About precision and recall during training

[鬼鬼]

When I am training model = 1, these two values are always 1, is this correct?Which loss in the logs_edge.dat file should you pay more attention to when training model = 1?

[Mahmoud El Hussieny]

class EdgeAccuracy(nn.Module):
    """
    Measures the accuracy of the edge map
    """
    def __init__(self, threshold=0.5):
        super(EdgeAccuracy, self).__init__()
        self.threshold = threshold

    def __call__(self, inputs, outputs):
        labels = (inputs > self.threshold)
        outputs = (outputs > self.threshold)

        relevant = torch.sum(labels.float())
        selected = torch.sum(outputs.float())

        if relevant == 0 and selected == 0:
            return torch.tensor(1), torch.tensor(1)

        true_positive = ((outputs == labels) * labels).float()
        recall = torch.sum(true_positive) / (relevant + 1e-8)
        precision = torch.sum(true_positive) / (selected + 1e-8)

        return precision, recall

according to this the relevant and selective must have a value greater than zero to get result. I thing the edge you have generated had values below threshold. make sure edges to have only values of 0 and 1