def floor_div_replacement(floor_div: Node):
graph = floor_div.graph
name = floor_div.soft_get('name', floor_div.id)
div = Div(graph, {'name': name + '/Div'}).create_node()
floor = Floor(graph, {'name': name}).create_node()
div.out_port(0).connect(floor.in_port(0))
div.in_port(0).connect(floor_div.in_port(0).get_source())
div.in_port(1).connect(floor_div.in_port(1).get_source())
floor_div.out_port(0).get_connection().set_source(floor.out_port(0))
graph.remove_node(floor_div.id)
rename_node(floor, name)
def replace_sub_graph(self, graph: Graph, match: dict):
div_sqrt = match['op']
div_sqrt_name = div_sqrt.soft_get('name', div_sqrt.id)
shape_node = Shape(graph,
dict(name=div_sqrt_name + '/Shape')).create_node()
data_out_port = div_sqrt.in_port(0).get_source()
shape_node.in_port(0).connect(data_out_port)
shape_values_node = node_to_get_shape_value_of_indices(
shape_node=shape_node, indices=[-1])
pow_node = AttributedPower(
graph, dict(name=div_sqrt_name + '/Sqrt',
power=mo_array(0.5))).create_node()
# Due to specification, Power must have inputs with the same data type.
convert_pow_input = Cast(
graph,
dict(dst_type=np.float32,
name=shape_values_node.name +
'/ConvertToFP32')).create_node()
div_node = Div(graph, dict(name="Div")).create_node()
shape_values_node.out_port(0).connect(convert_pow_input.in_port(0))
convert_pow_input.out_port(0).connect(pow_node.in_port(0))
div_sqrt.in_port(0).get_connection().set_destination(
div_node.in_port(0))
div_node.in_port(1).connect(pow_node.out_port(0))
div_sqrt.out_port(0).get_connection().set_source(div_node.out_port(0))
rename_nodes([(div_sqrt, div_sqrt_name + '/ShouldBeDeleted'),
(div_node, div_sqrt_name)])
def replace_tf_resize(graph: Graph, resize: Node, interpolation_mode: str):
resize_name = resize.soft_get('name', resize.id)
log.debug(
"Converting of {} to Interpolate-4 is triggered for node {}.".format(
resize.op, resize_name))
num_of_inputs = len([
port for port in resize.in_ports().values() if not port.disconnected()
])
assert num_of_inputs == 2, \
"Number of inputs of {} (with name {}) should be equal to 2".format(resize.op, resize_name)
attrs_msg = "If half_pixel_centers attribute of the node {} with op {} is True, " \
"the attribute align_corners must be False"
assert not resize.half_pixel_centers or (resize.half_pixel_centers and not resize.align_corners), \
attrs_msg.format(resize_name, resize.op)
shape = Shape(graph, {'name': resize_name + '/shapeof'}).create_node()
ss = create_op_with_const_inputs(graph, StridedSlice, {
1: int64_array([1]),
2: int64_array([3]),
3: int64_array([1])
}, {
'name': resize_name + '/StridedSlice',
'begin_mask': int64_array([1]),
'end_mask': int64_array([1]),
'new_axis_mask': int64_array([0]),
'shrink_axis_mask': int64_array([0]),
'ellipsis_mask': int64_array([0])
})
div_node = Div(graph, {'name': resize_name + '/Div'}).create_node()
shape_to_float = Cast(graph, dict(dst_type=np.float32)).create_node()
size_to_float = Cast(graph, dict(dst_type=np.float32)).create_node()
size_to_float.out_port(0).connect(div_node.in_port(0))
shape_to_float.out_port(0).connect(div_node.in_port(1))
ss.out_port(0).connect(shape_to_float.in_port(0))
shape.out_port(0).connect(ss.in_port(0))
align_corners = resize.align_corners
half_pixel_centers = resize.half_pixel_centers
nearest_mode = 'floor' if interpolation_mode == 'nearest' else 'round_prefer_floor'
if align_corners:
