EfficientAT mn10_as, ONNX
A modified version of the mn10_as checkpoint of EfficientAT
(mn10_as_mAP_471.pt from release v0.0.1, code at commit a425fdce92572e602a1d5634799bd9f1f2efa806),
© 2022 Florian Schmid, licensed under MIT. The licence is in LICENSE.
Changes from the original
- Exported to ONNX (opset 18) with the MobileNetV3 and a sigmoid, so the graph returns the 527 AudioSet class probabilities rather than logits.
- Feature extraction built in. The graph takes 32 kHz mono samples and holds the log-mel front
end of EfficientAT's
AugmentMelSTFTin eval mode: pre-emphasis, an 800-sample Hann window in a 1024-point frame, hop 320, 128 Kaldi mel banks from 0 to 15 kHz, log, (x + 4.5) / 5.
Weights are unchanged (float32). export_efficientat.py (with common.py) made this file and
checks it against EfficientAT's own pipeline: largest difference 2.5e-6.
File
efficientat-mn10-as.onnx, a single file of 20,200,590 bytes,
SHA-256 bec20620ab8be5e41297b45845f6fb116fbe42cd4da339ad7d44f27630b7a504.
| input | type | shape | meaning |
|---|---|---|---|
samples |
f32 | [W, N] |
W windows of mono 32 kHz audio, 0.5 s to 60 s long |
| output | type | shape | meaning |
|---|---|---|---|
probabilities |
f32 | [W, 527] |
AudioSet class probabilities per window |
The model was trained on 10 s clips; very short windows alone give poor answers.