def test_tf_conversion(self, qnode, tol):
"""Tests that the to_tf() function ignores QNodes that already
have the TF interface."""
converted_qnode = to_tf(qnode)
assert converted_qnode is qnode
x = tf.Variable(0.4)
with tf.GradientTape() as tape:
res = converted_qnode(x)
assert np.allclose(res, tf.cos(x), atol=tol, rtol=0)
grad = tape.gradient(res, x)
assert np.allclose(grad, -tf.sin(x), atol=tol, rtol=0)
def test_other_conversion(self, qnode, tol):
"""Tests that the to_tf() function correctly converts both torch and autograd qnodes
and QNodes with no interface."""
converted_qnode = to_tf(qnode)
assert converted_qnode is not qnode
assert converted_qnode._qnode is getattr(qnode, "_qnode", qnode)
x = tf.Variable(0.4)
with tf.GradientTape() as tape:
res = converted_qnode(x)
assert np.allclose(res, tf.cos(x), atol=tol, rtol=0)
grad = tape.gradient(res, x)
assert np.allclose(grad, -tf.sin(x), atol=tol, rtol=0)
def __init__(self,
qnode,
weight_shapes: dict,
output_dim: Iterable,
weight_specs: Optional[dict] = None,
*args,
**kwargs):
if not CORRECT_TF_VERSION:
raise ImportError(
"Quanv2D requires TensorFlow version 2 or above. The latest "
"version of TensorFlow can be installed using:\n"
"pip install tensorflow --upgrade\nAlternatively, visit "
"https://www.tensorflow.org/install for detailed instructions."
)
self.weight_shapes = {
weight: (tuple(size) if isinstance(size, Iterable) else
(size, ) if size > 1 else ())
for weight, size in weight_shapes.items()
}
if qml.tape_mode_active():
self._signature_validation_tape_mode(qnode, weight_shapes)
self.qnode = qnode
dtype = tf.float32 if tf.keras.backend.floatx(
) == tf.float32 else tf.float64
if self.qnode.diff_method != "backprop":
self.qnode.to_tf(dtype=dtype)
else:
self._signature_validation(qnode, weight_shapes)
self.qnode = to_tf(qnode, dtype=tf.keras.backend.floatx())
# Output dim must be an iterable with at least 2 dimensions (possibly a third for channels) TODO check this
if len(output_dim) != 3:
raise ValueError(
"Output must have three dimensions (the 2 image dimensions and one for the channels)"
)
else:
self.output_dim = output_dim
self.weight_specs = weight_specs if weight_specs is not None else {}
self.qnode_weights = {}
super().__init__(dynamic=True, *args, **kwargs)
def __init__(self,
qnode,
weight_shapes: dict,
output_dim,
weight_specs: Optional[dict] = None,
**kwargs):
if not CORRECT_TF_VERSION:
raise ImportError(
"KerasLayer requires TensorFlow version 2 or above. The latest "
"version of TensorFlow can be installed using:\n"
"pip install tensorflow --upgrade\nAlternatively, visit "
"https://www.tensorflow.org/install for detailed instructions."
)
self.weight_shapes = {
weight: (tuple(size) if isinstance(size, Iterable) else
(size, ) if size > 1 else ())
for weight, size in weight_shapes.items()
}
if qml.tape_mode_active():
self._signature_validation_tape_mode(qnode, weight_shapes)
self.qnode = qnode
dtype = tf.float32 if tf.keras.backend.floatx(
) == tf.float32 else tf.float64
if self.qnode.diff_method != "backprop":
self.qnode.to_tf(dtype=dtype)
else:
self._signature_validation(qnode, weight_shapes)
self.qnode = to_tf(qnode, dtype=tf.keras.backend.floatx())
# Allows output_dim to be specified as an int, e.g., 5, or as a length-1 tuple, e.g., (5,)
self.output_dim = output_dim[0] if isinstance(output_dim,
Iterable) else output_dim
self.weight_specs = weight_specs if weight_specs is not None else {}
self.qnode_weights = {}
super().__init__(dynamic=True, **kwargs)
def __init__(self,
qnode,
weight_shapes: dict,
output_dim,
weight_specs: Optional[dict] = None,
**kwargs):
if not CORRECT_TF_VERSION:
raise ImportError(
"KerasLayer requires TensorFlow version 2 or above. The latest "
"version of TensorFlow can be installed using:\n"
"pip install tensorflow --upgrade\nAlternatively, visit "
"https://www.tensorflow.org/install for detailed instructions."
)
self.sig = qnode.func.sig
if self.input_arg not in self.sig:
raise TypeError(
"QNode must include an argument with name {} for inputting data"
.format(self.input_arg))
if self.input_arg in set(weight_shapes.keys()):
raise ValueError(
"{} argument should not have its dimension specified in "
"weight_shapes".format(self.input_arg))
if set(weight_shapes.keys()) | {self.input_arg} != set(
self.sig.keys()):
raise ValueError(
"Must specify a shape for every non-input parameter in the QNode"
)
if qnode.func.var_pos:
raise TypeError(
"Cannot have a variable number of positional arguments")
if qnode.func.var_keyword:
raise TypeError(
"Cannot have a variable number of keyword arguments")
self.qnode = to_tf(qnode, dtype=tf.keras.backend.floatx())
self.weight_shapes = {
weight: (tuple(size) if isinstance(size, Iterable) else
(size, ) if size > 1 else ())
for weight, size in weight_shapes.items()
}
# Allows output_dim to be specified as an int, e.g., 5, or as a length-1 tuple, e.g., (5,)
self.output_dim = output_dim[0] if isinstance(output_dim,
Iterable) else output_dim
defaults = {
name
for name, sig in self.sig.items()
if sig.par.default != inspect.Parameter.empty
}
self.input_is_default = self.input_arg in defaults
if defaults - {self.input_arg} != set():
raise TypeError(
"Only the argument {} is permitted to have a default".format(
self.input_arg))
self.weight_specs = weight_specs if weight_specs is not None else {}
self.qnode_weights = {}
super().__init__(dynamic=True, **kwargs)