def fuse_conv_and_bias_callback(self, op, label_map_op_list):
"""
Callback function that handles fusion for Conv + bias pattern
"""
for (label_map, op) in label_map_op_list:
map_roles = label_map[self.map_roles_label]
conv_op = self.op_arg(map_roles, 0)
bias = label_map[self.conv_bias_label]
new_op_map = {}
if isinstance(conv_op, ConvolutionOp):
conv_new_op = ConvolutionOp(conv_op.conv_params, self.op_arg(conv_op, 0),
self.op_arg(conv_op, 1), bias, axes=conv_op.axes)
new_op_map[conv_op] = conv_new_op
# Create ops that are downstream to convolution but still upstream of the add 'op'
prev_op = self.op_arg(op, 0)
upstream_ops = []
while prev_op != conv_op:
upstream_ops.append(prev_op)
prev_op = self.op_arg(prev_op, 0)
for old_op in reversed(upstream_ops):
new_arg = new_op_map[self.op_arg(old_op, 0)]
if isinstance(old_op, MapRolesOp):
new_op_map[old_op] = MapRolesOp(new_arg, old_op.axes_map)
elif isinstance(old_op, TensorSliceOp):
new_op_map[old_op] = TensorSliceOp(new_arg, old_op.slices, old_op.axes)
elif isinstance(old_op, ReorderAxes):
new_op_map[old_op] = ReorderAxes(new_arg, old_op.axes)
self.replace_op(op, new_op_map[self.op_arg(op, 0)])
self.replace_op(conv_op, conv_new_op)
def visit(self, op):
x, y = op.args
x_reduction_axes = op.x_reduction_axes
y_reduction_axes = op.y_reduction_axes
out_axes = op.axes
if len(x_reduction_axes) == 0:
d = make_axis(1)
x_reduction_axes = make_axes((d, ))
y_reduction_axes = x_reduction_axes
x = broadcast(x, x.axes + x_reduction_axes)
y = broadcast(y, y_reduction_axes + y.axes)
if x.is_scalar:
temp = x
x = y
y = temp
if y.is_scalar:
if x.is_scalar:
out = x.scalar_op * y.scalar_op
if len(x_reduction_axes) > 0:
out = out * x_reduction_axes.size
out = broadcast(out, op.axes)
else:
out = Sum(x, x_reduction_axes) * y.scalar_op
out = broadcast(out, op.axes)
else:
x_rem_axes = x.axes - x_reduction_axes
x = axes_with_order(x, x_rem_axes + x_reduction_axes)
y_rem_axes = y.axes - y_reduction_axes
y = axes_with_order(y, y_reduction_axes + y_rem_axes)
x = flatten_at(x, len(x.axes) - len(x_reduction_axes))
y = flatten_at(y, len(y_reduction_axes))
if len(out_axes) == 0:
out = DotOneDimensional(x, y, axes=())
elif len(x.axes) == 1:
y = Transpose(y)
out = DotTwoByOne(y, x, axes=y.axes[0])
elif len(y.axes) == 1:
out = DotTwoByOne(x, y, axes=x.axes[0])
else:
out = DotTwoDimensional(
x,
y,
axes=([op.x_out_axes.flatten(),
op.y_out_axes.flatten()]))
out = unflatten(out)
out = ReorderAxes(out, out_axes)
self.replace_op(op, out)
def visit(self, op):
x, y = op.args
reduction_axes = op.reduction_axes
out_axes = op.axes
if len(reduction_axes) == 0:
# TODO: this is a weird case, should we really support it?
d = make_axis(1)
reduction_axes = make_axes((d, ))
x = broadcast(x, x.axes + reduction_axes)
y = broadcast(y, reduction_axes + y.axes)
if x.is_scalar:
x, y = y, x
if y.is_scalar:
if x.is_scalar:
out = x.scalar_op * y.scalar_op
if len(reduction_axes) > 0:
out = out * reduction_axes.size
out = broadcast(out, op.axes)
else:
out = Sum(x, reduction_axes) * y.scalar_op
out = broadcast(out, op.axes)
else:
# move reduction_axes to end
x = axes_with_order(x, (x.axes - reduction_axes) + reduction_axes)
# move reduction axes to front
y = axes_with_order(y, reduction_axes + (y.axes - reduction_axes))
# flatten non-reduction axes together and reduction axes together
x = flatten_at(x, len(x.axes) - len(reduction_axes))
# flatten non-reduction axes together and reduction axes together
y = flatten_at(y, len(reduction_axes))
if len(out_axes) == 0:
out = DotLowDimension(x, y, axes=())
elif len(x.axes) == 1:
y = Transpose(y)
out = DotLowDimension(y, x, axes=y.axes[0])
elif len(y.axes) == 1:
out = DotLowDimension(x, y, axes=x.axes[0])
else:
out = DotLowDimension(x,
y,
axes=([
op.x_out_axes.flatten(True),
op.y_out_axes.flatten(True)
]))
out = unflatten(out)
out = ReorderAxes(out, out_axes)
self.replace_op(op, out)