Evaluation metrics per class for multilabel sequence classification

[ehsan-soe]

Hi,

Does the repo support reporting evaluation metrics per class for multilabel sequence classification?

Thanks

[tbs17]

same ask here!

[rpoli40]

Here are the results that I got for toxic comments dataset using fast_bert: precision recall f1-score support

    toxic       0.50      0.93      0.65      6090
    severe_toxic       0.40      0.50      0.45       367
    obscene       0.61      0.84      0.71      3691
    threat       0.00      0.00      0.00       211
    insult       0.59      0.78      0.67      3427
    identity_hate       0.79      0.38      0.52       712

micro avg       0.55      0.82      0.66     14498
macro avg       0.48      0.57      0.50     14498
weighted avg       0.55      0.82      0.65     14498
samples avg       0.08      0.08      0.08     14498

Test F1 Accuracy: 0.6569613259668508 Test Flat Accuracy: 0.859779924348995

{'roc_auc': {'toxic': 0.747083913926278, 'severe_toxic': 0.7007573790439838, 'obscene': 0.79970607034752, 'threat': nan, 'insult': 0.7877764952425138, 'identity_hate': 0.8918915659392267}}

@kaushaltrivedi Can you please comment on it?