[Templates] Rewrite embeddings as operations (#1156)

* Remove old core

* more

* more tests

* more tests

* more tests

* fixed more tests

* more tests passing

* update jax test

* tests passing

* fix

* fix

* linting

* fix docs

* fix

* fix

* fix

* Update pennylane/tape/operation_recorder.py

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

* fix tests after changing observable underline

* Update tests/test_queuing.py

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

* Update pennylane/circuit_graph.py

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

* Update pennylane/tape/tape.py

* merge master

* Update pennylane/measure.py

Co-authored-by: Maria Schuld <mariaschuld@gmail.com>

* Update doc/code/qml_tape.rst

* Update doc/code/qml_tape.rst

* Update tests/interfaces/test_qnode_torch.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_qnode_torch.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_qnode_autograd.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_qnode_autograd.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_qnode_tf.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_tape_tf.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_qnode_tf.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_tape_autograd.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update tests/interfaces/test_tape_torch.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* Update pennylane/interfaces/torch.py

Co-authored-by: Christina Lee <christina@xanadu.ai>

* suggested changes

* Update pennylane/qnode.py

Co-authored-by: Theodor <theodor@xanadu.ai>

* suggested changes

* fix test

* rewrite amplitudeembedding

* add tests

* finish amplitude embedding

* finish angle embedding

* finish basis embedding

* finished basis embedding

* finished displacement

* finished squeezing

* finished all embeddings

* polish

* black

* removed merge conflict

* polish docstrings

* add interface tests beyond gradients

* typo

* improve angle emb tests

* improve tests further

* add list/tuple input test

* update docstring of basisembedding

* black

* fix bug

* Update .github/CHANGELOG.md

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

* typos

* Update tests/templates/test_embeddings/test_angle.py

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

* applied review suggestions

* black

Co-authored-by: Josh Izaac <josh146@gmail.com>
Co-authored-by: Chase Roberts <chase@xanadu.ai>
Co-authored-by: antalszava <antalszava@gmail.com>
Co-authored-by: Christina Lee <christina@xanadu.ai>
Co-authored-by: Theodor <theodor@xanadu.ai>
Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

.github/CHANGELOG.md

@@ -537,11 +537,13 @@
   1: ──RY(1.35)──╰X──RY(0.422)──╰X──┤   
   ```
 
-- The `QAOAEmbedding` and `BasicEntanglerLayers` are now classes inheriting 
+- The embedding templates, as well as `BasicEntanglerLayers`, are now classes inheriting 
   from `Operation`, and define the ansatz in their `expand()` method. This 
-  change does not affect the user interface. 
+  change does not affect the user interface.
+  [(#1138)](https://github.com/PennyLaneAI/pennylane/pull/1138)
+  [(#1156)](https://github.com/PennyLaneAI/pennylane/pull/1156)
 
-  For convenience, the class has a method that returns the shape of the 
+  For convenience, `BasicEntanglerLayers` has a method that returns the shape of the 
   trainable parameter tensor, i.e.,
 
   ```python

pennylane/templates/embeddings/amplitude.py

@@ -12,82 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 r"""
-Contains the ``AmplitudeEmbedding`` template.
+Contains the AmplitudeEmbedding template.
 """
 # pylint: disable-msg=too-many-branches,too-many-arguments,protected-access
 import warnings
 import numpy as np
 
 import pennylane as qml
-from pennylane.templates.decorator import template
+from pennylane.operation import Operation, AnyWires
 from pennylane.ops import QubitStateVector
 from pennylane.wires import Wires
 
 # tolerance for normalization
 TOLERANCE = 1e-10
 
 
-def _preprocess(features, wires, pad_with, normalize):
-    """Validate and pre-process inputs as follows:
-
-    * Check that the features tensor is one-dimensional.
-    * If pad_with is None, check that the first dimension of the features tensor
-      has length :math:`2^n` where :math:`n` is the number of qubits. Else check that the
-      first dimension of the features tensor is not larger than :math:`2^n` and pad features with value if necessary.
-    * If normalize is false, check that first dimension of features is normalised to one. Else, normalise the
-      features tensor.
-
-    Args:
-        features (tensor_like): input features to pre-process
-        wires (Wires): wires that template acts on
-        pad_with (float): constant used to pad the features tensor to required dimension
-        normalize (bool): whether or not to normalize the features vector
-
-    Returns:
-        tensor: pre-processed features
-    """
-
-    shape = qml.math.shape(features)
-
-    # check shape
-    if len(shape) != 1:
-        raise ValueError(f"Features must be a one-dimensional tensor; got shape {shape}.")
-
-    n_features = shape[0]
-    if pad_with is None and n_features != 2 ** len(wires):
-        raise ValueError(
-            f"Features must be of length {2 ** len(wires)}; got length {n_features}. "
-            f"Use the 'pad' argument for automated padding."
-        )
-
-    if pad_with is not None and n_features > 2 ** len(wires):
-        raise ValueError(
-            f"Features must be of length {2 ** len(wires)} or "
-            f"smaller to be padded; got length {n_features}."
-        )
-
-    # pad
-    if pad_with is not None and n_features < 2 ** len(wires):
-        padding = [pad_with] * (2 ** len(wires) - n_features)
-        features = qml.math.concatenate([features, padding], axis=0)
-
-    # normalize
-    norm = qml.math.sum(qml.math.abs(features) ** 2)
-
-    if not qml.math.allclose(norm, 1.0, atol=TOLERANCE):
-        if normalize or pad_with:
-            features = features / np.sqrt(norm)
-        else:
-            raise ValueError(
-                f"Features must be a vector of length 1.0; got length {norm}."
-                "Use 'normalize=True' to automatically normalize."
-            )
-
-    return features
-
-
-@template
-def AmplitudeEmbedding(features, wires, pad_with=None, normalize=False, pad=None):
+class AmplitudeEmbedding(Operation):
     r"""Encodes :math:`2^n` features into the amplitude vector of :math:`n` qubits.
 