coordinate_transformation_mode = 'align_corners'
if interpolation_mode == 'nearest':
nearest_mode = 'round_prefer_ceil'
elif half_pixel_centers:
coordinate_transformation_mode = 'tf_half_pixel_for_nn' if interpolation_mode == 'nearest' else 'half_pixel'
else:
coordinate_transformation_mode = 'asymmetric'
interpolate4 = create_op_with_const_inputs(
graph, Interpolate, {3: int64_array([1, 2])}, {
'name': resize_name + '/interpolate_4',
'mode': interpolation_mode,
'antialias': False,
'coordinate_transformation_mode': coordinate_transformation_mode,
'pads_begin': int64_array([0]),
'pads_end': int64_array([0]),
'nearest_mode': nearest_mode,
'cube_coeff': -0.75,
'shape_calculation_mode': 'sizes',
'version': 'opset4',
'in_ports_count': 4,
})
resize_input_connection = resize.in_port(0).get_connection()
resize_input_connection.set_destination(interpolate4.in_port(0))
resize_input_connection.get_source().connect(shape.in_port(0))
div_node.out_port(0).connect(interpolate4.in_port(2))
sizes_connection = resize.in_port(1).get_connection()
sizes_connection.set_destination(interpolate4.in_port(1))
sizes_connection.get_source().connect(size_to_float.in_port(0))
resize.out_port(0).get_connection().set_source(interpolate4.out_port(0))
rename_nodes([(resize, resize_name + '/delete'),
(interpolate4, resize_name)])
def find_and_replace_pattern(self, graph: Graph):
for embedding_segments_mean in graph.get_op_nodes(
op='EmbeddingSegmentsMean'):
embedding_segments_mean_name = embedding_segments_mean.soft_get(
'name', embedding_segments_mean.id)
embedding_table_input = embedding_segments_mean.in_port(0)
segment_ids_input = embedding_segments_mean.in_port(2)
num_segments_input = embedding_segments_mean.in_port(3)
# TODO: support EmbeddingSegmentsMean with specified weights vector.
# now this case has not appeared in models so far so EmbeddingSegmentsOperation fusion
# transformations do not handle it either
if embedding_segments_mean.is_in_port_connected(5):
return
# 1. compute indices membership matrix, i.e. which indices belong to some object
# the shape of this matrix is [num_segments, num_indices]
non_norm_range_1_to_num_segments = create_op_with_const_inputs(
graph, Range, {
0: int64_array(0),
2: int64_array(1)
}, {
'name':
embedding_segments_mean_name + '/Range1ToNumSegments',
'output_type': np.int64
})
num_segments_input.get_connection().add_destination(
non_norm_range_1_to_num_segments.in_port(1))
range_1_to_num_segments = ConvertLike(graph, {
'name':
embedding_segments_mean_name + '/Range1ToNumSegmentsNorm'
}).create_node()
range_1_to_num_segments.in_port(0).connect(
non_norm_range_1_to_num_segments.out_port(0))
num_segments_input.get_connection().add_destination(
range_1_to_num_segments.in_port(1))
unsqueeze_range_1_to_num_segments = create_op_with_const_inputs(
graph, Unsqueeze, {1: int64_array(1)}, {
'name':
embedding_segments_mean_name +
'/Range1ToNumSegmentsUnsqueeze'
})
unsqueeze_range_1_to_num_segments.in_port(0).connect(
range_1_to_num_segments.out_port(0))
unsqueeze_segment_ids = create_op_with_const_inputs(
graph, Unsqueeze, {1: int64_array(0)}, {
'name':
embedding_segments_mean_name + '/SegmentIdsUnsqueeze'
})
segment_ids_input.get_connection().add_destination(
unsqueeze_segment_ids.in_port(0))
boolean_membership_matrix = Equal(graph, {
'name':
embedding_segments_mean_name + '/BooleanMembershipMatrix'
}).create_node()
boolean_membership_matrix.in_port(0).connect(
