Source code for pennylane.workflow.interfaces.tensorflow_autograph
# 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 functions for adding the TensorFlow Autograph interface
to a PennyLane Device class.
"""
# pylint: disable=too-many-arguments,too-many-branches,too-many-statements
from functools import reduce
import numpy as np
import tensorflow as tf
import pennylane as qml
from pennylane.measurements import SampleMP, StateMP
from .tensorflow import (
_res_restructured,
_to_tensors,
set_parameters_on_copy,
)
def _compute_vjp(dy, jacs, multi_measurements, has_partitioned_shots):
# compute the vector-Jacobian product dy @ jac
# for a list of dy's and Jacobian matrices.
vjps = []
for dy_, jac_, multi in zip(dy, jacs, multi_measurements):
dy_ = dy_ if has_partitioned_shots else (dy_,)
jac_ = jac_ if has_partitioned_shots else (jac_,)
shot_vjps = []
for d, j in zip(dy_, jac_):
# see xfail test: test_tensorflow_qnode_default_qubit_2.py:test_autograph_adjoint_multi_out
# And Issue #5078
# pragma: no cover
if multi: # pragma: no cover
shot_vjps.append(qml.gradients.compute_vjp_multi(d, j))
else:
shot_vjps.append(qml.gradients.compute_vjp_single(d, j))
vjp = qml.math.sum(qml.math.stack(shot_vjps), 0)
vjps.extend(vjp)
return vjps
def _flatten_nested_list(x):
"""
Recursively flatten the list
"""
if not isinstance(x, (tuple, list)):
return [x]
return reduce(lambda a, y: a + _flatten_nested_list(y), x, [])
def _jac_restructured(jacs, tapes):
"""
Reconstruct the nested tuple structure of the jacobian of a list of tapes
"""
start = 0
jacs_nested = []
for tape in tapes:
num_meas = len(tape.measurements)
num_params = len(tape.trainable_params)
tape_jacs = tuple(jacs[start : start + num_meas * num_params])
tape_jacs = tuple(
tuple(tape_jacs[i * num_params : (i + 1) * num_params]) for i in range(num_meas)
)
while isinstance(tape_jacs, tuple) and len(tape_jacs) == 1:
tape_jacs = tape_jacs[0]
jacs_nested.append(tape_jacs)
start += num_meas * num_params
return tuple(jacs_nested)
[docs]def execute(
tapes,
device,
execute_fn,
gradient_fn,
gradient_kwargs,
_n=1,
max_diff=2,
grad_on_execution=None,
):
"""Execute a batch of tapes with TensorFlow parameters on a device.
Args:
tapes (Sequence[.QuantumTape]): batch of tapes to execute
device (pennylane.Device): Device to use to execute the batch of tapes.
If the device does not provide a ``batch_execute`` method,
by default the tapes will be executed in serial.
execute_fn (callable): The execution function used to execute the tapes
during the forward pass. This function must return a tuple ``(results, jacobians)``.
If ``jacobians`` is an empty list, then ``gradient_fn`` is used to
compute the gradients during the backwards pass.
gradient_kwargs (dict): dictionary of keyword arguments to pass when
determining the gradients of tapes
gradient_fn (callable): the gradient function to use to compute quantum gradients
_n (int): a positive integer used to track nesting of derivatives, for example
if the nth-order derivative is requested.
max_diff (int): If ``gradient_fn`` is a gradient transform, this option specifies
the maximum number of derivatives to support. Increasing this value allows
for higher order derivatives to be extracted, at the cost of additional
(classical) computational overhead during the backwards pass.
grad_on_execution (bool): Whether the gradients should be computed on execution or not.
Returns:
list[list[tf.Tensor]]: A nested list of tape results. Each element in
the returned list corresponds in order to the provided tapes.
"""
all_params = []
parameters = []
lens = []
trainable = []
output_types = []
# assumes all tapes have the same shot vector
has_partitioned_shots = tapes[0].shots.has_partitioned_shots
num_shot_copies = tapes[0].shots.num_copies or 1
for tape in tapes:
# store the trainable parameters
params = tape.get_parameters(trainable_only=False)
tape.trainable_params = qml.math.get_trainable_indices(params)
parameters += [p for i, p in enumerate(params) if i in tape.trainable_params]
all_params += params
trainable += (np.array(list(tape.trainable_params)) + sum(lens)).tolist()
lens.append(len(params))
o_types = []
for m in tape.measurements:
if isinstance(m, SampleMP):
o_types.append(tf.int64)
elif isinstance(m, StateMP):
o_types.append(tf.complex128)
else:
o_types.append(tf.float64)
output_types.extend(o_types * num_shot_copies)
total_measurements = sum(len(tape.measurements) for tape in tapes)
if grad_on_execution:
output_types += [tf.float64] * len(trainable)
output_types += [tf.int32] * (total_measurements * num_shot_copies)
def _nest_params(all_params):
count = 0
params_unwrapped = []
for s in lens:
params_unwrapped.append(all_params[count : count + s])
count += s
return params_unwrapped
def _forward(*all_params):
params_unwrapped = _nest_params(all_params)
output_sizes = []
new_tapes = set_parameters_on_copy(tapes, params_unwrapped)
# Forward pass: execute the tapes
res, jacs = execute_fn(new_tapes, **gradient_kwargs)
# flatten the results
res = _flatten_nested_list(res)
for i, r in enumerate(res):
# convert output to TensorFlow tensors
res[i] = _to_tensors(r)
output_sizes.append(tf.size(res[i]))
# flatten the jacobians
if jacs:
jacs = _flatten_nested_list(jacs)
for i, jac in enumerate(jacs):
jacs[i] = tf.convert_to_tensor(jac)
else:
jacs = []
return res + jacs + output_sizes
@tf.custom_gradient
def _execute(*all_params): # pylint:disable=unused-argument
res = tf.numpy_function(func=_forward, inp=all_params, Tout=output_types)
output_sizes = res[-total_measurements * num_shot_copies :]
if grad_on_execution:
jacs = res[total_measurements * num_shot_copies : -total_measurements * num_shot_copies]
res = res[: total_measurements * num_shot_copies]
# reconstruct the nested structure of res
res = _res_restructured(res, tapes)
def grad_fn(*dy, **tfkwargs):
"""Returns the vector-Jacobian product with given
parameter values and output gradient dy"""
# whether the tapes contain multiple measurements
multi_measurements = [len(tape.measurements) > 1 for tape in tapes]
dy = list(dy[: total_measurements * num_shot_copies])
if grad_on_execution:
# Jacobians were computed on the forward pass (grad_on_execution=True)
# No additional quantum evaluations needed; simply compute the VJPs directly.
def _backward(*args):
for tape in tapes:
for m in tape.measurements:
if m.numeric_type == complex:
raise NotImplementedError(
f"Tensorflow autograph only supports real valued measurements. Got {m}"
)
dy = args[: total_measurements * num_shot_copies]
jacs = args[total_measurements * num_shot_copies : -len(tapes)]
dy = _res_restructured(dy, tapes)
jacs = _jac_restructured(jacs, tapes)
return _compute_vjp(dy, jacs, multi_measurements, has_partitioned_shots)
vjps = tf.numpy_function(
func=_backward,
inp=dy + list(jacs) + multi_measurements,
Tout=[tf.float64] * len(parameters),
)
else:
# Need to compute the Jacobians on the backward pass (accumulation="backward")
if isinstance(gradient_fn, qml.transforms.core.TransformDispatcher):
# Gradient function is a gradient transform.
# Generate and execute the required gradient tapes
if _n == max_diff:
len_all_params = len(all_params)
def _backward(*all_params):
dy = all_params[len_all_params:]
all_params = all_params[:len_all_params]
params_unwrapped = _nest_params(all_params)
dy = _res_restructured(dy, tapes)
new_tapes = set_parameters_on_copy(tapes, params_unwrapped)
vjp_tapes, processing_fn = qml.gradients.batch_vjp(
new_tapes,
dy,
gradient_fn,
reduction=lambda vjps, x: vjps.extend(qml.math.unstack(x)),
gradient_kwargs=gradient_kwargs,
)
return processing_fn(execute_fn(vjp_tapes)[0])
vjps = tf.py_function(
func=_backward,
inp=list(all_params) + dy,
Tout=[tf.float64] * len(parameters),
)
else:
dy = _res_restructured(dy, tapes)
vjp_tapes, processing_fn = qml.gradients.batch_vjp(
tapes,
dy,
gradient_fn,
reduction="append",
gradient_kwargs=gradient_kwargs,
)
# This is where the magic happens. Note that we call ``execute``.
# This recursion, coupled with the fact that the gradient transforms
# are differentiable, allows for arbitrary order differentiation.
vjps = processing_fn(
execute(
vjp_tapes,
device,
execute_fn,
gradient_fn,
gradient_kwargs,
_n=_n + 1,
max_diff=max_diff,
grad_on_execution=grad_on_execution,
),
nums=output_sizes,
)
vjps = tf.unstack(tf.concat(vjps, 0), num=len(parameters))
else:
# Gradient function is not a gradient transform
# (e.g., it might be a device method).
# Note that unlike the previous branch:
#
# - there is no recursion here
# - gradient_fn is not differentiable
#
# so we cannot support higher-order derivatives.
len_all_params = len(all_params)
def _backward(*all_params):
dy = all_params[len_all_params : -len(tapes)]
multi_measurements = all_params[-len(tapes) :]
all_params = all_params[:len_all_params]
params_unwrapped = _nest_params(all_params)
new_tapes = set_parameters_on_copy(tapes, params_unwrapped)
jac = gradient_fn(new_tapes, **gradient_kwargs)
vjps = _compute_vjp(dy, jac, multi_measurements, has_partitioned_shots)
return vjps
vjps = tf.numpy_function(
func=_backward,
inp=list(all_params) + dy + multi_measurements,
Tout=[tf.float64] * len(parameters),
)
if not isinstance(vjps, (list, tuple)):
vjps = [vjps]
vjps = iter(vjps)
vjps = [next(vjps) if x in trainable else None for x in range(len(all_params))]
variables = tfkwargs.get("variables")
return (vjps, variables) if variables is not None else vjps
return res, grad_fn
return _execute(*all_params)
_modules/pennylane/workflow/interfaces/tensorflow_autograph
Download Python script
Download Notebook
View on GitHub