     By setting ``pad_with`` to a real or complex number, ``features`` is automatically padded to dimension
@@ -108,9 +48,8 @@ def AmplitudeEmbedding(features, wires, pad_with=None, normalize=False, pad=None
         gradients with respect to the features cannot be computed by PennyLane.
 
     Args:
-        features (tensor-like): input vector of length ``2^n``, or less if `pad_with` is specified
-        wires (Iterable or :class:`.wires.Wires`): Wires that the template acts on.
-            Accepts an iterable of numbers or strings, or a Wires object.
+        features (tensor_like): input tensor of dimension ``(2^n,)``, or less if `pad_with` is specified
+        wires (Iterable): wires that the template acts on
         pad_with (float or complex): if not None, the input is padded with this constant to size :math:`2^n`
         normalize (bool): whether to automatically normalize the features
         pad (float or complex): same as `pad`, to be deprecated
@@ -142,21 +81,7 @@ def circuit(f=None):
         **Differentiating with respect to the features**
 
         Due to non-trivial classical processing to construct the state preparation circuit,
-        the features argument is **not always differentiable**.
-
-        .. code-block:: python
-
-            from pennylane import numpy as np
-
-            @qml.qnode(dev)
-            def circuit(f):
-                AmplitudeEmbedding(features=f, wires=range(2))
-                return qml.expval(qml.PauliZ(0))
-
-        >>> g = qml.grad(circuit, argnum=0)
-        >>> f = np.array([1, 1, 1, 1], requires_grad=True)
-        >>> g(f)
-        ValueError: Cannot differentiate wrt parameter(s) {0, 1, 2, 3}.
+        the features argument is in general **not differentiable**.
 
         **Normalization**
 
@@ -216,17 +141,82 @@ def circuit(f=None):
 
     """
 
-    wires = Wires(wires)
+    num_params = 1
+    num_wires = AnyWires
+    par_domain = "A"
+
+    def __init__(self, features, wires, pad_with=None, normalize=False, pad=None, do_queue=True):
+
+        # pad is replaced with the more verbose pad_with
+        if pad is not None:
+            warnings.warn(
+                "The pad argument will be replaced by the pad_with option in future versions of PennyLane.",
+                PendingDeprecationWarning,
+            )
+            if pad_with is None:
+                pad_with = pad
+
+        wires = Wires(wires)
+        self.pad_with = pad_with
+        self.normalize = normalize
+
+        features = self._preprocess(features, wires, pad_with, normalize)
+        super().__init__(features, wires=wires, do_queue=do_queue)
+
+    def expand(self):
+
+        with qml.tape.QuantumTape() as tape:
+            QubitStateVector(self.parameters[0], wires=self.wires)
+
+        return tape
 
-    # pad is replaced with the more verbose pad_with
-    if pad is not None:
-        warnings.warn(
-            "The pad argument will be replaced by the pad_with option in future versions of PennyLane.",
-            PendingDeprecationWarning,
-        )
-        if pad_with is None:
-            pad_with = pad
+    @staticmethod
+    def _preprocess(features, wires, pad_with, normalize):
+        """Validate and pre-process inputs as follows:
 
-    features = _preprocess(features, wires, pad_with, normalize)
+        * Check that the features tensor is one-dimensional.
+        * If pad_with is None, check that the first dimension of the features tensor
+          has length :math:`2^n` where :math:`n` is the number of qubits. Else check that the
+          first dimension of the features tensor is not larger than :math:`2^n` and pad features with value if necessary.
+        * If normalize is false, check that first dimension of features is normalised to one. Else, normalise the
+          features tensor.
+        """
+
+        shape = qml.math.shape(features)
+
+        # check shape
+        if len(shape) != 1:
+            raise ValueError(f"Features must be a one-dimensional tensor; got shape {shape}.")
+
+        n_features = shape[0]
+        if pad_with is None and n_features != 2 ** len(wires):
+            raise ValueError(
+                f"Features must be of length {2 ** len(wires)}; got length {n_features}. "
+                f"Use the 'pad' argument for automated padding."
+            )
+
+        if pad_with is not None and n_features > 2 ** len(wires):
+            raise ValueError(
+                f"Features must be of length {2 ** len(wires)} or "
+                f"smaller to be padded; got length {n_features}."
+            )
 
