def replace_sub_graph(self, graph: Graph, match: [dict, SubgraphMatch]):
node = match['reduce']
connected_in_ports = [
port for port in node.in_ports().values()
if not port.disconnected()
]
if len(connected_in_ports) == 1:
node_name = node.soft_get('name', node.id)
# if the 'axis' is None then we still add a second input to the layer with a 1D array with 1 element equal
# to None. The infer function handles this case because the input shape is known at this stage only
if node.has_valid('axis'):
const = Const(graph, {
'name': node_name + '/axis',
'value': node.axis
}).create_node()
node.add_input_port(1, skip_if_exist=True)
const.out_port(0).connect(node.in_port(1))
del graph.node[node.id]['axis']
else:
# The default (if there is no 'axis') is to reduce over all the dimensions of the input tensor.
axes = create_op_with_const_inputs(
graph, Range, {
0: int64_array(0),
2: int64_array(1)
}, dict(name=node_name + '/axes'))
end_of_range = Rank(graph, dict(name=node_name +
'/range_end')).create_node()
node.in_port(0).get_connection().get_source().connect(
end_of_range.in_port(0))
end_of_range.out_port(0).connect(axes.in_port(1))
node.add_input_port(1, skip_if_exist=True)
axes.out_port(0).connect(node.in_port(1))
def convert_fft_to_dft(self, graph: Graph, mx_fft: Node):
mx_fft_name = mx_fft.soft_get('name', mx_fft.id)
unsqueeze_node = create_op_with_const_inputs(
graph, Unsqueeze, {1: int64_array([-1])},
{'name': mx_fft_name + '/Unsqueeze'})
rank_node = Rank(graph, {'name': mx_fft_name + '/Rank'}).create_node()
mx_fft_connection = mx_fft.in_port(0).get_connection()
mx_fft_connection.set_destination(unsqueeze_node.in_port(0))
mx_fft_connection.get_source().connect(rank_node.in_port(0))
add_node = create_op_with_const_inputs(graph, Add, {1: int64_array(1)},
{'name': mx_fft_name + '/Add'},
rank_node)
broadcast_node1 = create_op_with_const_inputs(
graph, Broadcast, {0: int64_array(0)},
{'name': mx_fft_name + '/Pad_broadcast'})
add_node.out_port(0).connect(broadcast_node1.in_port(1))
scatter_node = create_op_with_const_inputs(
graph, ScatterUpdate, {
2: int64_array(1),
3: int64_array(0)
}, {'name': mx_fft_name + '/ScatterUpdate'})
broadcast_node1.out_port(0).connect(scatter_node.in_port(0))
rank_node.out_port(0).connect(scatter_node.in_port(1))
pad_node = Pad(graph, {
'name': mx_fft_name + '/Pad',
'mode': 'constant'
}).create_node([unsqueeze_node, broadcast_node1, scatter_node])
dft_node = create_op_with_const_inputs(
graph, DFT, {1: int64_array([-1])}, {
'name': mx_fft_name + '/DFT',
'in_ports_count': 2
}, pad_node)
sub_node = create_op_with_const_inputs(graph, Sub, {1: int64_array(1)},
{'name': mx_fft_name + '/Sub'})
rank_node.out_port(0).connect(sub_node.in_port(0))
broadcast_node2 = create_op_with_const_inputs(
graph, Broadcast, {0: int64_array(0)},
{'name': mx_fft_name + '/Reshape_broadcast'})
sub_node.out_port(0).connect(broadcast_node2.in_port(1))
concat_node = create_op_with_const_inputs(
graph, Concat, {1: int64_array([-1, 2])}, {
'name': mx_fft_name + '/New_shape',
'in_ports_count': 2,
'axis': 0
}, broadcast_node2)
reshape_node = Reshape(graph, {}).create_node([dft_node, concat_node])
mx_fft.out_port(0).get_connection().set_source(
reshape_node.out_port(0))
rename_nodes([(mx_fft, mx_fft_name + '/to_be_removed'),
(reshape_node, mx_fft_name)])
def find_and_replace_pattern(self, graph: Graph):
for node in graph.get_op_nodes(op='MVNCaffe'):
