def extract(cls, node):
data_type = tf_dtype_extractor(node.pb.attr["T"].type)
AttributedPower.update_node_stat(node, {
'power': data_type(2),
'data_type': data_type
})
return cls.enabled
def extract(cls, node: Node):
scale = onnx_attr(node,
'scale',
'f',
default=np.array(1.0),
dst_type=lambda x: np.array(x))
AttributedPower.update_node_stat(node, {'scale': scale})
return cls.enabled
def extract(cls, node: Node):
pb = node.pb
assert pb, 'Protobuf layer can not be empty'
param = pb.power_param
attrs = {
'output_spatial_shape': None,
'power': param.power,
'scale': param.scale,
'shift': param.shift,
}
AttributedPower.update_node_stat(node, attrs)
return cls.enabled
def replace_pattern(graph: Graph, match: [str, Node]):
op = match['op']
op_type = op.type
if op.has_and_set('stop_value_propagation'):
return
const_port, tensor_port = get_value_in_port(op), get_tensor_in_port(op)
if const_port is None or tensor_port is None:
return
value = const_port.data.get_value()
assert value is not None
if value.size != 1:
return
value = value.item(0)
assert op_type in EltwisesWithScalarInputToPower.eltw_types
if op_type == 'Add':
delete_node = value == 0
AttributedPower.update_node_stat(op, {'shift': value})
elif op_type == 'Multiply':
delete_node = value == 1
AttributedPower.update_node_stat(op, {'scale': value})
elif op_type == 'Power':
delete_node = value == 1
AttributedPower.update_node_stat(op, {'power': value})
op.type_infer = AttributedPower.type_infer
const_port.disconnect()
if tensor_port.idx != 0:
tensor_port.get_connection().set_destination(op.in_port(0))
op.delete_input_port(1)
def extract(cls, node):
AttributedPower.update_node_stat(node, {'scale': 0})
return cls.enabled
def extract(cls, node):
AttributedPower.update_node_stat(node, {'power': -0.5})
return cls.enabled
def extract(cls, node: Node):
attrs = {
'power': node.module.exponent,
}
AttributedPower.update_node_stat(node, attrs)
return cls.enabled