def generate_sub_graph(self, graph: Graph, match: SubgraphMatch):
# IE DetectionOutput layer consumes flattened confidences and locations tensors.
# That is why we add reshapes before them.
locs_node = match.single_input_node(0)
conf_node = match.single_input_node(1)
prior_boxes_node = match.single_input_node(2)
locs_out_nodes = locs_node[0].out_nodes()
assert len(locs_out_nodes) == 1
locs_out_node = locs_out_nodes[list(locs_out_nodes.keys())[0]]
assert locs_out_node.op == "OpOutput", locs_out_node.op
graph.remove_node(locs_out_node.id)
conf_out_nodes = conf_node[0].out_nodes()
assert len(conf_out_nodes) == 1
conf_out_node = conf_out_nodes[list(conf_out_nodes.keys())[0]]
assert conf_out_node.op == "OpOutput", conf_out_node.op
graph.remove_node(conf_out_node.id)
# reshape operation to flatten confidence tensor
reshape_loc_op = Reshape(graph, {'dim': np.array([0, -1])})
reshape_loc_node = reshape_loc_op.create_node(
[locs_node], dict(name='DetectionOutput_Reshape_loc_'))
# reshape operation to flatten confidence tensor
reshape_conf_op = Reshape(graph, {'dim': np.array([0, -1])})
reshape_conf_node = reshape_conf_op.create_node(
[conf_node], dict(name='DetectionOutput_Reshape_conf_'))
# remove the OpOutput node after the priors node
assert prior_boxes_node[0].out_node().op == "OpOutput"
graph.remove_node(prior_boxes_node[0].out_node().id)
# reshape operation for prior boxes tensor
reshape_priors_op = Reshape(graph, {'dim': np.array([1, 2, -1])})
reshape_priors_node = reshape_priors_op.create_node(
[prior_boxes_node], dict(name='DetectionOutput_Reshape_priors_'))
# create Detection Output node with three inputs: locations, confidences and prior boxes
detection_output_op = DetectionOutput(
graph, match.custom_replacement_desc.custom_attributes)
detection_output_node = detection_output_op.create_node(
[reshape_loc_node, reshape_conf_node, reshape_priors_node],
dict(name=detection_output_op.attrs['type'] + '_'))
PermuteAttrs.set_permutation(reshape_priors_node,
detection_output_node, None)
# create Output node to mark DetectionOutput as a graph output operation
output_op = Output(graph)
output_op.create_node([detection_output_node], dict(name='sink_'))
return {}
def replace_sub_graph(self, graph: Graph, match: dict):
box_nms = match['box_nms']
top_k = box_nms.topk
nms_threshold = box_nms.overlap_thresh
ssd_concats = {}
concat_names = ['ssd_concat1', 'ssd_concat0', 'ssd_concat2']
for i, concat_match in enumerate(self.concats_pattern):
for matches in find_pattern_matches(graph, concat_match['nodes'], concat_match['edges'], None, None):
for match in matches:
if graph.has_node(match):
n = Node(graph, match)
if n.op == 'Concat':
ssd_concats.update({concat_names[i]: n})
break
assert concat_names[0] in ssd_concats
assert concat_names[1] in ssd_concats
assert concat_names[2] in ssd_concats
graph.remove_nodes_from(graph.get_nodes_with_attributes(op='Result'))
detection_output_node = DetectionOutput(graph, dict(name=graph.unique_id() + '/DetectionOutput_',
top_k=top_k, keep_top_k=top_k, nms_threshold=nms_threshold,
background_label_id=0, clip=0, decrease_label_id=1,
code_type="caffe.PriorBoxParameter.CENTER_SIZE",
confidence_threshold=0.01, share_location=1,
variance_encoded_in_target=0, normalized=1)).create_node()
reshape_node = create_op_node_with_second_input(graph, Reshape, int64_array([0, -1]),
dict(name=graph.unique_id() + '/DetectionOutput_'))
ssd_softmax_node = ssd_concats['ssd_concat0'].out_node().out_node()
ssd_softmax_node.out_port(0).disconnect()
ssd_softmax_node.out_port(0).connect(reshape_node.in_port(0))
reshape_node.out_port(0).connect(detection_output_node.in_port(1))
ssd_concats['ssd_concat2'].axis = 2
self.reshape_priorboxes(ssd_concats['ssd_concat2'])
ssd_concats['ssd_concat1'].out_port(0).get_connection().set_destination(detection_output_node.in_port(0))
ssd_concats['ssd_concat2'].out_port(0).get_connection().set_destination(detection_output_node.in_port(2))
