def infer(node: Node):
in_shape = node.in_port(0).data.get_shape()
if in_shape.size != 4:
raise Error('TensorFlow DepthToSpace operation is supported for 4D \'NHWC\' input layout only. '
'Current input shape is \'{}\''.format(in_shape))
layout = node.graph.graph['layout']
N = in_shape[get_batch_dim(layout, 4)]
H = in_shape[get_height_dim(layout, 4)]
W = in_shape[get_width_dim(layout, 4)]
C = in_shape[get_features_dim(layout, 4)]
block_size = node['block_size']
if C is not dynamic_dimension and C % (block_size ** 2):
raise Error('Feature dimensions of input tensor of DepthToSpace operation have to be divisible by square '
'of DepthToSpace \'block_size\' parameter. Input tensor shape = {}. Feature dimension = {}. '
'block_size = {}'.format(in_shape, C, block_size))
out_shape = shape_for_layout(layout,
batch=N,
features=C // (block_size * block_size),
height=H * block_size,
width=W * block_size)
if is_fully_defined(in_shape) and is_fully_defined(out_shape) and np.prod(in_shape) != np.prod(out_shape):
raise Error('Number of input elements "{}" is not equal to number of output elements "" for node "{}"'
''.format(in_shape, out_shape, node.soft_get('name', node.id)))
node.out_port(0).data.set_shape(out_shape)
def infer(node: Node):
in_shape = node.in_node().shape
if in_shape.size != 4:
raise Error('TensorFlow SpaceToDepth operation is supported for 4D \'NHWC\' input layout only. '
'Current input shape is \'{}\''.format(in_shape))
layout = node.graph.graph['layout']
N = in_shape[get_batch_dim(layout, 4)]
H = in_shape[get_height_dim(layout, 4)]
W = in_shape[get_width_dim(layout, 4)]
C = in_shape[get_features_dim(layout, 4)]
block_size = node['block_size']
if (H is not dynamic_dimension and H % block_size) or (W is not dynamic_dimension and W % block_size):
raise Error('Spatial dimensions of input tensor of SpaceToDepth operation have to be divisible by '
'SpaceToDepth \'block_size\' parameter. Input tensor shape = {}. Spatial dimensions = {},{}. '
'block_size = {}'.format(in_shape, H, W, block_size))
out_shape = shape_for_layout(layout,
batch=N,
features=C * (block_size ** 2),
height=H // block_size,
width=W // block_size)
node.out_port(0).data.set_shape(out_shape)
def regionyolo_infer(node: Node):
input_shape = node.in_port(0).data.get_shape()
axis = get_canonical_axis_index(input_shape, node.axis)
end_axis = get_canonical_axis_index(input_shape, node.end_axis)
node.axis = axis
node.end_axis = end_axis
if node.do_softmax:
dims_to_flatten = input_shape[axis:end_axis + 1]
if is_fully_defined(dims_to_flatten):
flat_dim = np.ma.prod(dims_to_flatten)
else:
flat_dim = dynamic_dimension_value
node.out_port(0).data.set_shape(
[*input_shape[:axis], flat_dim, *input_shape[end_axis + 1:]])
else:
layout = node.graph.graph['layout']
assert len(layout) == 4
node.out_port(0).data.set_shape(
shape_for_layout(layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=(node.classes + node.coords + 1) *
len(node.mask),
height=input_shape[get_height_dim(layout, 4)],
width=input_shape[get_width_dim(layout, 4)]))
def priorbox_clustered_infer(node: Node):
layout = node.graph.graph['layout']
data_shape = node.in_node(0).shape
num_ratios = len(node.width)
if node.has_and_set('V10_infer'):
assert node.in_node(0).value is not None
node.out_port(0).data.set_shape([2, np.prod(node.in_node(0).value) * num_ratios * 4])
else:
res_prod = data_shape[get_height_dim(layout, 4)] * data_shape[get_width_dim(layout, 4)] * num_ratios * 4
node.out_port(0).data.set_shape([1, 2, res_prod])
def upsample_infer(node: Node):
node_name = node.soft_get('name', node.id)
