Customizing parametrized tensors

[yuxuanzhang1995]

Hi, when constructing a TN with psi.gate_(), is it possible customize your tensor? (e.g., fix off diagonal terms or define certain functions in the tensor) Thanks!

[jcmgray]

Hi, I'm not totally sure I understand this question - with the gate function you can apply an arbitrary array already. Maybe you could give some more specific details?

[yuxuanzhang1995]

Yeah, let’s say if I want to implement something like: T = np.zeros([2,2]) T[0][0] = np.exp((J+h)) T[1][0] = np.exp(-1*(J)) T[0][1] = np.exp(-1*(J)) T[1][1] = np.exp((J-h)) where J and h are parameters; can I do that? Thanks!

…

On Mar 28, 2022, at 12:32 PM, Johnnie Gray ***@***.***> wrote: Hi, I'm not totally sure I understand this question - with the gate function you can apply an arbitrary array already. Maybe you could give some more specific details? — Reply to this email directly, view it on GitHub <#115 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANRXQL7PW3QZBJ4EH5BVHFDVCHUJPANCNFSM5RX36EVA>. You are receiving this because you authored the thread.

[jcmgray]

I think you're already there! just call psi.gate_(T, where) now?

[jcmgray]

Oh I think I didn't parse the parametrized in the title of this issue, is the aim to optimize the TN w.r.t to the parameters, and thus you need PTensor instances?

In which case have a look in circuit.py at how this is achieved for the preset gates. e.g.:

quimb/quimb/tensor/circuit.py

Lines 382 to 411 in 28dc9dd

	def fsim_param_gen(params):
	theta, phi = params[0], params[1]
	a_re = do('cos', theta)
	a_im = do('imag', a_re)
	a = do('complex', a_re, a_im)
	b_im = -do('sin', theta)
	b_re = do('imag', b_im)
	b = do('complex', b_re, b_im)
	c_im = -phi
	c_re = do('imag', c_im)
	c = do('exp', do('complex', c_re, c_im))
	data = [[[[1, 0], [0, 0]],
	[[0, a], [b, 0]]],
	[[[0, b], [a, 0]],
	[[0, 0], [0, c]]]]
	return ops.asarray(data)
	def apply_fsim(psi, theta, phi, i, j, parametrize=False, **gate_opts):
	mtags = _merge_tags('FSIM', gate_opts)
	if parametrize:
	G = ops.PArray(fsim_param_gen, (theta, phi))
	else:
	G = qu.fsim(theta, phi)
	psi.gate_(G, (int(i), int(j)), tags=mtags, **gate_opts)

that implements the function in a autoray compatible way (so many backends can be used), and creates a parametrized array which you could then optimize with respect to.

However, if you just want to construct the TN and contract it e.g., then the parametrized machinery is not necessary.

[yuxuanzhang1995]

Yeah that’s exactly what I needed. Thanks for the reference!

…

On Mar 28, 2022, at 3:57 PM, Johnnie Gray ***@***.***> wrote: Oh I think I didn't parse the parametrized in the title of this issue, is the aim to optimize the TN w.r.t to the parameters, and thus you need PTensor instances? In which case have a look in circuit.py at how this is achieved for the preset gates. e.g.: https://github.com/jcmgray/quimb/blob/28dc9dd222001b4336e33a147ad4bb442cb38455/quimb/tensor/circuit.py#L382-L411 <https://github.com/jcmgray/quimb/blob/28dc9dd222001b4336e33a147ad4bb442cb38455/quimb/tensor/circuit.py#L382-L411> that implements the function in a autoray compatible way (so many backends can be used), and creates a parametrized array which you could then optimize with respect to. However, if you just want to construct the TN and contract it e.g., then the parametrized machinery is not necessary. — Reply to this email directly, view it on GitHub <#115 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANRXQLZQUPYGIHVSQKU5TA3VCIMM7ANCNFSM5RX36EVA>. You are receiving this because you authored the thread.

[ghost]

Hi @jcmgray, I'm trying to create a custom parameterised gate for applying to a quantum circuit. I have tried to copy how 'fsim' is implemented in circuit.py:

...
def Givens_param_gen(theta): 

    a = do('cos', theta) 
    b = do('sin', theta) 

    data = [[1, 0, 0, 0], 
            [0, a, -b, 0], 
            [0, b, a, 0], 
            [0, 0, 0, 1]] 

    return ops.asarray(data) 

def apply_Givens(psi, theta, i, j, parametrize=False, **gate_opts): 
    mtags = _merge_tags('GIVENS', gate_opts) 
    if parametrize: 
        G = ops.PArray(Givens_param_gen, theta) 
    #else:
    #    qu.Givens(theta)
    psi.gate_(G, (int(i), int(j)), tags=mtags, **gate_opts)

from quimb.tensor.circuit import GATE_FUNCTIONS, ONE_QUBIT_PARAM_GATES, TWO_QUBIT_PARAM_GATES, ALL_PARAM_GATES
GATE_FUNCTIONS['GIVENS'] = apply_Givens
TWO_QUBIT_PARAM_GATES.add('GIVENS')
ALL_PARAM_GATES = ONE_QUBIT_PARAM_GATES | TWO_QUBIT_PARAM_GATES

def W_circ(n, **kwargs):
    
    circ = qtn.Circuit(n, **kwargs)

    circ.apply_gate('GIVENS', *qu.randn(1, dist='uniform'), 0, 1, parametrize=True)

    return circ
...

After plugging this into a loss-function/optimizer, the apply_gate is causing the error ValueError: The gate 'GIVENS' cannot be parametrized., I tried to solve this by explicitly adding 'GIVENS' to ALL_PARAM_GATES, but this didn't work. Am I missing something?

[jcmgray]

This line:

ALL_PARAM_GATES = ONE_QUBIT_PARAM_GATES | TWO_QUBIT_PARAM_GATES

redefines ALL_PARAM_GATES locally, not within circuit.py. It should work if you modify the circuit.py source (feel free to submit a PR with the new gate too), or you modify the variable inplace with ALL_PARAM_GATES.add('GIVENS').

Long term it might be worth adding adding an API so that we can add more custom gates easily.

[jcmgray]

As of 576edce, you should be able to just use:

def givens_param_gen(theta): 

    a = do('cos', theta) 
    b = do('sin', theta) 

    data = [[1, 0, 0, 0], 
            [0, a, -b, 0], 
            [0, b, a, 0], 
            [0, 0, 0, 1]] 

    return ops.asarray(data) 

register_param_gate('GIVENS', givens_param_gen, num_qubits=2)

I'd happily accept a PR adding this gate to quimb proper if you thought that would be useful.

[yuxuanzhang1995]

Oh, sorry; I guess the question is: does the autodiff part apply to customed functions like this?

…

On Mar 28, 2022, at 3:46 PM, Johnnie Gray ***@***.***> wrote: psi.gate_

[jcmgray]

Yes, any part of quimb that uses autoray.do like above supports autodiff.