unsqueeze_range_1_to_num_segments.out_port(0))
boolean_membership_matrix.in_port(1).connect(
unsqueeze_segment_ids.out_port(0))
shape_of_membership_matrix = Shape(graph, {
'name':
embedding_segments_mean_name + '/ShapeOfMembershipMatrix'
}).create_node([boolean_membership_matrix])
one_scalar_constant = Const(
graph, {
'name': embedding_segments_mean_name + '/OneScalar',
'value': int64_array([1])
}).create_node()
one_constant = Broadcast(graph, {
'name':
embedding_segments_mean_name + '/One'
}).create_node([one_scalar_constant, shape_of_membership_matrix])
zero_constant = Const(
graph, {
'name': embedding_segments_mean_name + '/Zero',
'value': int64_array(0)
}).create_node()
membership_matrix = Select(
graph, {
'name': embedding_segments_mean_name + '/MembershipMatrix',
'auto_broadcast': 'numpy'
}).create_node(
[boolean_membership_matrix, one_constant, zero_constant])
# 2. compute a number of indices belong to each object from the batch
# it computes the normalization coefficients
num_indices_per_object = create_op_with_const_inputs(
graph, ReduceSum, {1: int64_array(1)}, {
'name':
embedding_segments_mean_name + '/NumIndicesPerObject'
})
num_indices_per_object.in_port(0).connect(
membership_matrix.out_port(0))
# 3. replace zero coefficient (zero number of indices belong to an object) with one
# because for such object the single default embedding vector is used
where_zero_number = Equal(graph, {
'name':
embedding_segments_mean_name + '/WhereZeroIndicesNumber'
}).create_node([num_indices_per_object, zero_constant])
normalized_num_indices_per_object = Select(
graph, {
'name':
embedding_segments_mean_name + '/NormNumIndicesPerObject',
'auto_broadcast': 'numpy'
}).create_node([
where_zero_number, one_scalar_constant,
num_indices_per_object
])
# 4. cast normalized_num_indices_per_object to the same type as embedding vector table
norm_coefficients = ConvertLike(
graph, {
'name': embedding_segments_mean_name + '/NormCoefficients'
}).create_node()
norm_coefficients.in_port(0).connect(
normalized_num_indices_per_object.out_port(0))
embedding_table_input.get_connection().add_destination(
norm_coefficients.in_port(1))
# 5. replace EmbeddingSegmentMean with EmbeddingSegmentSum
embedding_segments_sum = EmbeddingSegmentsSum(
graph, {
'name':
embedding_segments_mean_name + '/EmbeddingSegmentsSum'
}).create_node()
for in_port in embedding_segments_mean.in_ports():
if embedding_segments_mean.is_in_port_connected(in_port):
embedding_segments_mean.in_port(
in_port).get_connection().set_destination(
embedding_segments_sum.in_port(in_port))
# 6. normalize EmbeddingSegmentSum results by computed coefficients
result_node = Div(graph, {
'name': embedding_segments_mean_name + '/Div'
}).create_node([embedding_segments_sum, norm_coefficients])
embedding_segments_mean.out_port(0).get_connection().set_source(
result_node.out_port(0))
rename_nodes([(embedding_segments_mean,
embedding_segments_mean_name + '/AbandonedName'),
(result_node, embedding_segments_mean_name)])
graph.remove_nodes_from([embedding_segments_mean.id])
def replace_resize(graph: Graph, resize: Node):
log.debug("Converting of ONNX Resize-11 to Interpolate-4 "
"is triggered for node {}.".format(
resize.soft_get('name', resize.id)))
input_shape = resize.in_port(0).data.get_shape()
input_rank = len(input_shape)
resize_name = resize.soft_get('name', resize.id)
if input_rank not in {4, 5}:
log.warning(
'The input shape is not 4D or 5D for op with name {}'.format(