node_name = node.soft_get('name', node.id)
start_axis = 2
if node['across_channels'] == 1:
start_axis = 1
rank = Rank(graph, {'name': node_name + '/Rank'}).create_node()
# create range of axes based on `start_axis` and rank of input
rng = create_op_with_const_inputs(graph, Range, {
0: int64_array(start_axis),
2: int64_array(1)
}, {
'name': node_name + '/Range',
'output_type': np.int64
})
rng.in_port(1).connect(rank.out_port(0))
new_mvn = MVN(
graph, {
'eps': node.soft_get('eps', 1e-9),
'eps_mode': 'inside_sqrt',
'normalize_variance': node.soft_get(
'normalize_variance', 1)
}).create_node([node.in_port(0).get_source().node, rng])
new_mvn.in_port(0).get_connection().add_destination(
rank.in_port(0))
node.out_port(0).get_connection().set_source(new_mvn.out_port(0))
rename_nodes([(node, node_name + '/tbd'), (new_mvn, node_name)])
graph.remove_node(node.id)
def convert_ifft_to_dft(self, graph: Graph, mx_fft: Node):
mx_fft_name = mx_fft.soft_get('name', mx_fft.id)
rank_node = Rank(graph, {'name': mx_fft_name + '/rank'}).create_node()
sub_node = create_op_with_const_inputs(graph, Sub, {1: int64_array(1)},
{'name': mx_fft_name + '/Sub'})
rank_node.out_port(0).connect(sub_node.in_port(0))
broadcast_node0 = create_op_with_const_inputs(
graph, Broadcast, {0: int64_array(0)},
{'name': mx_fft_name + '/broadcast'})
sub_node.out_port(0).connect(broadcast_node0.in_port(1))
concat_node = create_op_with_const_inputs(
graph, Concat, {1: int64_array([-1, 2])}, {
'name': mx_fft_name + '/new_shape',
'in_ports_count': 2,
'axis': 0
}, broadcast_node0)
reshape_node = Reshape(graph, {
'name': mx_fft_name + '/reshape'
}).create_node()
concat_node.out_port(0).connect(reshape_node.in_port(1))
mx_fft_connection = mx_fft.in_port(0).get_connection()
mx_fft_connection.set_destination(reshape_node.in_port(0))
mx_fft_connection.get_source().connect(rank_node.in_port(0))
dft_node = create_op_with_const_inputs(
graph, IDFT, {1: int64_array([-1])}, {
'name': mx_fft_name + '/idft',
'in_ports_count': 2
}, reshape_node)
split_node = create_op_with_const_inputs(
graph, Split, {1: int64_array(-1)}, {
'name': mx_fft_name + '/split',
'num_splits': 2
}, dft_node)
squeeze_node = create_op_with_const_inputs(graph, Squeeze,
{1: int64_array([-1])}, {},
split_node)
mx_fft.out_port(0).get_connection().set_source(
squeeze_node.out_port(0))
rename_nodes([(mx_fft, mx_fft_name + '/to_be_removed'),
(squeeze_node, mx_fft_name)])
def find_and_replace_pattern(self, graph: Graph):
global_poolings = graph.get_op_nodes(type='Pooling', global_pool=True)
if len(global_poolings) == 0:
return
layout = graph.graph['layout']
assert layout != 'NHWC', 'Global pooling transformation depends on layout (NHWC not enabled)'
for pooling in global_poolings:
name = pooling.soft_get('name', pooling.id)
assert pooling.has_valid(
'pool_method'
), 'Global Pooling {} has no `pool_method` attribute'.format(name)
method = pooling['pool_method']
assert method in self.pool_method_to_reduce_type, \
'Unexpected Global Pooling method `{}` for node `{}`'.format(method, name)
reduce_op_class = self.pool_method_to_reduce_type[method]
reduce = reduce_op_class(graph, {
'name': name + '/reduce',
'keep_dims': True
}).create_node()
pooling.out_port(0).get_connection().set_source(reduce.out_port(0))
src = pooling.in_port(0).get_connection().get_source()
reduce.in_port(0).get_connection().set_source(src)
