Make tape mode default #1040

josh146 · 2021-01-26T12:05:49Z

Context:

Makes tape mode default when importing PL

Description of the Change:

qml.enable_tape is placed at the bottom of the __init__.py file
Test fixtures are updated to take this into account
Various tests are updated to take this into account
The device method is removed from default.tensor.tf
Various bug fixes

Benefits:

Possible Drawbacks: The old core remains present!

Related GitHub Issues:

josh146 · 2021-01-26T12:06:20Z

pennylane/__init__.py

@@ -230,3 +230,6 @@ def circuit():
 def version():
    """Returns the PennyLane version number."""
    return __version__
+
+
+enable_tape()


This is placed at the bottom of the __init__ file, to allow the import logic to remain in non-tape mode.

Is that necessary so that non-tape mode can still be accessible through qml.disable_tape()?

No, this is because a lot of the logic that is executed above this isn't tape-mode compatible yet 😆

I initially added qml.enable_tape to the top of the __init__.py file, but it requires a lot of changes throughout the code base to get it to work. I figured it would be too hard to code-review, and might introduce bugs.

josh146 · 2021-01-26T12:06:31Z

pennylane/_device.py

+                if isinstance(o, qml.tape.MeasurementProcess) and o.obs is not None:
+                    observable_name = o.obs.name
+                else:
+                    observable_name = o.name


a bug uncovered

github-actions · 2021-01-26T12:06:41Z

Hello. You may have forgotten to update the changelog!
Please edit .github/CHANGELOG.md with:

A one-to-two sentence description of the change. You may include a small working example for new features.
A link back to this PR.
Your name (or GitHub username) in the contributors section.

josh146 · 2021-01-26T12:06:40Z

pennylane/circuit_drawer/representation_resolver.py

+        if isinstance(op, qml.tape.MeasurementProcess) and op.obs is not None:
+            op = op.obs
+


a bug uncovered

josh146 · 2021-01-26T12:07:00Z

pennylane/circuit_graph.py

@@ -253,7 +253,7 @@ def hash(self):
        """
        return hash(self.serialize())

-    def to_openqasm(self, rotations=True):
+    def to_openqasm(self, wires=None, rotations=True):


This method now has the same issue re: device wires as the circuit drawer :)

josh146 · 2021-01-26T12:08:32Z

pennylane/tape/__init__.py

@@ -40,6 +41,128 @@
 _mock_stack = []


+class TapeOperationRecorder(QuantumTape):


The OperationRecorder and @template decorator were never ported to work with the new queuing refactor.

As a result, in tape mode, they no longer work. They do not error, but instead just return empty lists, as they are not accessing the new AnnotatedQueue and QueuingContext. The versions below are written to work in tape mode, and dynamically patch the existing versions when enable_tape is called.

Wondering if these classes could be lost in the __init__.py file 🤔 Could it be worth having them in a separate file?

josh146 · 2021-01-26T12:08:49Z

pennylane/tape/measure.py

    def __repr__(self):
        """Representation of this class."""
        if self.obs is None:
-            return "{}(wires={})".format(self.return_type.value, self.wires)
+            return "{}(wires={})".format(self.return_type.value, self.wires.tolist())


Discovered a bug :)

pennylane/tape/qnode.py

pennylane/templates/layer.py

josh146 · 2021-01-26T12:12:52Z

pennylane/utils.py

+    if qml.tape_mode_active():
+        for op in operation_list:
+            try:
+                qml.tape.QueuingContext.remove(op)
+            except KeyError:
+                pass
+
+        with qml.tape.QuantumTape() as tape:
+            for o in operation_list:
+                o.queue()
+                if o.inverse:
+                    o.inv()
+
+        tape.inv()
+        return tape


This is too allow qml.inv() to work in tape mode. It was not previously re-written to support the new queuing system.

Here, we simply convert the list of operations into a sub/nested tape, remove the operations from the existing/parent tape, and invert the nested tape.

isn't qml.inv() more of a transform (take a circuit and return the inverted circuit), rather than something we want to do "in-place" 🤔

Yep, agree, however this existing function was already an in-place function.