resize_name))
return
assert (resize.is_in_port_connected(0) and (resize.is_in_port_connected(2) or resize.is_in_port_connected(3))), \
"Scales or sizes inputs must be connected to Node {} with op {}.".format(resize.soft_get("name", resize.id),
resize.op)
assert resize.soft_get('coordinate_transformation_mode') != 'tf_crop_and_resize', \
'Mode tf_crop_and_resize is not supported for op {} with name {}'.format(resize.op,
resize.soft_get("name", resize.id))
layout = graph.graph['layout']
if input_rank == 4:
begin_dim = get_height_dim(layout, input_rank)
end_dim = get_width_dim(layout, input_rank) + 1
else:
begin_dim = get_depth_dim(layout, input_rank)
end_dim = get_width_dim(layout, input_rank) + 1
sizes_ss = create_op_with_const_inputs(
graph, StridedSlice, {
1: int64_array([begin_dim]),
2: int64_array([end_dim]),
3: int64_array([1])
}, {
'name': resize_name + '/StridedSlice_sizes',
'begin_mask': int64_array([1]),
'end_mask': int64_array([1]),
'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([begin_dim]),
2: int64_array([end_dim]),
3: int64_array([1])
}, {
'name': resize_name + '/StridedSlice_scales',
'begin_mask': int64_array([1]),
'end_mask': int64_array([1]),
'new_axis_mask': int64_array([0]),
'shrink_axis_mask': int64_array([0]),
'ellipsis_mask': int64_array([0])
})
axes_node = Const(
graph, {
'name': resize_name + '/axis',
'value': int64_array(np.arange(begin_dim, end_dim))
}).create_node()
shape_calculation_mode = 'sizes' if resize.is_in_port_connected(
3) else 'scales'
interpolate_node = Interpolate(
graph, {
'version': 'opset4',
'mode': convert_mode(resize.mode),
'coordinate_transformation_mode':
resize.coordinate_transformation_mode,
'cube_coeff': resize.cube_coeff,
'nearest_mode': resize.nearest_mode,
'pads_begin': int64_array([0]),
'pads_end': int64_array([0]),
'antialias': 0,
'shape_calculation_mode': shape_calculation_mode,
'in_ports_count': 4
}).create_node()
axes_node.out_port(0).connect(interpolate_node.in_port(3))
shape_of = Shape(graph, {'name': resize_name + '/ShapeOf'}).create_node()
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
if not resize.is_in_port_connected(3):
cast_shape_to_float = Cast(graph, {
'dst_type': dst_dtype
}).create_node()
mul_node = Mul(graph, {'name': resize_name + '/Mul'}).create_node()
shape_of.out_port(0).connect(cast_shape_to_float.in_port(0))
cast_shape_to_float.out_port(0).connect(mul_node.in_port(0))
cast_add_result_to_int = Cast(graph, {
'dst_type': np.int64
}).create_node()
floor_node = Floor(graph, {
'name': resize_name + '/Floor'
}).create_node()
mul_node.out_port(0).connect(add_node.in_port(0))
add_node.out_port(0).connect(floor_node.in_port(0))
floor_node.out_port(0).connect(cast_add_result_to_int.in_port(0))
cast_add_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(2).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_scales.get_source().connect(mul_node.in_port(1))
else:
cast_shape_to_float = Cast(graph, {
'dst_type': dst_dtype
}).create_node()
cast_sizes_to_float = Cast(graph, {
'dst_type': dst_dtype
}).create_node()
div_node = Div(graph, {'name': resize_name + '/Div'}).create_node()
cast_sizes_to_float.out_port(0).connect(div_node.in_port(0))
cast_shape_to_float.out_port(0).connect(div_node.in_port(1))
shape_of.out_port(0).connect(cast_shape_to_float.in_port(0))
div_node.out_port(0).connect(add_node.in_port(0))
add_node.out_port(0).connect(scales_ss.in_port(0))