start = Const(graph, {'value': int64_array(2)}).create_node()
end = Rank(graph, {'name': name + '/input_rank'}).create_node()
delta = Const(graph, {'value': int64_array(1)}).create_node()
axis = Range(graph, {
'name': name + '/global_pooling_reduce_axis'
}).create_node()
axis.in_port(0).connect(start.out_port(0))
src.connect(end.in_port(0))
axis.in_port(1).connect(end.out_port(0))
axis.in_port(2).connect(delta.out_port(0))
axis.out_port(0).connect(reduce.in_port(1))
log.debug('Global {} pooling was converted to reduce: `{}`'.format(
method, name))
def replace_op(self, graph: Graph, node: Node):
name = node.soft_get('name', node.id)
# create range of axes for MVN based on `start_axis` and rank of input
rank = Rank(graph, {'name': name + '/Rank'}).create_node()
rng = create_op_with_const_inputs(graph, Range, {
0: int64_array(2),
2: int64_array(1)
}, {
'name': name + '/Range',
'output_type': np.int64
})
mvn = MVN(
graph, {
'eps': node.epsilon,
'eps_mode': 'inside_sqrt',
'normalize_variance': 1,
'name': name + '/Ins_Norm/MVN_',
}).create_node()
node.in_port(0).get_connection().set_destination(mvn.in_port(0))
rng.out_port(0).connect(mvn.in_port(1))
mul = Mul(graph, {
'axis': 1,
'name': name + '/Ins_Norm/mul_'
}).create_node()
mvn.out_port(0).connect(mul.in_port(0))
node.in_port(1).get_connection().set_destination(mul.in_port(1))
add = Add(graph, {
'axis': 1,
'name': name + '/Ins_Norm/add_'
}).create_node()
mul.out_port(0).connect(add.in_port(0))
node.in_port(2).get_connection().set_destination(add.in_port(1))
mvn.in_port(0).get_connection().add_destination(rank.in_port(0))
rng.in_port(1).connect(rank.out_port(0))
rename_nodes([(node, name + '/TBD'), (add, name)])
return [add.id]
def replace_resize(graph: Graph, resize: Node):
log.debug("Converting of ONNX Resize-10 to Interpolate-4 "
"is triggered for node {}.".format(
resize.soft_get('name', resize.id)))
resize_name = resize.soft_get('name', resize.id)
rank_node = Rank(graph, {'name': resize_name + '/max_axes'}).create_node()
range_node = create_op_with_const_inputs(graph, Range, {
0: int64_array(2),
2: int64_array(1)
}, {'name': resize_name + '/axes'})
sizes_ss = create_op_with_const_inputs(graph, StridedSlice, {
1: int64_array([2]),
2: int64_array([0]),
3: int64_array([1])
}, {
'name': resize_name + '/sizes_ss',
'begin_mask': int64_array([1]),
'end_mask': int64_array([0]),
'new_axis_mask': int64_array([0]),
'shrink_axis_mask': int64_array([0]),
'ellipsis_mask': int64_array([0])
})
scales_ss = create_op_with_const_inputs(
graph, StridedSlice, {
1: int64_array([2]),
2: int64_array([0]),
3: int64_array([1])
}, {
'name': resize_name + '/scales_ss',
'begin_mask': int64_array([1]),
'end_mask': int64_array([0]),
'new_axis_mask': int64_array([0]),
'shrink_axis_mask': int64_array([0]),
'ellipsis_mask': int64_array([0])
})
rank_node.out_port(0).connect(range_node.in_port(1))
interpolate_node = Interpolate(
graph, {
'version': 'opset4',
'mode': 'linear_onnx' if resize.mode == 'linear' else 'nearest',
'coordinate_transformation_mode': 'asymmetric',
'cube_coeff': -0.75,
'nearest_mode': 'simple',
'pads_begin': int64_array([0]),
'pads_end': int64_array([0]),
'antialias': 0,
'shape_calculation_mode': 'scales',
'in_ports_count': 4
}).create_node()
range_node.out_port(0).connect(interpolate_node.in_port(3))
shape_of = Shape(graph, {'name': resize_name + '/ShapeOf'}).create_node()