layout = node.graph.graph['layout']
assert len(
layout
) == 4, 'Input tensor rank must be equal to 4 for node "{}"'.format(
node_name)
input_shape = node.in_port(0).data.get_shape()
if len(node.in_nodes()) == 1:
in_height = input_shape[get_height_dim(layout, 4)]
in_width = input_shape[get_width_dim(layout, 4)]
assert node.has('width_scale') is not None and node.has(
'height_scale') is not None
if in_height is not dynamic_dimension:
out_height = math.floor(in_height * node.height_scale)
else:
out_height = dynamic_dimension
if in_width is not dynamic_dimension:
out_width = math.floor(in_width * node.width_scale)
else:
out_width = dynamic_dimension
node.out_port(0).data.set_shape(
shape_for_layout(layout,
batch=input_shape[get_batch_dim(layout, 4)],
features=input_shape[get_features_dim(
layout, 4)],
height=out_height,
width=out_width))
else:
scales = node.in_port(1).data.get_value()
assert scales is not None, 'The input with scales for node "{}" is not constant'.format(
node_name)
eps = 1e-5 # This is to make rounding in case of very close number to round to closest instead of down
# generic output shape calculation to support 5D input shape case
output_shape = shape_array(
[dynamic_dimension for _ in range(len(input_shape))])
for idx in range(len(output_shape)):
if input_shape[idx] is not dynamic_dimension:
output_shape[idx] = int(
(input_shape[idx] + eps) * scales[idx])
else:
output_shape[idx] = dynamic_dimension_value
node.out_port(0).data.set_shape(output_shape)
def priorbox_infer(node: Node):
layout = node.graph.graph['layout']
data_shape = node.in_node(0).shape
# calculate all different aspect_ratios (the first one is always 1)
# in aspect_ratio 1/x values will be added for all except 1 if flip is True
ar_seen = [1.0]
ar_seen.extend(node.aspect_ratio.copy())
if node.flip:
for s in node.aspect_ratio:
ar_seen.append(1.0 / s)
ar_seen = np.unique(mo_array(ar_seen).round(decimals=6))
num_ratios = 0
if len(node.min_size) > 0:
num_ratios = len(ar_seen) * len(node.min_size)
if node.has_valid('fixed_size') and len(node.fixed_size) > 0:
num_ratios = len(ar_seen) * len(node.fixed_size)
if node.has_valid('density') and len(node.density) > 0:
for d in node.density:
if node.has_valid('fixed_ratio') and len(node.fixed_ratio) > 0:
num_ratios = num_ratios + len(
node.fixed_ratio) * (pow(d, 2) - 1)
else:
num_ratios = num_ratios + len(ar_seen) * (pow(d, 2) - 1)
num_ratios = num_ratios + len(node.max_size)
if node.has_and_set('V10_infer'):
assert node.in_node(0).value is not None
node.out_port(0).data.set_shape(
[2, np.prod(node.in_node(0).value) * num_ratios * 4])
else:
res_prod = data_shape[get_height_dim(
layout, 4)] * data_shape[get_width_dim(layout,
4)] * num_ratios * 4
node.out_port(0).data.set_shape([1, 2, res_prod])
def replace_pattern(self, graph: Graph, match: Dict[str, Node]):
log.debug('UpsampleToResample is triggered')
upsample = match['upsample']
upsample_name = upsample.soft_get('name', upsample.id)
input_shape = upsample.in_port(0).data.get_shape()
input_shape_rank = len(input_shape)
if input_shape_rank not in [4, 5]:
log.warning('The input shape is not 4D or 5D for op {}'.format(
upsample.soft_get('name')))
return
depth_scale = None
layout = graph.graph['layout']
if len(upsample.in_nodes()) == 2:
if upsample.in_node(1).value is None:
return
scales = upsample.in_node(1).value
assert len(scales) in (
4, 5
), 'Supported scales rank is 4 or 5, but it is {} for node {}'.format(
len(scales), upsample_name)
if not (math.isclose(scales[0], 1, rel_tol=1e-5)
and math.isclose(scales[1], 1, rel_tol=1e-5)):
return
