It looks like every parameter is counted twice. Is there a way to report correct results for Siamese and Triplet models?