Refactoring interfaces in preparation for Jacobian Product work (#4415

)

* preparation cleaning to interface files

* move conversion to numpy inside execute_fn

* fixing some failures

* hopefully fixing all the tests

* fixing more tests

* fixing legacy mistake

* Update pennylane/devices/qubit/sampling.py

* Update pennylane/optimize/qnspsa.py

* Update tests/transforms/test_batch_transform.py

* Update pennylane/interfaces/torch.py

* changelog and black

* fix failures

* Update pennylane/interfaces/execution.py

Co-authored-by: Matthew Silverman <matthews@xanadu.ai>

* respond to feedback

* oops

* add recommendation to warning

* Update pennylane/interfaces/execution.py

* black

* Update pennylane/gradients/finite_difference.py

* Update pennylane/interfaces/execution.py

* black and more pragma: no-cover

---------

Co-authored-by: Matthew Silverman <matthews@xanadu.ai>
Co-authored-by: Romain Moyard <rmoyard@gmail.com>

doc/releases/changelog-dev.md

@@ -165,6 +165,13 @@ array([False, False])
 * Provide users access to the logging configuration file path and improve the logging configuration structure.
  [(#4377)](https://github.com/PennyLaneAI/pennylane/pull/4377)
 
+* Refactoring of `pennylane/interfaces`. The `execute_fn` passed to the machine learning framework boundaries 
+ is now responsible for converting parameters to numpy. The gradients module can now handle tensorflow parameters,
+ but gradient tapes now retain the original dtype instead of converting to float64. This may cause instability 
+ with finite diff and float32 parameters. The ml boundary functions are now uncoupled from their legacy
+ counterparts.
+ [(#4415)](https://github.com/PennyLaneAI/pennylane/pull/4415)
+
 * `qml.transforms.adjoint_metric_tensor` now uses the simulation tools in `pennylane.devices.qubit` instead of
  private methods of `pennylane.devices.DefaultQubit`.
  [(#4456)](https://github.com/PennyLaneAI/pennylane/pull/4456)
@@ -181,6 +188,10 @@ array([False, False])
 
 <h3>Breaking changes 💔</h3>
 
+* Gradient transforms no longer implicitly cast `float32` parameters to `float64`. Finite diff
+ with float32 parameters may no longer give accurate results.
+ [(#4415)](https://github.com/PennyLaneAI/pennylane/pull/4415)
+
 * Support for Python 3.8 is dropped.
  [(#4453)](https://github.com/PennyLaneAI/pennylane/pull/4453)
 

pennylane/gradients/finite_difference.py

@@ -18,6 +18,7 @@
 # pylint: disable=protected-access,too-many-arguments,too-many-branches,too-many-statements
 import functools
 from collections.abc import Sequence
+from warnings import warn
 
 import numpy as np
 from scipy.special import factorial
@@ -329,6 +330,13 @@ def finite_diff(
  f0=f0,
  validate_params=validate_params,
  )
+
+ if any(qml.math.get_dtype_name(p) == "float32" for p in tape.get_parameters()):
+ warn(
+ "Finite differences with float32 detected. Answers may be inaccurate. float64 is recommended.",
+ UserWarning,
+ )
+
  if argnum is None and not tape.trainable_params:
  return _no_trainable_grad(tape)
 

pennylane/gradients/general_shift_rules.py

@@ -396,8 +396,12 @@ def _copy_and_shift_params(tape, indices, shifts, multipliers, cast=False):
 
  # Shift copied parameter
  new_params = list(op.data)
- new_params[p_idx] = new_params[p_idx] * qml.math.convert_like(multiplier, new_params[p_idx])
- new_params[p_idx] = new_params[p_idx] + qml.math.convert_like(shift, new_params[p_idx])
+ multiplier = qml.math.convert_like(multiplier, new_params[p_idx])
+ multiplier = qml.math.cast_like(multiplier, new_params[p_idx])
+ shift = qml.math.convert_like(shift, new_params[p_idx])
+ shift = qml.math.cast_like(shift, new_params[p_idx])
+ new_params[p_idx] = new_params[p_idx] * multiplier
+ new_params[p_idx] = new_params[p_idx] + shift
  if cast:
  dtype = getattr(new_params[p_idx], "dtype", float)
  new_params[p_idx] = qml.math.cast(new_params[p_idx], dtype)

pennylane/interfaces/autograd.py

@@ -303,16 +303,6 @@ def execute(tapes, device, execute_fn, gradient_fn, gradient_kwargs, _n=1, max_d
  list[list[float]]: A nested list of tape results. Each element in
  the returned list corresponds in order to the provided tapes.
  """
- if not qml.active_return():
- return _execute_legacy(
- tapes,
- device,
- execute_fn,
- gradient_fn,
- gradient_kwargs,
- _n=_n,
- max_diff=max_diff,
- )
  # pylint: disable=unused-argument
  for tape in tapes:
  # set the trainable parameters
@@ -385,8 +375,7 @@ def _execute(
  "::L".join(str(i) for i in inspect.getouterframes(inspect.currentframe(), 2)[1][1:3]),
  )
 
- unwrapped_tapes = tuple(convert_to_numpy_parameters(t) for t in tapes)
- res, jacs = execute_fn(unwrapped_tapes, **gradient_kwargs)
+ res, jacs = execute_fn(tapes, **gradient_kwargs)
 
  return res, jacs
 
@@ -461,9 +450,7 @@ def partial_gradient_fn(tape):
  return gradient_fn(tape, **gradient_kwargs)
 
  g_tapes, fn = qml.transforms.map_batch_transform(partial_gradient_fn, tapes)
- unwrapped_tapes = tuple(convert_to_numpy_parameters(g_t) for g_t in g_tapes)
-
- res, _ = execute_fn(unwrapped_tapes, **gradient_kwargs)
+ res, _ = execute_fn(g_tapes, **gradient_kwargs)
 
  jacs = fn(res)
  cached_jac["jacobian"] = jacs
@@ -507,9 +494,8 @@ def grad_fn(dy):
 
  # Generate and execute the required gradient tapes
  if _n == max_diff:
- unwrapped_tapes = tuple(convert_to_numpy_parameters(t) for t in tapes)
  vjp_tapes, processing_fn = qml.gradients.batch_vjp(
- unwrapped_tapes,
+ tapes,
  dy,
  gradient_fn,
  reduction="append",
@@ -551,8 +537,7 @@ def grad_fn(dy):
  # - gradient_fn is not differentiable
  #
  # so we cannot support higher-order derivatives.
- unwrapped_tapes = tuple(convert_to_numpy_parameters(t) for t in tapes)
- jacs = gradient_fn(unwrapped_tapes, **gradient_kwargs)
+ jacs = gradient_fn(tapes, **gradient_kwargs)
 
  vjps = _compute_vjps_autograd(jacs, dy, multi_measurements, has_partitioned_shots)