Source code for pennylane.devices.default_qubit_tf

# Copyright 2018-2021 Xanadu Quantum Technologies Inc.

# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at

#     http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""This module contains a TensorFlow implementation of the :class:`~.DefaultQubitLegacy`
reference plugin.
"""
import itertools
import numpy as np
import semantic_version

import pennylane as qml

try:
    import tensorflow as tf

    if tf.__version__[0] == "1":  # pragma: no cover
        raise ImportError("default.qubit.tf device requires TensorFlow>=2.0")

    from tensorflow.python.framework.errors_impl import InvalidArgumentError

    SUPPORTS_APPLY_OPS = semantic_version.match(">=2.3.0", tf.__version__)

except ImportError as e:  # pragma: no cover
    raise ImportError("default.qubit.tf device requires TensorFlow>=2.0") from e


from pennylane.math.single_dispatch import _ndim_tf
from . import DefaultQubitLegacy
from .default_qubit_legacy import tolerance


[docs]class DefaultQubitTF(DefaultQubitLegacy): """Simulator plugin based on ``"default.qubit.legacy"``, written using TensorFlow. **Short name:** ``default.qubit.tf`` This device provides a pure-state qubit simulator written using TensorFlow. As a result, it supports classical backpropagation as a means to compute the Jacobian. This can be faster than the parameter-shift rule for analytic quantum gradients when the number of parameters to be optimized is large. To use this device, you will need to install TensorFlow: .. code-block:: console pip install tensorflow>=2.0 **Example** The ``default.qubit.tf`` is designed to be used with end-to-end classical backpropagation (``diff_method="backprop"``) with the TensorFlow interface. This is the default method of differentiation when creating a QNode with this device. Using this method, the created QNode is a 'white-box', and is tightly integrated with your TensorFlow computation: >>> dev = qml.device("default.qubit.tf", wires=1) >>> @qml.qnode(dev, interface="tf", diff_method="backprop") ... def circuit(x): ... qml.RX(x[1], wires=0) ... qml.Rot(x[0], x[1], x[2], wires=0) ... return qml.expval(qml.Z(0)) >>> weights = tf.Variable([0.2, 0.5, 0.1]) >>> with tf.GradientTape() as tape: ... res = circuit(weights) >>> print(tape.gradient(res, weights)) tf.Tensor([-2.2526717e-01 -1.0086454e+00 1.3877788e-17], shape=(3,), dtype=float32) Autograph mode will also work when using classical backpropagation: >>> @tf.function ... def cost(weights): ... return tf.reduce_sum(circuit(weights)**3) - 1 >>> with tf.GradientTape() as tape: ... res = cost(weights) >>> print(tape.gradient(res, weights)) tf.Tensor([-3.5471588e-01 -1.5882589e+00 3.4694470e-17], shape=(3,), dtype=float32) There are a couple of things to keep in mind when using the ``"backprop"`` differentiation method for QNodes: * You must use the ``"tf"`` interface for classical backpropagation, as TensorFlow is used as the device backend. * Only exact expectation values, variances, and probabilities are differentiable. Creation of a backpropagation QNode with finite shots raises an error. If you do try and take a derivative with finite shots on this device, the gradient will be ``None``. If you wish to use a different machine-learning interface, or prefer to calculate quantum gradients using the ``parameter-shift`` or ``finite-diff`` differentiation methods, consider using the ``default.qubit`` device instead. Args: wires (int, Iterable[Number, str]): Number of subsystems represented by the device, or iterable that contains unique labels for the subsystems as numbers (i.e., ``[-1, 0, 2]``) or strings (``['ancilla', 'q1', 'q2']``). Default 1 if not specified. shots (None, int): How many times the circuit should be evaluated (or sampled) to estimate the expectation values. Defaults to ``None`` if not specified, which means that the device returns analytical results. If ``shots > 0`` is used, the ``diff_method="backprop"`` QNode differentiation method is not supported and it is recommended to consider switching device to ``default.qubit`` and using ``diff_method="parameter-shift"``. """ name = "Default qubit (TensorFlow) PennyLane plugin" short_name = "default.qubit.tf" _asarray = staticmethod(tf.convert_to_tensor) _dot = staticmethod(lambda x, y: tf.tensordot(x, y, axes=1)) _abs = staticmethod(tf.abs) _reduce_sum = staticmethod(tf.reduce_sum) _reshape = staticmethod(tf.reshape) _flatten = staticmethod(lambda tensor: tf.reshape(tensor, [-1])) _gather = staticmethod(tf.gather) _einsum = staticmethod(tf.einsum) _cast = staticmethod(tf.cast) _transpose = staticmethod(tf.transpose) _tensordot = staticmethod(tf.tensordot) _conj = staticmethod(tf.math.conj) _real = staticmethod(tf.math.real) _imag = staticmethod(tf.math.imag) _roll = staticmethod(tf.roll) _stack = staticmethod(tf.stack) _size = staticmethod(tf.size) _ndim = staticmethod(_ndim_tf) @staticmethod def _const_mul(constant, array): return constant * array @staticmethod def _asarray(array, dtype=None): if isinstance(array, tf.Tensor): if dtype is None or dtype == array.dtype: return array return tf.cast(array, dtype) try: res = tf.convert_to_tensor(array, dtype) except InvalidArgumentError: axis = 0 res = tf.concat([tf.reshape(i, [-1]) for i in array], axis) if dtype is not None: res = tf.cast(res, dtype) return res def __init__(self, wires, *, shots=None, analytic=None): r_dtype = tf.float64 c_dtype = tf.complex128 super().__init__(wires, shots=shots, r_dtype=r_dtype, c_dtype=c_dtype, analytic=analytic) # prevent using special apply method for this gate due to slowdown in TF implementation del self._apply_ops["CZ"] # Versions of TF before 2.3.0 do not support using the special apply methods as they # raise an error when calculating the gradient. For versions of TF after 2.3.0, # special apply methods are also not supported when using more than 8 wires due to # limitations with TF slicing. if not SUPPORTS_APPLY_OPS or self.num_wires > 8: self._apply_ops = {}
[docs] @classmethod def capabilities(cls): capabilities = super().capabilities().copy() capabilities.update(passthru_interface="tf") return capabilities
@staticmethod def _scatter(indices, array, new_dimensions): indices = np.expand_dims(indices, 1) return tf.scatter_nd(indices, array, new_dimensions) def _get_batch_size(self, tensor, expected_shape, expected_size): """Determine whether a tensor has an additional batch dimension for broadcasting, compared to an expected_shape. Differs from QubitDevice implementation by the exception made for abstract tensors.""" try: size = self._size(tensor) ndim = qml.math.ndim(tensor) if ndim > len(expected_shape) or size > expected_size: return size // expected_size except (ValueError, tf.errors.OperatorNotAllowedInGraphError) as err: # This except clause covers the usage of tf.function, which is not compatible # with `DefaultQubit._get_batch_size` if not qml.math.is_abstract(tensor): raise err return None def _apply_state_vector(self, state, device_wires): """Initialize the internal state vector in a specified state. Args: state (array[complex]): normalized input state of length ``2**len(wires)`` or broadcasted state of shape ``(batch_size, 2**len(wires))`` device_wires (Wires): wires that get initialized in the state This implementation only adds a check for parameter broadcasting when initializing a quantum state on subsystems of the device. """ # translate to wire labels used by device device_wires = self.map_wires(device_wires) dim = 2 ** len(device_wires) state = self._asarray(state, dtype=self.C_DTYPE) batch_size = self._get_batch_size(state, (dim,), dim) output_shape = [2] * self.num_wires if batch_size: output_shape.insert(0, batch_size) if not (state.shape in [(dim,), (batch_size, dim)]): raise ValueError("State vector must have shape (2**wires,) or (batch_size, 2**wires).") if not qml.math.is_abstract(state): norm = qml.math.linalg.norm(state, axis=-1, ord=2) if not qml.math.allclose(norm, 1.0, atol=tolerance): raise ValueError("Sum of amplitudes-squared does not equal one.") if len(device_wires) == self.num_wires and sorted(device_wires) == device_wires: # Initialize the entire device state with the input state self._state = self._reshape(state, output_shape) return # generate basis states on subset of qubits via the cartesian product basis_states = np.array(list(itertools.product([0, 1], repeat=len(device_wires)))) # get basis states to alter on full set of qubits unravelled_indices = np.zeros((2 ** len(device_wires), self.num_wires), dtype=int) unravelled_indices[:, device_wires] = basis_states # get indices for which the state is changed to input state vector elements ravelled_indices = np.ravel_multi_index(unravelled_indices.T, [2] * self.num_wires) if batch_size: # This is the only logical branch that differs from DefaultQubitLegacy raise NotImplementedError( "Parameter broadcasting is not supported together with initializing the state " "vector of a subsystem of the device when using DefaultQubitTF." ) # The following line is unchanged in the "else"-clause in DefaultQubitLegacy's implementation state = self._scatter(ravelled_indices, state, [2**self.num_wires]) state = self._reshape(state, output_shape) self._state = self._asarray(state, dtype=self.C_DTYPE)