def optimize_pair(self, op1: ElementwiseAdd, op2: ScalarMul):
c1, v1 = _get_constant_and_variable(op1, "x0", "x1")
if c1 is None:
return False
y2 = op2.outputs["y"]
op2.remove_all()
op1.remove_all()
y = (v1 * op2.value) + (c1 * op2.value) # type: Variable
y.replace(y2)
return True
def _optimize_ElementwiseAdd_ScalarMul(op1: ElementwiseAdd,
c1: ConstantVariable, v1: Variable,
op2: Operator):
if not isinstance(op2, ScalarMul):
return False
y2 = op2.outputs["y"]
op2.remove_all()
op1.remove_all()
y = (v1 * op2.value) + (c1 * op2.value)
y.replace(y2)
return True
def optimize_pair(self, op1: ElementwiseAdd, op2: ScalarAdd):
if isinstance(op1.inputs["x0"], ConstantVariable):
c1 = op1.inputs["x0"], v1 = op1.inputs["x1"]
elif isinstance(op1.inputs["x1"], ConstantVariable):
c1 = op1.inputs["x1"], v1 = op1.inputs["x0"]
else:
return False
y2 = op2.outputs["y"]
op2.remove_all()
op1.remove_all()
y = v1 + (c1 + op2.value) # type: Variable
y.replace(y2)
return True
def optimize_pair(self, op1: ElementwiseAdd, op2: ElementwiseDiv):
c1, v1 = _get_constant_and_variable(op1, "x0", "x1")
if c1 is None:
return False
c2, v2 = _get_constant_and_variable(op2, "x0", "x1")
if c2 is None:
return False
y2 = op2.outputs["y"]
op2.remove_all()
op1.remove_all()
y = (v1 / c2) + (c1 / c2) # type: Variable
y.replace(y2, with_assert=False)
return True
def optimize_pair(self, op1: ScalarAdd, op2: ElementwiseAdd):
x0 = op1.inputs["x0"]
y1 = op1.outputs["y"]
if y1 == op2.inputs["x0"]:
w = op2.inputs["x1"]
else:
w = op2.inputs["x0"]
y2 = op2.outputs["y"]
op2.remove_all()
op1.remove_all()
y = (x0 + w) + op1.value # type: Variable
y.replace(y2)
return True
def optimize_pair(self, op1: ElementwiseAdd, op2: ElementwiseMul):
c1, v1 = _get_constant_and_variable(op1, "x0", "x1")
if c1 is None:
return False
c2, v2 = _get_constant_and_variable(op2, "x0", "x1")
if c2 is None:
return False
y2 = op2.outputs["y"]
op2.remove_all()
op1.remove_all()
y = (v1 * c2) + (c1 * c2)
y.replace(y2)
return True
def optimize_pair(self, graph: Graph, op1: ElementwiseAdd, op2: ElementwiseDiv):
c1, v1 = _get_constant_and_variable(op1, "x0", "x1")
if c1 is None:
return False
c2, v2 = _get_constant_and_variable(op2, "x0", "x1")
if c2 is None:
return False
y2 = op2.outputs["y"]
op2.remove_all()
op1.remove_all()
y = (v1 / c2) + (c1 / c2)
OptimizeRule.replace_variable(graph, y2, y.change_order(y2.order))
return True
def _convert_average(converter: KerasConverter, k_op: "keras.layers.Average"):
xs = [converter.get_variable(tensor) for tensor in converter.get_input_tensor(k_op)]
for x in xs[1:]:
xs[0].order.unify(x.order)
y = ElementwiseAdd(None)(*xs)[0] / len(xs)
converter.set_variable(converter.get_output_tensor(k_op)[0], y)
def _convert_add(converter: KerasConverter, k_op: "keras.layers.Add"):
xs = [
converter.get_variable(tensor)
for tensor in converter.get_input_tensor(k_op)
]
y, = ElementwiseAdd(None)(*xs)
converter.set_variable(converter.get_output_tensor(k_op)[0], y)
def _convert_average(converter: KerasConverter, k_op: "keras.layers.Average"):
xs = [
converter.get_variable(tensor)
for tensor in converter.get_input_tensor(k_op)
]
# FIXME: More effective implementation
y = ElementwiseAdd(None)(*xs)[0] / len(xs)
converter.set_variable(converter.get_output_tensor(k_op)[0], y)
def _optimize_ElementwiseAdd_ElementwiseMul(op1: ElementwiseAdd,
c1: ConstantVariable, v1: Variable,
op2: Operator):
if not isinstance(op2, ElementwiseMul):
return False
x0 = op2.inputs["x0"]
x1 = op2.inputs["x1"]
y2 = op2.outputs["y"]
if isinstance(x0, ConstantVariable):
c2 = x0
elif isinstance(x1, ConstantVariable):
c2 = x1
else:
return False
op2.remove_all()
op1.remove_all()
y = (v1 * c2) + (c1 * c2)
y.replace(y2)
return True
def optimize(self, graph: Graph) -> Tuple[Graph, bool]:
flag_changed = False
for op1 in traverse.filter_nodes(traverse.listup_operators(graph),
Concat): # type: Concat
if len(op1.inputs) != 2:
continue
x0 = op1.inputs["x0"]
x1 = op1.inputs["x1"]
y = op1.outputs["y"]
op2 = x0.output_from
op3 = x1.output_from
if isinstance(op2, ElementwiseAdd) and isinstance(
op3, ElementwiseAdd) and len(x0.input_to) == 1 and len(
x1.input_to) == 1:
"""
before)
v1 -+
+-[op2: ElementwiseAdd]-> x0 -+
c1 -+ |
+-[op1: Concat]-> y
v2 -+ |
+-[op3: ElementwiseAdd]-> x1 -+
c2 -+
after)
v1 -+
+-[Concat]-> x6 -+
v2 -+ |
+-[ElementwiseAdd]-> y
|
c3 -+
"""
x2 = op2.inputs["x0"]
x3 = op2.inputs["x1"]
x4 = op3.inputs["x0"]
x5 = op3.inputs["x1"]
if isinstance(x2, ConstantVariable):
c1 = x2
v1 = x3
elif isinstance(x3, ConstantVariable):
c1 = x3
v1 = x2
else:
continue
if isinstance(x4, ConstantVariable):
c2 = x4
v2 = x5
elif isinstance(x5, ConstantVariable):
c2 = x5
v2 = x4
else:
continue
if not (c1.order == c2.order == Order([op1.axis])):
continue
op1.remove_all()
op2.remove_all()
op3.remove_all()
c3 = ConstantVariable(np.hstack([c1.data, c2.data]), c1.order)
x6, = Concat(None, axis=op1.axis)(v1, v2)
y_dummy, = ElementwiseAdd(None)(x6, c3)
y_dummy.replace(y)
flag_changed = True
return graph, flag_changed