height_scale = scales[get_height_dim(layout, input_shape_rank)]
width_scale = scales[get_width_dim(layout, input_shape_rank)]
if len(scales) == 5:
depth_scale = scales[get_depth_dim(layout, input_shape_rank)]
else:
height_scale = upsample['height_scale']
width_scale = upsample['width_scale']
if 1 in upsample.in_ports() and not upsample.in_port(1).disconnected():
upsample.in_port(1).disconnect()
upsample_name = upsample.soft_get('name', upsample.id)
shape = Shape(graph, {'name': upsample_name + '/0_port'}).create_node()
layout = graph.graph['layout']
if input_shape_rank == 4:
begin_value = int64_array(
[get_height_dim(layout, input_shape_rank)])
factor_value = float32_array([height_scale, width_scale])
else:
begin_value = int64_array(
[get_depth_dim(layout, input_shape_rank)])
factor_value = float32_array(
[depth_scale, height_scale, width_scale])
ss = create_op_with_const_inputs(
graph, StridedSlice, {
1: begin_value,
2: int64_array([get_width_dim(layout, input_shape_rank) + 1]),
3: int64_array([1])
}, {
'name': upsample_name + '/ss_0_port',
'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])
})
mul = create_op_node_with_second_input(
graph, Mul, factor_value, {'name': upsample_name + '/factor_mul'})
source = upsample.in_port(0).get_connection().get_source()
source.connect(shape.in_port(0))
shape.out_port(0).connect(ss.in_port(0))
ss.out_port(0).connect(mul.in_port(0))
# Create Interpolate operation
if input_shape_rank == 4:
axes = int64_array([
get_height_dim(layout, input_shape_rank),
get_width_dim(layout, input_shape_rank)
])
else:
axes = int64_array([
get_depth_dim(layout, input_shape_rank),
get_height_dim(layout, input_shape_rank),
get_width_dim(layout, input_shape_rank)
])
axes_node = Const(graph, {
'name': upsample_name + '/axis',
'value': axes
}).create_node()
interpolate = Interpolate(
graph, {
'mode': upsample.attrs()['mode'],
'antialias': 0,
'pads_begin': int64_array([0]),
'pads_end': int64_array([0]),
'coordinate_transformation_mode': 'half_pixel',
'nearest_mode': 'round_prefer_floor',
'cube_coeff': -0.75,
'shape_calculation_mode': 'scales',
'version': 'opset4',
'in_ports_count': 4
}).create_node()
upsample.add_input_port(1, skip_if_exist=True)
assert upsample.in_port(1).disconnected()
mul.out_port(0).connect(interpolate.in_port(1))
axes_node.out_port(0).connect(interpolate.in_port(3))
scales_node = Const(graph, {
'name': upsample_name + '/scales',
'value': factor_value
}).create_node()
scales_node.out_port(0).connect(interpolate.in_port(2))
upsample.in_port(0).get_connection().set_destination(
interpolate.in_port(0))
upsample.out_port(0).get_connection().set_source(
interpolate.out_port(0))
rename_nodes([(upsample, upsample_name + '/delete'),
(interpolate, upsample_name)])
convert_to_float = Cast(graph, dict(dst_type=np.float32)).create_node()
convert_to_int = Cast(graph, dict(dst_type=np.int64)).create_node()
mul.in_port(0).get_connection().insert_node(convert_to_float)
mul.out_port(0).get_connection().insert_node(convert_to_int)
def test_get_height_dim_NDHWC(self):
self.assertEqual(get_height_dim('NHWC', 5), 2)
def test_get_height_dim_NCDHW(self):
self.assertEqual(get_height_dim('NCHW', 5), 3)
def test_get_height_dim_NHWC(self):
self.assertEqual(get_height_dim('NHWC', 4), 1)
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 interp_infer(node: Node):
layout = node.graph.graph['layout']
assert len(layout) == 4
if len(node.in_nodes()) == 2:
src_shape = node.in_node(0).shape
dst_shape = node.in_node(1).shape
# in Caffe can be 2 inputs too, but shape should be got from shape of the second input
if node.parse_2nd_input == 'shape':
