SoftTriple Loss: Deep Metric Learning Without Triplet Sampling