Result(graph, {'name': detection_output_node.id + '/Result'}).create_node([detection_output_node])
def generate_sub_graph(self, graph: Graph, match: SubgraphMatch):
reshape_classes_node = create_op_node_with_second_input(
graph, Reshape, int64_array([0, -1]),
dict(name='do_reshape_classes'),
match.single_input_node(1)[0])
priors_node = match.single_input_node(2)[0]
placeholder = [
Node(graph, node_id) for node_id in graph.nodes()
if Node(graph, node_id).op == 'Parameter'
][0]
im_height = placeholder.shape[1]
im_width = placeholder.shape[2]
# scale prior boxes to the [0, 1] interval
priors_scale_const_node = Const(
graph, {
'value':
np.array(
[1 / im_width, 1 / im_height, 1 / im_width, 1 / im_height])
}).create_node([])
priors_scale_node = Mul(graph, {
'name': 'scale_priors'
}).create_node([priors_node, priors_scale_const_node])
# calculate prior boxes widths and heights
split_node = SplitV(graph, {
'axis': 2,
'size_splits': [1, 1, 1, 1],
'out_ports_count': 4
}).create_node([priors_scale_node])
priors_width_node = Sub(
graph, dict(name=split_node.name + '/sub_2-0_')).create_node([
(split_node, 2), (split_node, 0)
])
priors_height_node = Sub(graph, dict(name=split_node.name +
'/sub_3-1_')).create_node([
(split_node, 3),
(split_node, 1)
])
# concat weights and heights into a single tensor and multiple with the box coordinates regression values
concat_width_height_node = Concat(graph, {
'name': 'concat_priors_width_height',
'axis': -1,
'in_ports_count': 4
}).create_node([
priors_width_node, priors_height_node, priors_width_node,
priors_height_node
])
applied_width_height_regressions_node = Mul(graph, {
'name': 'final_regressions'
}).create_node(
[concat_width_height_node,
match.single_input_node(0)[0]])
# reshape to 2D tensor as Inference Engine Detection Output layer expects
reshape_regression_node = create_op_node_with_second_input(
graph, Reshape, int64_array([0, -1]),
dict(name='reshape_regression'),
applied_width_height_regressions_node)
detection_output_op = DetectionOutput(
graph, match.custom_replacement_desc.custom_attributes)
detection_output_op.attrs['old_infer'] = detection_output_op.attrs[
'infer']
detection_output_op.attrs['infer'] = __class__.do_infer
detection_output_node = detection_output_op.create_node(
[reshape_regression_node, reshape_classes_node, priors_scale_node],
dict(name=detection_output_op.attrs['type'],
clip=1,
normalized=1,
variance_encoded_in_target=0))
return {'detection_output_node': detection_output_node}
def generate_sub_graph(self, graph: Graph, match: SubgraphMatch):
reshape_classes_node = create_op_node_with_second_input(
graph, Reshape, int64_array([0, -1]),
dict(name='do_reshape_classes'),
match.single_input_node(1)[0])
initial_priors_node = match.single_input_node(2)[0]
priors_name = initial_priors_node.soft_get('name',
initial_priors_node.id)
# model calculates identical prior boxes for each batch, so we take first slice of them
begin = Const(graph, {
'value': np.array([0, 0, 0], dtype=np.int32)
}).create_node()
end = Const(graph, {
'value': np.array([1, 0, 0], dtype=np.int32)
}).create_node()
stride = Const(graph, {
'value': np.array([1, 1, 1], dtype=np.int32)
}).create_node()
priors_node = StridedSlice(
graph, {
'name': priors_name + '/0_batch_slice',
'begin_mask': np.array([1, 1, 1], dtype=np.int32),
'end_mask': np.array([1, 0, 0], dtype=np.int32),
'new_axis_mask': np.array([0], dtype=np.int32),
'shrink_axis_mask': np.array([0], dtype=np.int32),
'ellipsis_mask': np.array([0], dtype=np.int32)
}).create_node()
initial_priors_node.out_port(0).connect(priors_node.in_port(0))
begin.out_port(0).connect(priors_node.in_port(1))
end.out_port(0).connect(priors_node.in_port(2))
stride.out_port(0).connect(priors_node.in_port(3))
placeholders = graph.get_op_nodes(type='Parameter')
assert len(placeholders) == 1, "{} replacer requires model to have one Placeholder, but current model has " \
"{} placeholders".format(self.replacement_id, len(placeholders))
