-upstream.sh retrieves and verifies both the model and the pinned reference source.

Golden fixture

golden/ contains a deterministic model-ready input tensor and its expected output. It catches changes in model bytes, tensor names, shapes, dtypes, or ONNX Runtime results. It does not test the raw-audio preprocessing path.

uv run --project golden --frozen --python 3.11.15 python scripts/verify_golden.py

The generation seed, runtime versions, hashes, and tolerances are recorded in golden/manifest.json.

Licence and attribution

This mirror is distributed under Apache License 2.0, matching the upstream repository. The upstream LICENSE and NOTICE files are copied here unchanged.

Limitations

Kitware has not authorized or endorsed this mirror.
The available upstream material does not fully document every regional or training-data limit.
RGB input is not equivalent to the BGR training contract.
A matching checksum proves byte identity. It does not establish biological accuracy or deployment fitness.

legojoey17/batai-mobilenet-onnx