dst_shape = [dst_shape[get_height_dim(layout, 4)], dst_shape[get_width_dim(layout, 4)]]
else:
# it is TF case
dst_shape = node.in_node(1).value
if src_shape is None or dst_shape is None or len(src_shape) != 4 or len(dst_shape) != 2:
log.error(
'Node {} with op {} cannot be converted to Resample layer because there is no enough info about '
'src/dst shapes: src_shape = {}, dst_shape = {}'.format(node.name, node.op, src_shape, dst_shape))
node.type = None # prevent translation to a valid IE layer
return
in_height = src_shape[get_height_dim(layout, 4)]
in_width = src_shape[get_width_dim(layout, 4)]
out_height = dst_shape[0]
out_width = dst_shape[1]
node.factor = factor_update(
node.factor,
[float(out_height) / in_height, float(out_width) / in_width],
[in_height, in_width],
[out_height, out_width],
node.soft_get('name')
)
if node.factor is None:
node['width'] = out_width
node['height'] = out_height
node.out_node().shape = shape_for_layout(layout,
batch=src_shape[get_batch_dim(layout, 4)],
features=src_shape[get_features_dim(layout, 4)],
height=out_height,
width=out_width)
node.graph.remove_edge(node.in_node(1).id, node.id)
else:
outn = node.out_node(0)
in_shape = node.in_node(0)
num_ = in_shape.shape[get_batch_dim(layout, 4)]
channels_ = in_shape.shape[get_features_dim(layout, 4)]
height_in_ = in_shape.shape[get_height_dim(layout, 4)]
width_in_ = in_shape.shape[get_width_dim(layout, 4)]
height_out_ = height_in_ + node.pad_beg + node.pad_end
width_out_ = width_in_ + node.pad_beg + node.pad_end
if node.shrink_factor != 1 and node.zoom_factor == 1:
shrink_factor = node.shrink_factor
if shrink_factor < 1:
log.error('Shrink factor should be positive in node {}'.format(node.id))
return None
height_out_ = (height_out_ - 1) / shrink_factor + 1
width_out_ = (width_out_ - 1) / shrink_factor + 1
elif node.shrink_factor == 1 and node.zoom_factor != 1:
zoom_factor = node.zoom_factor
if zoom_factor < 1:
log.error('Zoom factor should be positive in node {}'.format(node.id))
return None
node['debug_message'] = 'Interp layer shape inference function may be wrong, please, try to update ' \
'layer shape inference function in the file (openvino/tools/mo/ops/interp.op at the ' \
'line {}).'.format(inspect.currentframe().f_lineno) + refer_to_faq_msg(100)
# Reshape methods can be different in some cases
# Commented out section represents reshape that used in deeplab-caffe
# Uncomment the following lines, if your model was trained with deeplab-caffe
# or have the same reshape method
# height_out_ = height_out_ + (height_out_ - 1) * (zoom_factor - 1)
# width_out_ = width_out_ + (width_out_ - 1) * (zoom_factor - 1)
# Comment out the following lines if you use the reshape method from previous section
height_out_ = height_out_ * zoom_factor
width_out_ = width_out_ * zoom_factor
elif node.width != 0 and node.height != 0:
height_out_ = node.height
width_out_ = node.width
elif node.shrink_factor != 1 and node.zoom_factor != 1:
shrink_factor = node.shrink_factor
zoom_factor = node.zoom_factor
if shrink_factor < 1:
log.error('Shrink factor should be positive in node {}'.format(node.id))
return None
if zoom_factor < 1:
log.error('Zoom factor should be positive in node {}'.format(node.id))
return None
height_out_ = (height_out_ - 1) / shrink_factor + 1
width_out_ = (width_out_ - 1) / shrink_factor + 1
height_out_ = height_out_ + (height_out_ - 1) * (zoom_factor - 1)
width_out_ = width_out_ + (width_out_ - 1) * (zoom_factor - 1)
outn.shape = shape_for_layout(layout,
batch=num_,
features=channels_,
height=height_out_,
width=width_out_)