Created by: sneaxiy

The relu op would output 0, which causes np.allclose fails randomly (two numbers which are all nearly 0s may fail in np.allclose). This PR fixes the double grad unittest by replacing relu with leaky_relu.