Hi,

def accuracy_multilabel(y_pred:Tensor, y_true:Tensor, sigmoid:bool=True):
    if sigmoid: y_pred = y_pred.sigmoid()
    outputs = np.argmax(y_pred, axis=1)
    real_vals = np.argmax(y_true, axis=1)
    return np.mean(outputs.numpy() == real_vals.numpy())

in this block.

This piece of code seems incorrect as the shape of y_pred and y_true is (Batch_size, class_space).
Doing a np.argmax with axis=1 returns a single class index value for each sample.
This is what we do for multi-class classification.

However in multi-class classification we don't normally use sigmoid on y_pred, although it is not wrong.

This function seems much like accuracy_multiclass rather than accuracy_multilabel