When I tried the 3dcnn pytorch version, can't reproduce the perasonr mentioned in the paper. The training performance on refine_minus_core set is 0.95(pearsonr), the test performance on core set is 0.59(pearsonr). Any suggestions?