-    QubitStateVector(features, wires=wires)
+        # pad
+        if pad_with is not None and n_features < 2 ** len(wires):
+            padding = [pad_with] * (2 ** len(wires) - n_features)
+            features = qml.math.concatenate([features, padding], axis=0)
+
+        # normalize
+        norm = qml.math.sum(qml.math.abs(features) ** 2)
+
+        if not qml.math.allclose(norm, 1.0, atol=TOLERANCE):
+            if normalize or pad_with:
+                features = features / np.sqrt(norm)
+            else:
+                raise ValueError(
+                    f"Features must be a vector of length 1.0; got length {norm}."
+                    "Use 'normalize=True' to automatically normalize."
+                )
+
+        features = qml.math.cast(features, np.complex128)
+        return features

pennylane/templates/embeddings/angle.py

@@ -16,40 +16,14 @@
 """
 # pylint: disable-msg=too-many-branches,too-many-arguments,protected-access
 import pennylane as qml
-from pennylane.templates.decorator import template
-from pennylane.templates import broadcast
-from pennylane.wires import Wires
+from pennylane.ops import RX, RY, RZ
+from pennylane.operation import Operation, AnyWires
 
 
-def _preprocess(features, wires):
-    """Validate and pre-process inputs as follows:
+ROT = {"X": RX, "Y": RY, "Z": RZ}
 
-    * Check that the features tensor is one-dimensional.
-    * Check that the first dimension of the features tensor
-      has length :math:`n` or less, where :math:`n` is the number of qubits.
 
-    Args:
-        features (tensor_like): input features to pre-process
-        wires (Wires): wires that template acts on
-
-    Returns:
-        int: number of features
-    """
-    shape = qml.math.shape(features)
-
-    if len(shape) != 1:
-        raise ValueError(f"Features must be a one-dimensional tensor; got shape {shape}.")
-
-    n_features = shape[0]
-    if n_features > len(wires):
-        raise ValueError(
-            f"Features must be of length {len(wires)} or less; got length {n_features}."
-        )
-    return n_features
-
-
-@template
-def AngleEmbedding(features, wires, rotation="X"):
+class AngleEmbedding(Operation):
     r"""
     Encodes :math:`N` features into the rotation angles of :math:`n` qubits, where :math:`N \leq n`.
 
@@ -66,26 +40,41 @@ def AngleEmbedding(features, wires, rotation="X"):
     ``features`` than rotations, the circuit does not apply the remaining rotation gates.
 
     Args:
-        features (array): input array of shape ``(N,)``, where N is the number of input features to embed,
+        features (tensor_like): input tensor of shape ``(N,)``, where N is the number of input features to embed,
             with :math:`N\leq n`
-        wires (Iterable or Wires): Wires that the template acts on. Accepts an iterable of numbers or strings, or
-            a Wires object.
+        wires (Iterable): wires that the template acts on
         rotation (str): type of rotations used
-
     """
 
-    wires = Wires(wires)
-    n_features = _preprocess(features, wires)
-    wires = wires[:n_features]
+    num_params = 1
+    num_wires = AnyWires
+    par_domain = "A"
+
+    def __init__(self, features, wires, rotation="X", do_queue=True):
+
+        if rotation not in ROT:
+            raise ValueError(f"Rotation option {rotation} not recognized.")
+        self.rotation = ROT[rotation]
+
+        shape = qml.math.shape(features)
+        if len(shape) != 1:
+            raise ValueError(f"Features must be a one-dimensional tensor; got shape {shape}.")
+        n_features = shape[0]
+        if n_features > len(wires):
+            raise ValueError(
+                f"Features must be of length {len(wires)} or less; got length {n_features}."
+            )
+
+        wires = wires[:n_features]
+        super().__init__(features, wires=wires, do_queue=do_queue)
+
+    def expand(self):
 
-    if rotation == "X":
-        broadcast(unitary=qml.RX, pattern="single", wires=wires, parameters=features)
+        features = self.parameters[0]
 
-    elif rotation == "Y":
-        broadcast(unitary=qml.RY, pattern="single", wires=wires, parameters=features)
+        with qml.tape.QuantumTape() as tape:
 
-    elif rotation == "Z":
-        broadcast(unitary=qml.RZ, pattern="single", wires=wires, parameters=features)
+            for i in range(len(self.wires)):
+                self.rotation(features[i], wires=self.wires[i])
 
-    else:
-        raise ValueError(f"Rotation option {rotation} not recognized.")
+        return tape