# Specifiy the evaluation function
def compute_metrics(eval_pred):
    logits, labels = eval_pred
    probs = torch.sigmoid(torch.tensor(logits)).numpy()
    preds = (probs >= 0.5).astype(int)
    labels = labels.astype(int)

    from sklearn.metrics import accuracy_score

    exact_accuracy  = accuracy_score(labels, preds)
    macro_f1        = f1_score(labels, preds, average="macro", zero_division=0)
    micro_f1        = f1_score(labels, preds, average="micro", zero_division=0)
    macro_precision = precision_score(labels, preds, average="macro", zero_division=0)
    macro_recall    = recall_score(labels, preds, average="macro", zero_division=0)

    per_class_f1        = f1_score(labels, preds, average=None, zero_division=0)
    per_class_recall    = recall_score(labels, preds, average=None, zero_division=0)
    per_class_precision = precision_score(labels, preds, average=None, zero_division=0)
    per_class_accuracy  = (preds == labels).mean(axis=0)

    per_class_metrics = {}
    for i, emotion in enumerate(EMOTION_LABELS):
        per_class_metrics[f"f1_{emotion}"]        = float(per_class_f1[i])
        per_class_metrics[f"recall_{emotion}"]    = float(per_class_recall[i])
        per_class_metrics[f"precision_{emotion}"] = float(per_class_precision[i])
        per_class_metrics[f"accuracy_{emotion}"]  = float(per_class_accuracy[i])

    return {
        "exact_accuracy":   exact_accuracy,
        "macro_f1":         macro_f1,
        "micro_f1":         micro_f1,
        "macro_precision":  macro_precision,
        "macro_recall":     macro_recall,
        **per_class_metrics,
    }