Confusion Matrix on BYOM

Question/Issue:
Confusion Matrix on detailed like on Edge Impulse Paltfrom when run code from download model features (BYOM)

Project ID:

Context/Use case:
i want to get details of raw confusion matrix at the end of training models

Steps Taken:

  1. train the model

Expected Outcome:
at the end of training models
calculate_fomo_metrics {“support”: 192, “confusion_matrix”: [[27046, 12, 10, 3, 2], [17, 183, 0, 0, 0], [14, 0, 169, 0, 0], [1, 0, 0, 101, 0], [5, 0, 0, 0, 85]], “non_background”: {“precision”: 0.952212389380531, “recall”: 0.9356521739130435, “f1”: 0.943859649122807}, “classification_report”: {“0”: {“precision”: 0.998633829339438, “recall”: 0.9990026964134008, “f1-score”: 0.9988182288204446, “support”: 27073}, “1”: {“precision”: 0.9384615384615385, “recall”: 0.915, “f1-score”: 0.9265822784810127, “support”: 200}, “2”: {“precision”: 0.9441340782122905, “recall”: 0.9234972677595629, “f1-score”: 0.9337016574585636, “support”: 183}, “3”: {“precision”: 0.9711538461538461, “recall”: 0.9901960784313726, “f1-score”: 0.9805825242718447, “support”: 102}, “4”: {“precision”: 0.9770114942528736, “recall”: 0.9444444444444444, “f1-score”: 0.96045197740113, “support”: 90}, “accuracy”: 0.9976851851851852, “macro avg”: {“precision”: 0.9658789572839973, “recall”: 0.9544280974097561, “f1-score”: 0.9600273332865991, “support”: 27648}, “weighted avg”: {“precision”: 0.9976660602033497, “recall”: 0.9976851851851852, “f1-score”: 0.9976725210976868, “support”: 27648}}, “class_names”: [“_background”, “mata_terbuka”, “mata_terpejam”, “menguap”, “tidak_menguap”]}

Actual Outcome:
Just this matrix each epoch
Epoch Train Validation
Loss Loss Precision Recall F1
59 0.03492 0.04800 0.95 0.94 0.94

Reproducibility:

  • [ x] Always
  • [ ] Sometimes
  • [ ] Rarely
1 Like

Hi @nugro

This sounds like a feature request, could you provide some additional examples of the output you’re expecting? For instance, you mentioned wanting raw confusion matrices at the end of the training, but how would you like to visualize or interact with this data? Would you expect detailed metrics per class (e.g., precision, recall, F1-score, support), similar to what’s currently provided during Edge Impulse’s live classification?

Is this for an ongoing enterprise contract? If you do not already have a solutions engineer already let me know otherwise it could be a good time to Contact Sales

Best

Eoin

Hi @Eoin

Actually this project is for my thesis research.

Yep, the details is like you provide in logs if train using edge impulse platfrom.

So, why there are no same matrix between if we training in locally using download features and using edge impulse platfrom?

This is just displayed on the UI for configuration of the model params before training. Locally you cannot configure them.

Hope this helps

Eoin