# When we calculate 'sizes' input as floor(input_shape * scales), we can get incorrect 'sizes' if, e.g.,
# scales = [1.0, 1.0, 1.33333, 2.0], input_shape = [1, 3, 30, 200], because
# input_shape * scales = [1, 3, 39.9999, 400], and floor(input_shape * scales)[2] == 39, not 40.
# Maybe we need to calculate 'sizes' input as floor(input_shape * scales + eps), where eps is some small
# floating point number, e.g. 1.0e-5. But, in this case, if scales = [1.0, 1.0, 1.333333, 2.0],
# input_shape = [1, 3, 30, 200], floor(input_shape * scales + eps) = 39, not 40, because
# input_shape[2] * scales[2] + 1.0e-5 = 39.99991.
# Hence, we need to calculate 'sizes' as floor(input_shape * (scales + eps)).
add_node = create_op_with_const_inputs(graph, Add,
{1: float_array([1.0e-5])},
{'name': resize_name + '/Add'})
dst_dtype = np.float32 # even if data_type=FP16 use float32 for shape values
cast_shape_to_float = Cast(graph, {'dst_type': dst_dtype}).create_node()
shape_of.out_port(0).connect(cast_shape_to_float.in_port(0))
mul_node = Mul(graph, {
'name': resize_name + '/Mul'
}).create_node([cast_shape_to_float, add_node])
floor_node = Floor(graph, {
'name': resize_name + '/Floor'
}).create_node([mul_node])
cast_mul_result_to_int = Cast(graph, {
'dst_type': np.int64
}).create_node([floor_node])
cast_mul_result_to_int.out_port(0).connect(sizes_ss.in_port(0))
sizes_ss.out_port(0).connect(interpolate_node.in_port(1))
scales_ss.out_port(0).connect(interpolate_node.in_port(2))
connection_of_resize_input = resize.in_port(0).get_connection()
connection_of_resize_input.set_destination(interpolate_node.in_port(0))
connection_of_scales = resize.in_port(1).get_connection()
connection_of_scales.set_destination(scales_ss.in_port(0))
connection_of_resize_input.get_source().connect(shape_of.in_port(0))
connection_of_resize_input.get_source().connect(rank_node.in_port(0))
connection_of_scales.get_source().connect(add_node.in_port(0))
rename_nodes([(resize, resize_name + '/delete'),
(interpolate_node, resize_name)])
resize.out_port(0).get_connection().set_source(
interpolate_node.out_port(0))
def mxrepeat_decomposition(node: Node):
graph = node.graph
name = node.soft_get('name', node.id)
rename_node(node, name + '/to_be_removed')
# Unqueeze
input_rank = Rank(graph, {'name': name + '/Rank'}).create_node()
node.in_port(0).get_source().connect(input_rank.in_port(0))
axis = get_canonical_axis_index_node(input_rank, node.axis)
unsqueeze_axis = create_op_node_with_second_input(
graph,
Add,
int64_array([1]), {'name': name + '/Unsqueeze/Axis'},
input_node=axis)
unsqueeze = Unsqueeze(graph, {
'name': name + '/Unsqueeze'
}).create_node()
unsqueeze.in_port(1).connect(unsqueeze_axis.out_port(0))
# Tile (1, 1, ..., repeats, ..., 1)
# we generate tile array according to the following table:
# parts: | first | repeats | second |
# i: | 0, 1, ..., axis,| axis + 1,| ..., rank+1 |
# tile_array: | 1, 1, ..., 1 ,| repeats ,| ..., 1 |
one = Const(graph, {
'name': name + '/Broadcast/One',
'value': int64_array([1])
}).create_node()
first_ones = Broadcast(graph, {
'name': name + '/Broadcast/Ones_first_part'
}).create_node()
first_ones.in_port(0).connect(one.out_port(0))
first_ones.in_port(1).connect(unsqueeze_axis.out_port(0))
repeats = Const(graph, {
'name': name + '/repeats',
'value': int64_array([node.repeats])
}).create_node()
second_ones = Broadcast(graph, {
'name': name + '/Broadcast/Ones_second_part'
}).create_node()
second_part_broadcast_shape = Sub(
graph, {
'name': name + '/Broadcast/Shape/second_part'
}).create_node()
second_part_broadcast_shape.in_port(0).connect(input_rank.out_port(0))
second_part_broadcast_shape.in_port(1).connect(
unsqueeze_axis.out_port(0))
second_ones.in_port(0).connect(one.out_port(0))
second_ones.in_port(1).connect(second_part_broadcast_shape.out_port(0))
tile_repeats = new_shape_node_from_shape_nodes(
[first_ones, repeats, second_ones])
tile = Tile(graph, {'name': name + '/Tile'}).create_node()
tile.in_port(1).connect(tile_repeats.out_port(0))
# Reshape (input_shape[:axis], input_shape[axis] * repeats, input_shape[axis+1:])
# we generate reshape dim array according to the following table:
# parts: | first | rep | second |
# i: | 0, 1, ... ,| axis, | ..., rank |
# dim_array: | inp_sh[i] ,| input_shape[axis] * repeats ,| inp_sh[i] |
input_shape = Shape(graph, {'name': name + '/Shape'}).create_node()
node.in_port(0).get_source().connect(input_shape.in_port(0))
first_input_shape_part = get_shape_values_by_range_idxs(
input_shape,
input_rank,
begin=0,
end=node.axis,
include_begin=True,
include_end=False)
original_axis_dim = create_op_with_const_inputs(
graph,
Gather, {2: int64_array(0)}, {'name': name + '/OriginalDim'},
input_node=input_shape)
original_axis_dim.in_port(1).connect(axis.out_port(0))
repeated_dimention = Mul(graph, {
'name': name + '/RepeatedDim'
}).create_node()
repeated_dimention.in_port(0).connect(original_axis_dim.out_port(0))
repeated_dimention.in_port(1).connect(repeats.out_port(0))
second_input_shape_part = get_shape_values_by_range_idxs(
input_shape,
input_rank,
begin=node.axis,
end=-1,
include_begin=False,
include_end=True)
output_shape = new_shape_node_from_shape_nodes([
first_input_shape_part, repeated_dimention, second_input_shape_part
])
reshape = Reshape(graph, {'name': name}).create_node()
rename_node(reshape, name)
reshape.in_port(1).connect(output_shape.out_port(0))
# Final connections
node.in_port(0).get_connection().set_destination(unsqueeze.in_port(0))
tile.in_port(0).connect(unsqueeze.out_port(0))
reshape.in_port(0).connect(tile.out_port(0))
node.out_port(0).get_connection().set_source(reshape.out_port(0))