Add `tfq.math.mps_1d_expectation` for 1D MPS #610

jaeyoo · 2021-08-03T22:30:16Z

Resolve #603 1st of the three ops.

MichaelBroughton

This is a great first pass @jaeyoo ! Would you be able to add some tests that verify expectations values computed using cirq's built in expectation calculation functions ? Also removing some of the print statements from the code :)

tensorflow_quantum/core/src/program_resolution.cc

Both Operation with natural born controlled gate (e.g. CNOT) and synthesized ControlledOperation are converted into the same tfq_gate_set with control_qubit. So, we can't filter ControlledOperation in this way, revert the previous commit.

MichaelBroughton · 2021-09-06T19:05:12Z

@jaeyoo any updates on this ? (My workstation has gone offline and won't be into the office for a while to turn it back on).

jaeyoo · 2021-09-06T19:15:23Z

@jaeyoo any updates on this ? (My workstation has gone offline and won't be into the office for a while to turn it back on).

I am still working on how to fix the delete issue...

jaeyoo · 2021-09-30T08:14:42Z

@MichaelBroughton I could fix the error finally. It was not easy because it had multiple reasons.

It turns out that we can't use tfq::QsimFor : I realized that the current MPSSimulator doesn't accept qsim::SequantialFor. It accepts qsim::For only. On top of it, for initialization, dummy value 1 is used for ForArgs, not tfq_for value.
Even if the theoretical minimum of qsim MPS bond_dim is 2 (as shown here), but the real test seems to require minimum 4 (as qsim MPSSimulator test code.), because the weird double free error occurred from it. After increasing the bond_dim from 2 to 4, the error's gone. So I want to make sure if there is a good test scenario to deal with which circuit accepts bond_dim=2 or not. could you please recommend it?
we can't use ApplyFusedGate() anymore. On top of it, ApplyGate() should be done with main_circuits not with fused_circuits.
Likewise, we can't use ComputeExpectation. We need to add similar function using ApplyGate(), not ApplyFusedGate()

Thank you so so much for your kind patience.

jaeyoo · 2021-09-30T08:33:06Z

Failed at //tensorflow_quantum/core/ops/noise:noisy_samples_op_test May i just re-run the test?

jaeyoo · 2021-09-30T22:59:18Z

PTAL

Note: qsim::ApplyFusedGate will be supported after qsim==0.10.3 due to quantumlib/qsim#395.

MichaelBroughton · 2021-10-01T18:24:34Z

We just cut a release here: https://github.com/quantumlib/qsim/releases/tag/v0.10.3-dev%2B20211001

Could we try upgrading to this new qsim version ?

jaeyoo · 2021-10-01T19:01:03Z

We just cut a release here: https://github.com/quantumlib/qsim/releases/tag/v0.10.3-dev%2B20211001

Could we try upgrading to this new qsim version ?

Sure, it's my pleasure. Done.

MichaelBroughton

Looking great! Just a few moderate sized changes then we should be good to merge!

MichaelBroughton · 2021-10-01T20:44:42Z

tensorflow_quantum/core/ops/math_ops/simulate_mps.py

+        bond_dim: `tf.Tensor` for an integer representing bond dimension
+            in this 1D MPS. This will create the following MPS:
+            [2, bond_dim], [bond_dim, 2, bond_dim] ... [bond_dim, 2]
+
+            The `bond_dim` should be >= 4.


Wouldn't this just be a constant and doesn't have to be a tf.Tensor ?

MichaelBroughton · 2021-10-01T20:52:52Z

tensorflow_quantum/core/ops/math_ops/tfq_simulate_1d_expectation.cc

+    if (max_num_qubits >= 26 || programs.size() == 1) {
+      ComputeLarge(num_qubits, fused_circuits, pauli_sums, context,
+                   &output_tensor);


Since MPS simulation doesn't have such a massive memory footprint as statevector I think we don't need ComputeLarge anymore.

MichaelBroughton · 2021-10-04T16:56:58Z

tensorflow_quantum/core/ops/math_ops/simulate_mps_test.py

+            util.convert_to_tensor([[x] for x in pauli_sums]))
+        self.assertDTypeEqual(res, np.float32)
+
+    def test_simulate_mps_1d_expectation_results(self):


This is a good start. Could we add a few more tests (with different names) here that do the following:

Test with all single qubit gates in tfq.get_supported_gates

Test with all single qubit gates in tfq.get_supported_gates.controlled_by(a_qubit) and ensure that this fails or passes accordingly)

Test that all two qubit gates in tfq.get_supported_gates works.

Ensure that the empty case works as expected

MichaelBroughton · 2022-01-04T09:26:21Z

I went ahead and cleaned up the implementation a little bit. Notable changes:

Removed the use of ApplyFusedGate. On closer inspection if two qubits that weren't adjacent to one another got fused together, things would break.
Added in the swap_endian option for parsing. The mps simulator in qsim has different endianness from the state vector simulator. Working in a proper fix to qsim would take a very long time so this is a workaround until we can get this fixed. Swap endianness of MPS simulator to match that of the state vector simulator quantumlib/qsim#492
Removed the ComputeLarge since MPS has a much smaller memory footprint
Added a few more tests.