placeholder = placeholders[0]
# scale prior boxes to the [0, 1] interval
node_with_scales_for_prior_boxes = self.placeholder_scales(placeholder)
priors_scale_node = Mul(graph, {'name': 'scale_priors'}).create_node()
broadcast = Broadcast(graph, {
'name': 'scales_broadcast'
}).create_node()
shape_of_priors = Shape(graph, {'name': 'priors_shape'}).create_node()
priors_node.out_port(0).connect(shape_of_priors.in_port(0))
broadcast.in_port(1).connect(shape_of_priors.out_port(0))
broadcast.in_port(0).connect(
node_with_scales_for_prior_boxes.out_port(0))
priors_scale_node.in_port(0).connect(priors_node.out_port(0))
priors_scale_node.in_port(1).connect(broadcast.out_port(0))
try:
variance = match.custom_replacement_desc.custom_attributes[
'variance']
except:
raise Error(
'There is no variance attribute in {} replacement config file `custom_attributes`'
''.format(self.replacement_id))
priors = self.append_variances(priors_scale_node, variance)
# calculate prior boxes widths and heights
split_node = create_op_with_const_inputs(graph, VariadicSplit, {
1: int64_array(2),
2: int64_array([1, 1, 1, 1])
}, {'out_ports_count': 4}, priors_scale_node)
priors_width_node = Sub(
graph, dict(name=split_node.name + '/sub_2-0_')).create_node([
(split_node, 2), (split_node, 0)
])
priors_height_node = Sub(graph, dict(name=split_node.name +
'/sub_3-1_')).create_node([
(split_node, 3),
(split_node, 1)
])
# concat weights and heights into a single tensor and multiple with the box coordinates regression values
# WA with 3 Concats instead of 1 for keeping model reshapable
# concat_width_height_node = Concat(graph, {'name': 'concat_priors_width_height', 'axis': -1,
# 'in_ports_count': 4}).create_node(
# [priors_width_node, priors_height_node, priors_width_node, priors_height_node])
concat_1 = Concat(graph, {
'name': 'concat_width_height',
'axis': -1,
'in_ports_count': 2
}).create_node([priors_width_node, priors_height_node])
concat_2 = Concat(graph, {
'name': 'concat_width_height_width',
'axis': -1,
'in_ports_count': 2
}).create_node([concat_1, priors_width_node])
concat_width_height_node = Concat(graph, {
'name': 'concat_priors_width_height',
'axis': -1,
'in_ports_count': 2
}).create_node([concat_2, priors_height_node])
applied_width_height_regressions_node = Mul(graph, {
'name': 'final_regressions'
}).create_node(
[concat_width_height_node,
match.single_input_node(0)[0]])
# reshape to 2D tensor as Inference Engine Detection Output layer expects
reshape_regression_node = create_op_node_with_second_input(
graph, Reshape, int64_array([0, -1]),
dict(name='reshape_regression'),
applied_width_height_regressions_node)
detection_output_op = DetectionOutput(
graph, match.custom_replacement_desc.custom_attributes)
# get nms from the original network
iou_threshold = None
nms_nodes = graph.get_op_nodes(op='NonMaxSuppression')
if len(nms_nodes) > 0:
# it is highly unlikely that for different classes NMS has different
# moreover DetectionOutput accepts only scalar values for iou_threshold (nms_threshold)
iou_threshold = nms_nodes[0].in_node(3).value
if iou_threshold is None:
raise Error(
'During {} `iou_threshold` was not retrieved from RetinaNet graph'
.format(self.replacement_id))
detection_output_node = detection_output_op.create_node(
[reshape_regression_node, reshape_classes_node, priors],
dict(name=detection_output_op.attrs['type'],
nms_threshold=iou_threshold,
clip_after_nms=1,
normalized=1,
variance_encoded_in_target=0,
background_label_id=1000))
return {'detection_output_node': detection_output_node}
def transform_graph(self, graph: Graph, replacement_descriptions: dict):
parameter_node = graph.get_op_nodes(op='Parameter')[0]
parameter_node['data_type'] = data_type_str_to_np(
parameter_node.graph.graph['cmd_params'].data_type)
parameter_node.out_port(0).disconnect()
# remove existing Result operations to remove unsupported sub-graph
graph.remove_nodes_from(
[node.id
for node in graph.get_op_nodes(op='Result')] + ['detections'])
