def test_uniform_unit_scaling_can_initialize(self):
tensor = torch.zeros([10, 6])
uniform_unit_scaling(tensor, "linear")
assert tensor.data.max() < math.sqrt(3 / 10)
assert tensor.data.min() > -math.sqrt(3 / 10)
# Check that it gets the scaling correct for relu (1.43).
uniform_unit_scaling(tensor, "relu")
assert tensor.data.max() < math.sqrt(3 / 10) * 1.43
assert tensor.data.min() > -math.sqrt(3 / 10) * 1.43
def test_uniform_unit_scaling_can_initialize(self):
tensor = Variable(torch.zeros([10, 6]))
uniform_unit_scaling(tensor, "linear")
assert tensor.data.max() < math.sqrt(3/10)
assert tensor.data.min() > -math.sqrt(3/10)
# Check that it gets the scaling correct for relu (1.43).
uniform_unit_scaling(tensor, "relu")
assert tensor.data.max() < math.sqrt(3/10) * 1.43
assert tensor.data.min() > -math.sqrt(3/10) * 1.43