I heavily considered just deleting this function (not sure if anyone ever uses it, I had completely forgotten it existed 😆 ) but realised I could make it work with existing logic with these few added lines.

But replacing this with a transform is definitely on the to-do list

So this is now the only operation except from decompositions that introduces nesting, right? So if I use inv together with decompositions, the resulting tape can have nesting depth 3? And that is accounted for everywhere by the tape.expand function?

josh146 · 2021-01-26T12:13:13Z

tests/circuit_graph/test_circuit_graph.py

@@ -174,16 +174,16 @@ def test_layers(self, parameterized_circuit, wires):
        assert layers[1].ops == [ops[3]]
        assert layers[1].param_inds == [3]
        assert layers[2].ops == [ops[x] for x in [5, 6]]
-        assert layers[2].param_inds == [4, 5]
+        assert layers[2].param_inds == [6, 7]


This is because tape mode numbers parameters differently to non-tape mode.

josh146 · 2021-01-26T12:14:04Z

tests/devices/test_default_gaussian.py

+            if "State" not in g:
+                qml.Displacement(a, 0, wires=[0])


In tape mode, state preparations are now not allowed to be placed after unitary operations.

tests/interfaces/test_autograd.py

josh146 · 2021-01-26T12:15:19Z

tests/test_numpy_wrapper.py

-            # Test that the gate parameter is not a PennyLane tensor, but a
-            # float
-            assert not isinstance(rec.queue[0].parameters[0], np.tensor)
-            assert isinstance(rec.queue[0].parameters[0], float)


This was a feature specifically of BaseQNode, and is now no longer the case.

josh146 · 2021-01-26T12:15:46Z

tests/test_operation.py

-    def test_operation_inside_context_do_queue_false(self, qnode):
-        """Test that an operation does not get added to the QNode queue when do_queue=False"""


do_queue has been removed from tape mode

josh146 · 2021-01-26T12:16:12Z

tests/test_optimize.py

@@ -47,7 +47,6 @@
 # function arguments in various formats
 mixed_list = [(0.2, 0.3), np.array([0.4, 0.2, 0.4]), 0.1]
 mixed_tuple = (np.array([0.2, 0.3]), [0.4, 0.2, 0.4], 0.1)
-nested_list = [[[0.2], 0.3], [0.1, [0.4]], -0.1]


ragged arrays as input is no longer allowed in tape mode

pennylane/tape/__init__.py

mariaschuld · 2021-01-26T12:21:09Z

pennylane/tape/measure.py

@@ -72,16 +72,24 @@ def __init__(self, return_type, obs=None, wires=None, eigvals=None):
        # Below, we imitate an identity observable, so that the
        # device undertakes no action upon recieving this observable.
        self.name = "Identity"
-        self.diagonalizing_gates = lambda: []


A bit surprised this changes?

This was just so that the TestCircuitGraphDiagonalizingGates test passes 🤔

Alternatively, I could just delete that test class?

I didn't know the circuit graph has functionality for diagonalisation in the first place...

…into make-tape-default

mariaschuld

I think another approval would be necessary, but all changes make sense to me, and you addressed all concerns/questions I had.

The only thing one might want to consider is to delete non-tape tests. They have basically no meaning any more if we know we reproduce them in the new tape tests. And the delete will obviously not introduce bugs :)

josh146 · 2021-01-27T11:46:58Z

Thanks @mariaschuld!

The only thing one might want to consider is to delete non-tape tests. They have basically no meaning any more if we know we reproduce them in the new tape tests. And the delete will obviously not introduce bugs :)

I thought about this, but I figured it doesn't make sense to delete the old tests until we delete the old core. Plus, it will lead to a massive (temporary) decrease in our coverage rating.

mariaschuld · 2021-01-28T06:04:02Z

pennylane/templates/decorator.py

+        if qml.tape_mode_active():
+            recorder_class = qml.tape.TapeOperationRecorder
+
+        with recorder_class() as rec:


Just to understand, why is it not

if qml.tape_mode_active(): import pennylane as qml recorder_class = qml.tape.TapeOperationRecorder else: recorder_class = OperationRecorder

If tape mode is active to reimport, else you use the old one?