# determine if the op which is a input/final result of mean value and scale applying to the input tensor
# then connect it to the input of the first convolution of the model, so we remove the image pre-processing
# which includes padding and resizing from the model
preprocessing_input_node_id = replacement_descriptions[
'preprocessing_input_node']
assert preprocessing_input_node_id in graph.nodes, 'The node with name "{}" is not found in the graph. This ' \
'should be a last node before image normalization and is specified' \
' in the json file.'.format(preprocessing_input_node_id)
preprocessing_input_node = Node(graph, preprocessing_input_node_id)
consumer_node = preprocessing_input_node.out_port(
0).get_connection().get_destination().node
consumer_node.in_port(0).get_connection().set_source(
parameter_node.out_port(0))
preprocessing_output_node_id = replacement_descriptions[
'preprocessing_output_node']
assert preprocessing_output_node_id in graph.nodes, 'The node with name "{}" is not found in the graph. This ' \
'node should provide scaled image output and is specified' \
' in the json file.'.format(preprocessing_output_node_id)
preprocessing_output_node = Node(graph, preprocessing_output_node_id)
preprocessing_output_node.out_port(0).disconnect()
convolution_nodes = [
n for n in graph.pseudo_topological_sort()
if n.soft_get('type') == 'Convolution'
]
convolution_nodes[0].in_port(0).get_connection().set_source(
preprocessing_output_node.out_port(0))
# create prior boxes (anchors) generator
aspect_ratios = replacement_descriptions['aspect_ratios']
assert len(aspect_ratios) % 2 == 0
aspect_ratios = list(zip(aspect_ratios[::2], aspect_ratios[1::2]))
priors_generator = self.AnchorGenerator(
min_level=int(replacement_descriptions['min_level']),
aspect_ratios=aspect_ratios,
num_scales=int(replacement_descriptions['num_scales']),
anchor_scale=replacement_descriptions['anchor_scale'])
prior_boxes = []
for i in range(100):
inp_name = 'box_net/box-predict{}/BiasAdd'.format('_%d' %
i if i else '')
if inp_name not in graph:
break
widths, heights = priors_generator.get(i)
prior_box_op = PriorBoxClusteredOp(
graph, {
'width': np.array(widths),
'height': np.array(heights),
'clip': 0,
'flip': 0,
'variance': replacement_descriptions['variance'],
'offset': 0.5
})
prior_boxes.append(
prior_box_op.create_node(
[Node(graph, inp_name), parameter_node]))
# concatenate prior box operations
concat_prior_boxes = Concat(graph, {'axis': -1}).create_node()
for idx, node in enumerate(prior_boxes):
concat_prior_boxes.add_input_port(idx)
concat_prior_boxes.in_port(idx).connect(node.out_port(0))
conf = Sigmoid(graph, dict(name='concat/sigmoid')).create_node(
[Node(graph, 'concat')])
reshape_size_node = Const(graph, {
'value': int64_array([0, -1])
}).create_node([])
logits = Reshape(graph, dict(name=conf.name + '/Flatten')).create_node(
[conf, reshape_size_node])
deltas = Reshape(graph, dict(name='concat_1/Flatten')).create_node(
[Node(graph, 'concat_1'), reshape_size_node])
# revert convolution boxes prediction weights from yxYX to xyXY (convolutions share weights and bias)
weights = Node(graph, 'box_net/box-predict/pointwise_kernel')
weights.value = weights.value.reshape(-1, 4)[:, [1, 0, 3, 2]].reshape(
weights.shape)
bias = Node(graph, 'box_net/box-predict/bias')
bias.value = bias.value.reshape(-1,
4)[:, [1, 0, 3, 2]].reshape(bias.shape)
detection_output_node = DetectionOutput(
graph,
dict(
name='detections',
num_classes=int(replacement_descriptions['num_classes']),
share_location=1,
background_label_id=int(
replacement_descriptions['num_classes']) + 1,
nms_threshold=replacement_descriptions['nms_threshold'],
confidence_threshold=replacement_descriptions[
'confidence_threshold'],
top_k=100,
keep_top_k=100,
code_type='caffe.PriorBoxParameter.CENTER_SIZE',
)).create_node([deltas, logits, concat_prior_boxes])
output_op = Result(graph, dict(name='output'))
output_op.create_node([detection_output_node])