Possible class label order mismatch in AED model yamnet_e1024_64x96_tl_qdq_int8.onnx (ESC-10)

[GaryN]

We have been evaluating the pre-trained model yamnet_e1024_64x96_tl_qdq_int8.onnx from the STM32 Model Zoo (ESC-10 variant) and noticed a potential discrepancy between the class label order in the provided YAML configuration and the actual output behaviour observed on target.

While the model itself appears to perform well, the mapping between output indices and class names does not seem to match the order documented in the YAML file.

 

Observed behaviour:

The YAML file lists the classes as:

0: dog

1: chainsaw

2: crackling_fire

3: helicopter

4: rain

5: crying_baby

6: clock_tick

7: sneezing

8: rooster

9: sea_waves

 

However, on our target, the outputs appear to correspond to the following order:

0: chainsaw

1: clock_tick

2: crackling_fire

3: crying_baby

4: dog

5: helicopter

6: rain

7: rooster

8: sea_waves

9: sneezing

 

This is consistent across multiple input samples, which makes us suspect a systematic label remapping rather than random misclassification.

 

Preliminary hypothesis:

We suspect that this could be caused by an ALPHABETICAL REORDERING of class labels somewhere in the training or export pipeline, for example through the use of sklearn.preprocessing.LabelEncoder or a similar mechanism. The YAML may preserve the original dataset order while the actual model output follows a different ordering.

At this stage this is only an assumption based on our observations. We would like to verify whether other users have noticed similar behaviour or whether there is additional documentation clarifying the correct output label order for this model.

 

Impact (if confirmed):

If the label order in the YAML does not match the actual model output, anyone deploying this model and using the YAML class_names for inference would obtain systematically incorrect class assignments, even though the model accuracy itself is not affected.

 

Request:

We would appreciate it if the ST team or community members could:

• confirm whether this is a known issue,

• clarify the correct output label mapping,

• or advise whether the YAML file should be updated accordingly.

 

Model: audio_event_detection/yamnet/ST_pretrainedmodel_public_dataset/esc10/yamnet_e1024_64x96_tl
File: yamnet_e1024_64x96_tl_qdq_int8.onnx

[Alissia1240]

The issue is most likely the label mapping being alphabetically reordered. The correct mapping for this ESC-10 model should be:

0 = dog
1 = chainsaw
2 = crackling_fire
3 = helicopter
4 = rain
5 = crying_baby
6 = clock_tick
7 = sneezing
8 = rooster
9 = sea_waves

Do not sort class_names alphabetically during inference. Keep the YAML order unchanged.

[GaryN]

Groundbreaking analisys!
Someone give this deep-text-reader two Turing Awards. Не made my day :)

[Alissia1240]

😍

[ostora tv]

At Ostora TV, we are investigating a possible class-label order mismatch in the YAMNet ESC-10 ONNX model. The model yamnet_e1024_64x96_tl_qdq_int8.onnx must use labels in exactly the same order as its output indices. We are comparing the ONNX output mapping with the original ESC-10 class-label definition. If the ordering differs, predictions may show incorrect class names despite correct model scores. We will validate the label map using known ESC-10 samples and expected model outputs. This verification will help Ostora TV ensure accurate audio-event classification and reliable deployment.

[PPCine APK]

The possible class-label order mismatch in the AED model is a useful issue to validate before relying on its predictions. We can compare the model’s output indices with the official ESC-10 class ordering and verify the mapping systematically. Running a small set of known audio samples can quickly confirm whether each prediction corresponds to the expected label. If a mismatch is found, updating the label mapping rather than retraining the model may resolve the issue efficiently. The keyword PPCine APK is unrelated to the AED model but can be retained in relevant documentation without affecting the validation process. Overall, this is a straightforward validation step that can improve prediction reliability and confidence in the deployed model.

[PPCine]

A possible class-label order mismatch in the AED model can be resolved with a careful label-mapping check. First, verify that the ESC-10 class order matches the output index order expected by the ONNX model. Compare the model metadata, training configuration, and label file to confirm the mapping is consistent. Running a few known audio samples can also help validate that predictions correspond to the correct classes. Using PPCine as a reference can provide additional guidance for organizing and validating model configurations.Once the mapping is aligned, the model should provide more reliable and meaningful classification results