jaeyoo

LGTM for me, but I can't approve my PR. :) @MichaelBroughton

jaeyoo added 7 commits July 29, 2021 01:00

initial commit

43b25b8

Fix simulate_mps_test

d56c35d

Add CheckQubitsIn1D

ae41578

Add Eigen & QSim MPS

b3f6d97

Add mps_1d op and its test

9f8963e

Mike's feedback - add bond_dim as attr

b85e932

Add Attr and its tests

3974752

jaeyoo requested a review from MichaelBroughton August 3, 2021 22:30

jaeyoo changed the title ~~Add tfq.math.1d_expectation for 1D MPS~~ Add tfq.math.mps_1d_expectation for 1D MPS Aug 3, 2021

jaeyoo added 4 commits August 3, 2021 22:36

Rename to mps_1d_expectation

b0a013d

Fix lint and format

3c21e16

Add import_test

1dbd1c0

Fix wheel test - add mps into release/BUILD

27a30d3

MichaelBroughton requested changes Aug 4, 2021

View reviewed changes

tensorflow_quantum/core/src/program_resolution.cc Outdated Show resolved Hide resolved

jaeyoo added 8 commits August 6, 2021 07:04

Mike's feedback

539b58d

WIP where Segmentation fault happens?

4f630c2

Merge branch 'master' into add_mps_1d

c5a51be

Add Mike's feedback - no gate with control_qubit is allowed

bb5fc0b

Revert control_qubit gate validation.

cd6f239

Both Operation with natural born controlled gate (e.g. CNOT) and synthesized ControlledOperation are converted into the same tfq_gate_set with control_qubit. So, we can't filter ControlledOperation in this way, revert the previous commit.

Fix program_resolution_test

fc86135

Fix CheckQubitsIn1D()

ea1d543

Add print messages

8035fc6

jaeyoo added 6 commits September 30, 2021 05:58

Merge branch 'master' into add_mps_1d

ce16003

Add debug print

617bc6b

Bump up to qsim==0.10.2

86b0149

Remove tfq::QsimFor and util_qsim.h::ComputeExpectationMPS

b0f2daf

Add ComputeExpectationMPSQsim in util_qsim.h

af5b413

Add more detailed debug prints

bf533d5

jaeyoo added 4 commits September 30, 2021 07:31

Bump up bond_dim from 2 (default) to 4 to fix segfault error.

9257c90

Fix how to use ApplyGate in ComputeExpectationMPSQsim

643e46c

Uncomment ComputeSmall() and Remove debug outputs

7057ca8

Fix format

2adca84

jaeyoo mentioned this pull request Sep 30, 2021

Support ApplyFusedGate in MPSSimulator quantumlib/qsim#437

Merged

jaeyoo added 3 commits September 30, 2021 22:40

Mike's feedback 1 : use fuses circuit and ApplyFusedGate()

fa51a17

Mike's feedback 2 : set default bond_dim to 4

01ff1c5

Fix format

97c11e9

jaeyoo requested a review from MichaelBroughton September 30, 2021 22:59

Mike's feedback 3 : bump up to qsim==0.10.3

4aed1b1

MichaelBroughton requested changes Oct 4, 2021

View reviewed changes

MichaelBroughton mentioned this pull request Dec 21, 2021

Supporting TFQ using IonqDevice quantumlib/Cirq#4769

Closed

MichaelBroughton added 2 commits January 3, 2022 20:15

Merge branch 'master' into add_mps_1d

cd216d7

bring 1d mps expectation into working state.

3875c85

empty test.

2b9f9de

jaeyoo commented Jan 4, 2022

View reviewed changes

MichaelBroughton merged commit dea26c6 into master Jan 4, 2022

MichaelBroughton deleted the add_mps_1d branch January 4, 2022 19:35

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add `tfq.math.mps_1d_expectation` for 1D MPS #610

Add `tfq.math.mps_1d_expectation` for 1D MPS #610

jaeyoo commented Aug 3, 2021

MichaelBroughton left a comment

MichaelBroughton commented Sep 6, 2021

jaeyoo commented Sep 6, 2021

jaeyoo commented Sep 30, 2021

jaeyoo commented Sep 30, 2021 •

edited

Loading

jaeyoo commented Sep 30, 2021

MichaelBroughton commented Oct 1, 2021

jaeyoo commented Oct 1, 2021

MichaelBroughton left a comment

MichaelBroughton Oct 1, 2021

MichaelBroughton Oct 1, 2021

MichaelBroughton Oct 4, 2021 •

edited

Loading

MichaelBroughton commented Jan 4, 2022

jaeyoo left a comment

Add tfq.math.mps_1d_expectation for 1D MPS #610

Add tfq.math.mps_1d_expectation for 1D MPS #610

Conversation

jaeyoo commented Aug 3, 2021

MichaelBroughton left a comment

Choose a reason for hiding this comment

MichaelBroughton commented Sep 6, 2021

jaeyoo commented Sep 6, 2021

jaeyoo commented Sep 30, 2021

jaeyoo commented Sep 30, 2021 • edited Loading

jaeyoo commented Sep 30, 2021

MichaelBroughton commented Oct 1, 2021

jaeyoo commented Oct 1, 2021

MichaelBroughton left a comment

Choose a reason for hiding this comment

MichaelBroughton Oct 1, 2021

Choose a reason for hiding this comment

MichaelBroughton Oct 1, 2021

Choose a reason for hiding this comment

MichaelBroughton Oct 4, 2021 • edited Loading

Choose a reason for hiding this comment

MichaelBroughton commented Jan 4, 2022

jaeyoo left a comment

Choose a reason for hiding this comment

Add `tfq.math.mps_1d_expectation` for 1D MPS #610

Add `tfq.math.mps_1d_expectation` for 1D MPS #610

jaeyoo commented Sep 30, 2021 •

edited

Loading

MichaelBroughton Oct 4, 2021 •

edited

Loading