Packaging Cirq 0.7 for Nix: serialization failures

[drewrisinger]

Hi all,
#2727 is higher priority.

For NixOS/nixpkgs#78779 I tried building Cirq tag 0.7.0, and got several test failures. Are these in the main tree or local to me? And any tips on resolving them? (note: this is after disabling the failures from #2727, which also appear here).

Failure 2 looks like same issue as #2727.

________________________ test_serialize_sympy_constants ________________________

    def test_serialize_sympy_constants():
        proto = _arg_to_proto(sympy.pi, arg_function_language='')
        packed = json_format.MessageToDict(proto,
                                           including_default_value_fields=True,
                                           preserving_proto_field_name=True,
                                           use_integers_for_enums=True)
>       assert packed == {'arg_value': {'float_value': float(np.float32(sympy.pi))}}
E       AssertionError: assert {'arg_value':...592653589793}} == {'arg_value':...927410125732}}
E         Differing items:
E         {'arg_value': {'float_value': 3.141592653589793}} != {'arg_value': {'float_value': 3.1415927410125732}}
E         Use -v to get the full diff

cirq/google/arg_func_langs_test.py:108: AssertionError
___________________ test_serialize_conversion[value1-proto1] ___________________

value = array([ True, False])
proto = {'arg_value': {'bool_values': {'values': [True, False]}}}

    @pytest.mark.parametrize('value,proto', [
        ((True, False), {
            'arg_value': {
                'bool_values': {
                    'values': [True, False]
                }
            }
        }),
        (np.array([True, False], dtype=np.bool), {
            'arg_value': {
                'bool_values': {
                    'values': [True, False]
                }
            }
        }),
    ])
    def test_serialize_conversion(value: ARG_LIKE, proto: v2.program_pb2.Arg):
        msg = v2.program_pb2.Arg()
        json_format.ParseDict(proto, msg)
        packed = json_format.MessageToDict(_arg_to_proto(value,
>                                                        arg_function_language=''),
                                           including_default_value_fields=True,
                                           preserving_proto_field_name=True,
                                           use_integers_for_enums=True)

cirq/google/arg_func_langs_test.py:138: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
cirq/google/arg_func_langs.py:131: in _arg_to_proto
    msg.arg_value.bool_values.values.extend(value)
/nix/store/gajvrh1azp0vh6q531b695qrcak24zmq-python3.7-protobuf-3.7.1/lib/python3.7/site-packages/google/protobuf/internal/containers.py:279: in extend
    new_values = [self._type_checker.CheckValue(elem) for elem in elem_seq_iter]
/nix/store/gajvrh1azp0vh6q531b695qrcak24zmq-python3.7-protobuf-3.7.1/lib/python3.7/site-packages/google/protobuf/internal/containers.py:279: in <listcomp>
    new_values = [self._type_checker.CheckValue(elem) for elem in elem_seq_iter]
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <google.protobuf.internal.type_checkers.TypeCheckerWithDefault object at 0x7fffd76f45d0>
proposed_value = True

    def CheckValue(self, proposed_value):
      """Type check the provided value and return it.
    
      The returned value might have been normalized to another type.
      """
      if not isinstance(proposed_value, self._acceptable_types):
        message = ('%.1024r has type %s, but expected one of: %s' %
                   (proposed_value, type(proposed_value), self._acceptable_types))
>       raise TypeError(message)
E       TypeError: True has type <class 'numpy.bool_'>, but expected one of: ((<class 'bool'>, <class 'numbers.Integral'>),)

/nix/store/gajvrh1azp0vh6q531b695qrcak24zmq-python3.7-protobuf-3.7.1/lib/python3.7/site-packages/google/protobuf/internal/type_checkers.py:109: TypeError

Thanks in advance!

[mpharrigan]

The first issue seems to be a boring floating-point-comparison error. Replacing with np.isclose() in the test would work

If I had to guess on the second, it looks like something in the google protobuf libraries. What versions are you using and what is pinned in cirq's requirements?

[drewrisinger]

Yes, I realized that about the first error. Does it appear in your CI tests?

I'm using protobuf3.8. Nix's default is 3.7, which I also tested with. Error persisted in both configs.
Cirq pinned 3.8.

[vtomole]

@drewrisinger

pinned 3.8

Python 3.8? We don't run that on our CI yet.

[mpharrigan]

I think they're talking about protobuf 3.8

[dabacon]

The first one just looks like accuracy issue which should be fixed.

The second one, about bool, is a duplicate of #2727

[dabacon]

For the first we will need a version of np.is_close or np.are_almost_equal that works for python dicts.

[drewrisinger]

Failure 2 is resolved by #2986. Failure 1 is resolved by changing the following line:

Cirq/cirq/google/arg_func_langs_test.py

Line 108 in 91e919a

assert packed == {'arg_value': {'float_value': float(np.float32(sympy.pi))}}

to assert packed == {'arg_value': pytest.approx({'float_value': float(np.float32(sympy.pi))})}

I've also found a similar issue to Failure 1 in test_convert_to_ion_gates() when I test (cirq v0.7, v0.8) via Travis: https://travis-ci.org/github/drewrisinger/nur-packages/jobs/685881776#L6573-L6621.
I'm going to try to fix this one as well via pytest.approx. EDIT: I don't think this one can easily be fixed by pytest.approx b/c we would have to approximate the class cirq.PhasedXPowGate

[balopat]

Hi @drewrisinger is this still an issue for Cirq 0.8?

[drewrisinger]

Yes, it looks like it's been resolved. Likely due to Nixpkgs updating their protobuf version.

Relevant test config:

• python 3.8
• cirq==0.8.2
• protobuf==3.13.0
• numpy==1.19.1
• networkx 2.4