If PennyLane is already imported in the Python session, import pennylane as qml won't actually do anything (e.g., the __init__.py won't be re-executed).

The import pennylane as qml is just to make the qml namespace available in this file, so that we can access the tape_mode_active and TapeOperationRecorder functions/classes.

I couldn't include this import at the top of the file due to circular imports, so that's the only reason it is inside the decorator :(

Makes sense. I never understood circular imports, maybe we should solve the issue once and for all...

pennylane/_qubit_device.py

chaserileyroberts · 2021-01-28T14:49:17Z

pennylane/tape/qnode.py

-        if res_type_namespace == "jax":
-            return __import__(res_type_namespace).numpy.squeeze(res)
-        return __import__(res_type_namespace).squeeze(res)
+        return qml.math.squeeze(res)


chaserileyroberts · 2021-01-28T14:51:15Z

pennylane/templates/decorator.py


    @wraps(func)
    def wrapper(*args, **kwargs):
-        with OperationRecorder() as rec:
+        import pennylane as qml


This import feels hacky and cyclical. Can't we import the needed modules at the top of this file?

hahaha I tried so many combinations (including having it at the top of the file), and constantly ran into cyclical import errors. This was the only combination that didn't result in a cyclical import error.

I'm okay with this for now, since

we can remove it next week when we rip out the old core

PennyLane has already been imported by this point, so there is no overhead here, it simply makes the namespace available to the function.

chaserileyroberts · 2021-01-28T14:55:36Z

pennylane/utils.py

+            for o in operation_list:
+                o.queue()
+                if o.inverse:
+                    o.inv()


Yeah this feels fragile. We'll definitely want to rework how to do this in a nonmutable way.

antalszava

Looks good to me! 💯 🥇 🍾

antalszava · 2021-01-28T15:44:47Z

doc/code/qml_tape.rst

@@ -173,4 +173,4 @@ For more details and examples, please see the tape documentation.
 .. automodapi:: pennylane.tape
    :no-main-docstr:
    :include-all-objects:
-    :skip: enable_tape, disable_tape
+    :skip: QNode, qnode, enable_tape, disable_tape


How come that we are skipping these two? There seems to be other new objects too like TapeCircuitGraph, would we like to skip to show info about them once we deprecate the old core?

Because of all the mocking done by qml.enable_tape(), qml.qnode, qml.QNode, and qml.CircuitGraph all actually point to the pennylane/tape versions 🙂

So I am skipping them from showing at the tape module level, as they will instead be displayed at the qml level

pennylane/circuit_graph.py

tests/beta/test_default_tensor_tf.py

* get hash if tape in execute * Remove hash def for tape; it is accessible through tape.graph.hash * hash move around * format * hash * test * update qubit_device file * Move updating the wires for the state measurement to the vanilla circuit graph class * format * same hash test * comment update

Co-authored-by: antalszava <antalszava@gmail.com>

Co-authored-by: Chase Roberts <chase@xanadu.ai>

josh146 added 6 commits January 26, 2021 17:12

fixes so far

4df4002

more

5951d93

more

973dc5b

Update

016ea50

openqasm

eea1885

inverse

573daa3

josh146 requested a review from mariaschuld January 26, 2021 12:05

josh146 commented Jan 26, 2021

View reviewed changes

pennylane/tape/qnode.py Show resolved Hide resolved

josh146 commented Jan 26, 2021

View reviewed changes

pennylane/templates/layer.py Show resolved Hide resolved

black

ff4aa03

josh146 commented Jan 26, 2021

View reviewed changes

tests/interfaces/test_autograd.py Show resolved Hide resolved

josh146 commented Jan 26, 2021

View reviewed changes