Why make torch.tanh(pre_out) * np.pi / 2.0?

PennyLane Help
1

Hello,
Here example: Quantum transfer learning

Why make this operation? The tutorial say " A constant np.pi/2.0 scaling", what is special about this?

Thanks you

2

Hey @wing_chen,

There’s a factor of \pi / 2 being multiplied by the output of torch.tanh (range of -1 to 1) in the forward function. This variable, q_in, then goes into quantum_net, where it then gets used as the angle of rotation for a qml.RY gate. So, the angles that qml.RY will receive are [-\pi/2, \pi/2], which covers the range of Pauli Y rotations. It’s a similar thing to why when you integrate over something in spherical coordinates, the domain of the azimuthal angle that you integrate over is [0, \pi] (equivalent to [-\pi / 2, \pi / 2]).

Hope that helps!

3

Thanks you. I have anothr question.
If making quantum convolution, and output is qml.sample or qml.exp_vals, should I still use np.pi/2?

4

Hey @wing_chen,

I’m not sure I understand the question exactly. But, if your quantum neural network has RY gates and you want to ensure that they receive angles that cover the full range, then the same interval can be used :slight_smile: