Daily rc sync to master (#4281)

* Support `HardwareHamiltonian` pulses in `stoch_pulse_grad` (#4215) * single out gradient transform checks * rename stochastic pulse gradient file * unify gradient_analysis and grad_method_validation * continue restructure of analysis+validation * CV * black * modularize more * more modularizing * black * tiny [skip ci] * [skip ci] lint * remove dummy test * test fix * add test file to linting test file * test fixes, docstrings * code review * docstring gradient_analysis_and_grad_method_validation * move first fun * code review:move functions * test regex * regexs * move and promote reorder_grads * tmp * more tmp * test cases, contractions * lint * docstring * even more tmp * cleanup * black * tmp * lint * move stoch_pulse_gradient.. files back to pulse_gradient... * move stoch_pulse_gradient.. files back to pulse_gradient... * lint * rename * extend functions and tests * lint and black * changelog * improve * update example to include non-Pauli word generator * add jit test with pauli sentence * tmp * debugging, docstring, extend test * review * optimize for Pauli words * Apply suggestions from code review Co-authored-by: Romain Moyard <[email protected]> * test cases code review * fix parametrization * drafting * working prototype * finish merge; cleanup * changelog * comments * [skip ci] * raising an error; cleanup [skip ci] * Apply suggestions from code review Co-authored-by: Korbinian Kottmann <[email protected]> * change contraction idea * typo in docs * tests * remove prints * fix test * test descriptions * fix merge * format * code review; test coverage * coverage reordering * fix * trigger CI * trigger * clear caches * trigger * trigger --------- Co-authored-by: Romain Moyard <[email protected]> Co-authored-by: Korbinian Kottmann <[email protected]> Co-authored-by: Korbinian Kottmann <[email protected]> * Fix batching of derivative tapes in autograd (#4245) * Fix `expval` of `Sum` with broadcasting (#4275) * fix bug and add test * changelog addition * Various doc fixes (#4268) * exclude files from pr --------- Co-authored-by: David Wierichs <[email protected]> Co-authored-by: Romain Moyard <[email protected]> Co-authored-by: Korbinian Kottmann <[email protected]> Co-authored-by: Korbinian Kottmann <[email protected]> Co-authored-by: Christina Lee <[email protected]> Co-authored-by: Edward Jiang <[email protected]> Co-authored-by: GitHub Actions Bot <> Co-authored-by: Matthew Silverman <[email protected]>
PennyLaneAI · Jun 22, 2023 · 111606e · 111606e
1 parent b5613af
commit 111606e
Show file tree

Hide file tree

Showing 6 changed files with 58 additions and 37 deletions.
diff --git a/doc/releases/changelog-0.31.0.md b/doc/releases/changelog-0.31.0.md
@@ -447,6 +447,7 @@
 
 * Allow for `Sum` observables with trainable parameters.
   [(#4251)](https://github.com/PennyLaneAI/pennylane/pull/4251)
+  [(#4275)](https://github.com/PennyLaneAI/pennylane/pull/4275)
 
 <h3>Contributors ✍️</h3>
 

diff --git a/pennylane/devices/default_qubit.py b/pennylane/devices/default_qubit.py
@@ -570,10 +570,13 @@ def expval(self, observable, shot_range=None, bin_size=None):
             Hamiltonian is not NumPy or Autograd
 
         """
+        is_state_batched = self._ndim(self.state) == 2
         # intercept Sums
         if isinstance(observable, Sum) and not self.shots:
             return measure(
-                ExpectationMP(observable.map_wires(self.wire_map)), self._pre_rotated_state
+                ExpectationMP(observable.map_wires(self.wire_map)),
+                self._pre_rotated_state,
+                is_state_batched,
             )
 
         # intercept other Hamiltonians
@@ -592,7 +595,7 @@ def expval(self, observable, shot_range=None, bin_size=None):
 
         if backprop_mode:
             # TODO[dwierichs]: This branch is not adapted to broadcasting yet
-            if self._ndim(self.state) == 2:
+            if is_state_batched:
                 raise NotImplementedError(
                     "Expectation values of Hamiltonians for interface!=None are "
                     "not supported together with parameter broadcasting yet"
@@ -632,7 +635,7 @@ def expval(self, observable, shot_range=None, bin_size=None):
             Hmat = observable.sparse_matrix(wire_order=self.wires)
 
             state = qml.math.toarray(self.state)
-            if self._ndim(state) == 2:
+            if is_state_batched:
                 res = qml.math.array(
                     [
                         csr_matrix.dot(

diff --git a/pennylane/pauli/utils.py b/pennylane/pauli/utils.py
@@ -51,8 +51,7 @@ def _wire_map_from_pauli_pair(pauli_word_1, pauli_word_2):
     return {label: i for i, label in enumerate(wire_labels)}
 
 
-@singledispatch
-def is_pauli_word(observable):  # pylint:disable=unused-argument
+def is_pauli_word(observable):
     """
     Checks if an observable instance consists only of Pauli and Identity Operators.
 
@@ -93,36 +92,45 @@ def is_pauli_word(observable):  # pylint:disable=unused-argument
     >>> is_pauli_word(4 * qml.PauliX(0) @ qml.PauliZ(0))
     True
     """
+    return _is_pauli_word(observable)
+
+
+@singledispatch
+def _is_pauli_word(observable):  # pylint:disable=unused-argument
+    """
+    Private implementation of is_pauli_word, to prevent all of the
+    registered functions from appearing in the Sphinx docs.
+    """
     return False
 
 
-@is_pauli_word.register(PauliX)
-@is_pauli_word.register(PauliY)
-@is_pauli_word.register(PauliZ)
-@is_pauli_word.register(Identity)
+@_is_pauli_word.register(PauliX)
+@_is_pauli_word.register(PauliY)
+@_is_pauli_word.register(PauliZ)
+@_is_pauli_word.register(Identity)
 def _is_pw_pauli(
     observable: Union[PauliX, PauliY, PauliZ, Identity]
 ):  # pylint:disable=unused-argument
     return True
 
 
-@is_pauli_word.register
+@_is_pauli_word.register
 def _is_pw_tensor(observable: Tensor):
     pauli_word_names = ["Identity", "PauliX", "PauliY", "PauliZ"]
     return set(observable.name).issubset(pauli_word_names)
 
 
-@is_pauli_word.register
+@_is_pauli_word.register
 def _is_pw_ham(observable: Hamiltonian):
     return False if len(observable.ops) != 1 else is_pauli_word(observable.ops[0])
 
 
-@is_pauli_word.register
+@_is_pauli_word.register
 def _is_pw_prod(observable: Prod):
     return all(is_pauli_word(op) for op in observable)
 
 
-@is_pauli_word.register
+@_is_pauli_word.register
 def _is_pw_sprod(observable: SProd):
     return is_pauli_word(observable.base)
 

diff --git a/pennylane/transforms/core/transform.py b/pennylane/transforms/core/transform.py
@@ -58,9 +58,9 @@ def post_processing_fn(results):
 
             return [tape1, tape2], post_processing_fn
 
-    Of course, we want to be able to apply this transform on `qfunc` and `qnodes`. That's where the `transform` function
+    Of course, we want to be able to apply this transform on ``qfunc`` and ``qnodes``. That's where the ``transform`` function
     comes into play. This function validates the signature of your quantum transform and dispatches it on the different
-    object. Let's define a circuit as a qfunc and as qnode.
+    object. Let's define a circuit as a qfunc and as a qnode.
 
         .. code-block:: python
 
@@ -85,7 +85,7 @@ def qnode_circuit(a):
 
     Now you can use the dispatched transform directly on qfunc and qnodes.
 
-    For QNodes, the dispatched transform populates the `TransformProgram` of your QNode. The transform and its
+    For QNodes, the dispatched transform populates the ``TransformProgram`` of your QNode. The transform and its
     processing function are applied in the execution.
 
     >>> transformed_qnode = dispatched_transform(qfunc_circuit)

diff --git a/pennylane/transforms/decompositions/single_qubit_unitary.py b/pennylane/transforms/decompositions/single_qubit_unitary.py
@@ -417,10 +417,10 @@ def _zxz_decomposition(U, wire, return_global_phase=False):
 def one_qubit_decomposition(U, wire, rotations="ZYZ", return_global_phase=False):
     r"""Decompose a one-qubit unitary :math:`U` in terms of elementary operations. (batched operation)
 
-    Any one qubit unitary operation can be implemented upto a global phase by composing RX, RY,
+    Any one qubit unitary operation can be implemented up to a global phase by composing RX, RY,
     and RZ gates.
 
-    Currently supported values for `rotations` are "ZYZ", "XYX", and "ZXZ".
+    Currently supported values for ``rotations`` are "ZYZ", "XYX", and "ZXZ".
 
     Args:
         U (tensor): A :math:`2 \times 2` unitary matrix.
@@ -431,7 +431,7 @@ def one_qubit_decomposition(U, wire, rotations="ZYZ", return_global_phase=False)
 
     Returns:
         list[Operation]: Returns a list of gates which when applied in the order of appearance in
-        the list is equivalent to the unitary :math:`U` up to a global phase. If `return_global_phase=True`,
+        the list is equivalent to the unitary :math:`U` up to a global phase. If ``return_global_phase=True``,
         the global phase is returned as the last element of the list.
 
     **Example**

diff --git a/tests/devices/test_default_qubit.py b/tests/devices/test_default_qubit.py
@@ -19,6 +19,7 @@
 # pylint: disable=protected-access,cell-var-from-loop
 import math
 
+from functools import partial
 import pytest
 
 import pennylane as qml
@@ -2361,22 +2362,28 @@ def test_Hamiltonian_filtered_from_rotations(self, mocker):
         assert qml.equal(call_args.measurements[0], qml.expval(qml.PauliX(0)))
 
 
+@pytest.mark.parametrize("is_state_batched", [False, True])
 class TestSumSupport:
     """Tests for custom Sum support in DefaultQubit."""
 
-    expected_grad = [-np.sin(1.3), np.cos(1.3)]
+    @staticmethod
+    def expected_grad(is_state_batched):
+        if is_state_batched:
+            return [[-np.sin(1.3), -np.sin(0.4)], [np.cos(1.3), np.cos(0.4)]]
+        return [-np.sin(1.3), np.cos(1.3)]
 
     @staticmethod
-    def circuit(y, z):
-        qml.RX(1.3, 0)
+    def circuit(y, z, is_state_batched):
+        rx_param = [1.3, 0.4] if is_state_batched else 1.3
+        qml.RX(rx_param, 0)
         return qml.expval(
             qml.sum(
                 qml.s_prod(y, qml.PauliY(0)),
                 qml.s_prod(z, qml.PauliZ(0)),
             )
         )
 
-    def test_super_expval_not_called(self, mocker):
+    def test_super_expval_not_called(self, is_state_batched, mocker):
         """Tests basic expval result, and ensures QubitDevice.expval is not called."""
         dev = qml.device("default.qubit", wires=1)
         spy = mocker.spy(qml.QubitDevice, "expval")
@@ -2385,49 +2392,51 @@ def test_super_expval_not_called(self, mocker):
         spy.assert_not_called()
 
     @pytest.mark.autograd
-    def test_trainable_autograd(self):
+    def test_trainable_autograd(self, is_state_batched):
         """Tests that coeffs passed to a sum are trainable with autograd."""
+        if is_state_batched:
+            pytest.skip(msg="Broadcasting, qml.jacobian and new return types do not work together")
         dev = qml.device("default.qubit", wires=1)
         qnode = qml.QNode(self.circuit, dev, interface="autograd")
         y, z = np.array([1.1, 2.2])
-        actual = qml.grad(qnode)(y, z)
-        assert np.allclose(actual, self.expected_grad)
+        actual = qml.grad(qnode, argnum=[0, 1])(y, z, is_state_batched)
+        assert np.allclose(actual, self.expected_grad(is_state_batched))
 
     @pytest.mark.torch
-    def test_trainable_torch(self):
+    def test_trainable_torch(self, is_state_batched):
         """Tests that coeffs passed to a sum are trainable with torch."""
         import torch
 
         dev = qml.device("default.qubit", wires=1)
         qnode = qml.QNode(self.circuit, dev, interface="torch")
         y, z = torch.tensor(1.1, requires_grad=True), torch.tensor(2.2, requires_grad=True)
-        qnode(y, z).backward()
-        actual = [y.grad, z.grad]
-        assert np.allclose(actual, self.expected_grad)
+        _qnode = partial(qnode, is_state_batched=is_state_batched)
+        actual = torch.stack(torch.autograd.functional.jacobian(_qnode, (y, z)))
+        assert np.allclose(actual, self.expected_grad(is_state_batched))
 
     @pytest.mark.tf
-    def test_trainable_tf(self):
+    def test_trainable_tf(self, is_state_batched):
         """Tests that coeffs passed to a sum are trainable with tf."""
         import tensorflow as tf
 
         dev = qml.device("default.qubit", wires=1)
         qnode = qml.QNode(self.circuit, dev, interface="tensorflow")
         y, z = tf.Variable(1.1, dtype=tf.float64), tf.Variable(2.2, dtype=tf.float64)
         with tf.GradientTape() as tape:
-            res = qnode(y, z)
-        actual = tape.gradient(res, [y, z])
-        assert np.allclose(actual, self.expected_grad)
+            res = qnode(y, z, is_state_batched)
+        actual = tape.jacobian(res, [y, z])
+        assert np.allclose(actual, self.expected_grad(is_state_batched))
 
     @pytest.mark.jax
-    def test_trainable_jax(self):
+    def test_trainable_jax(self, is_state_batched):
         """Tests that coeffs passed to a sum are trainable with jax."""
         import jax
 
         dev = qml.device("default.qubit", wires=1)
         qnode = qml.QNode(self.circuit, dev, interface="jax")
         y, z = jax.numpy.array([1.1, 2.2])
-        actual = jax.grad(qnode, argnums=[0, 1])(y, z)
-        assert np.allclose(actual, self.expected_grad)
+        actual = jax.jacobian(qnode, argnums=[0, 1])(y, z, is_state_batched)
+        assert np.allclose(actual, self.expected_grad(is_state_batched))
 
 
 class TestGetBatchSize: