def std_symbt_pred(arr: "SparseToDenseArrow", port_attr: PortAttributes):
# sparse_indices: value
# output_shape: value
# sparse_values: SymbolicTensor
a = ports_has(arr.in_ports()[0:2], 'value', port_attr)
b = port_has(arr.in_port(2), 'symbolic_tensor', port_attr)
# Hack, a FIXME to propagate should stop repropagation
c = not port_has(arr.out_port(0), 'symbolic_tensor', port_attr)
return a and b and c
def add_pred1(arr: "AddArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def div_pred1(arr: "DivArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr) and np.all(
port_attr[arr.in_port(1)]['value'])
def mul_pred3(arr: "MulArrow", port_attr: PortAttributes):
ports = [arr.in_ports()[1], arr.out_ports()[0]]
# TODO: think harder about the zeros case
return ports_has(ports, 'value', port_attr) and np.all(
port_attr[ports[0]]['value'])
def std_pred2(arr: "SparseToDenseArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def broadcast_pred(arr: Arrow, port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'shape', port_attr)
def neg_fwd_pred(arr: "NegArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def pow_pred2(arr: "PowArrow", port_attr: PortAttributes):
ports = [arr.in_ports()[0], arr.out_ports()[0]]
return ports_has(ports, 'value', port_attr)
def reshape_pred3(arr: "ReshapeArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports()[1:2], 'value', port_attr)
def reshape_pred2(arr: "ReshapeArrow", port_attr: PortAttributes):
return ports_has(arr.out_ports()[:1], 'shape', port_attr)
def snd_pred4(arr: "ScatterNdArrow", port_attr: PortAttributes):
return ports_has(arr.out_ports(), 'shape', port_attr)
def snd_pred2(arr: "ScatterNdArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def std_pred4(arr: "SparseToDenseArrow", port_attr: PortAttributes):
return ports_has(arr.out_ports(), 'shape', port_attr)
def std_pred3(arr: "SparseToDenseArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports()[0:1], 'shape', port_attr)
def div_pred2(arr: "DivArrow", port_attr: PortAttributes):
ports = [arr.in_ports()[0], arr.out_ports()[0]]
return ports_has(ports, 'value', port_attr) and np.all(
port_attr[arr.out_port(0)]['value'])
def fwd_pred(arr: "SelectArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def pow_pred1(arr: "PowArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def gathernd_shape_pred(arr: "GatherArrow", port_attr: PortAttributes):
# FIXME: Can we infer shaep from output or aoutput and one input?
return ports_has(arr.in_ports(), 'shape', port_attr)
def log_bwd_pred(arr: "LogArrow", port_attr: PortAttributes):
return ports_has(arr.out_ports(), 'value', port_attr)
def mul_pred1(arr: "MulArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def add_pred3(arr: "AddArrow", port_attr: PortAttributes):
ports = [arr.in_ports()[1], arr.out_ports()[0]]
return ports_has(ports, 'value', port_attr)
def mul_pred2(arr: "MulArrow", port_attr: PortAttributes):
ports = [arr.in_ports()[0], arr.out_ports()[0]]
return ports_has(ports, 'value', port_attr) and np.all(
port_attr[ports[0]]['value'])
def sub_pred1(arr: "SubArrow", port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'value', port_attr)
def constant_pred(arr, port_attr: PortAttributes):
return ports_has(arr.in_ports(), 'constant', port_attr)