scales_ss.out_port(0).connect(interpolate_node.in_port(2))
sizes_ss.out_port(0).connect(interpolate_node.in_port(1))
connection_of_resize_input = resize.in_port(0).get_connection()
connection_of_resize_input.set_destination(interpolate_node.in_port(0))
connection_of_sizes = resize.in_port(3).get_connection()
connection_of_sizes.set_destination(sizes_ss.in_port(0))
connection_of_resize_input.get_source().connect(shape_of.in_port(0))
connection_of_sizes.get_source().connect(
cast_sizes_to_float.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 dequantize_data(fake_quantize: Node, dst_type: type,
quantized_type: type) -> Node:
graph = fake_quantize.graph
quantized_data = fake_quantize.in_port(0).get_source().node
name = fake_quantize.soft_get('name', fake_quantize.id)
assert quantized_data.soft_get('type') == 'Convert' and quantized_data.dst_type == quantized_type, \
'Weights aren`t compressed as expected for node {}'.format(fake_quantize.soft_get('name', fake_quantize.id))
dequantizing_cast = Cast(
graph,
dict(name=quantized_data.name +
"/to_{}".format(np_data_type_to_destination_type(dst_type)),
dst_type=dst_type,
stop_value_propagation=True)).create_node()
fake_quantize.in_port(0).get_connection().set_destination(
dequantizing_cast.in_port(0))
# limits of dequantize
in_low = fake_quantize.in_port(1).get_source()
in_high = fake_quantize.in_port(2).get_source()
out_low = fake_quantize.in_port(3).get_source()
out_high = fake_quantize.in_port(4).get_source()
# scale calculation
output_range = Sub(graph, {
'name': name + '/output_range'
}).create_node()
output_range.in_port(0).connect(out_high)
output_range.in_port(1).connect(out_low)
input_range = Sub(graph, {'name': name + '/input_range'}).create_node()
input_range.in_port(0).connect(in_high)
input_range.in_port(1).connect(in_low)
scale = Div(graph, {'name': name + '/scale'}).create_node()
scale.in_port(0).connect(output_range.out_port(0))
scale.in_port(1).connect(input_range.out_port(0))
# shift calculation
descaled_output_low = Div(graph, {
'name': name + '/descaled_output_low'
}).create_node()
descaled_output_low.in_port(0).connect(out_low)
descaled_output_low.in_port(1).connect(scale.out_port(0))
shift = Sub(graph, {'name': name + '/shift'}).create_node()
shift.in_port(0).connect(in_low)
shift.in_port(1).connect(descaled_output_low.out_port(0))
zero = Const(graph, {
'name': name + '/zero',
'value': mo_array(0, dtype=dst_type)
}).create_node()
scale_eq_zero = Equal(graph, {
'name': name + '/scale_eq_zero'
}).create_node()
scale_eq_zero.in_port(0).connect(scale.out_port(0))
scale_eq_zero.in_port(1).connect(zero.out_port(0))
zero_point = Select(graph, {
'name': name + '/zero_point'
}).create_node()
zero_point.in_port(0).connect(scale_eq_zero.out_port(0))
zero_point.in_port(1).connect(zero.out_port(0))
zero_point.in_port(2).connect(shift.out_port(0))
# DeQuantize(x) == Mul(Sub(x, zero_point), scale)
sub_zp = Sub(graph, {'name': name + '/minus_zp'}).create_node()
sub_zp.in_port(0).connect(dequantizing_cast.out_port(0))
sub_zp.in_port(1).connect(zero_point.out_port(0))
mul_scale = Mul(graph, {
'name': name + '/mulpiply_by_scale'
}).create_node()
mul_scale.in_port(0).connect(sub_zp.out_port(0))
mul_scale.in_port(1).connect(scale.out_port(0))
fake_quantize.out_port(0).get_connection().set_source(
mul_scale.out_port(0))
graph.remove_nodes_from([fake_quantize.id, fake_quantize.out_node(0)])