def find_subgraph_match_to_pattern(graph: Graph, body_pattern: dict):
"""
Finds sub-graph matches corresponding pattern in graph
:param graph: a graph where to search for matched sub-graph
:param body_pattern: a pattern
:return: a list of sub-graph matches
"""
matches = []
for match in find_pattern_matches(graph, **body_pattern):
match = inverse_dict(match)
for k in match:
match[k] = Node(graph, match[k])
matches.append(match)
return matches
def transform_graph(self, graph: Graph, replacement_descriptions: dict):
matches = find_pattern_matches(graph, self.pattern_nodes,
self.pattern_edges)
for match in list(matches):
inverse_match = inverse_dict(match)
interpolate = Node(graph, inverse_match['interpolate'])
transpose_1 = Node(graph, inverse_match['transpose_1'])
transpose_2 = Node(graph, inverse_match['transpose_2'])
# because we remove Transpose layers the ResizeNearestNeighbor should be updated for NCHW layout
interpolate.axes = int64_array([2, 3])
transpose_1.in_port(0).get_connection().set_destination(
interpolate.in_port(0))
transpose_2.out_port(0).get_connection().set_source(
interpolate.out_port(0))
graph.remove_nodes_from([transpose_1.id, transpose_2.id])
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 transform_graph(self, graph: Graph, replacement_descriptions: dict):
matches = find_pattern_matches(graph, self.pattern_nodes,
self.pattern_edges)
for match in list(matches):
inverse_match = inverse_dict(match)
interpolate = Node(graph, inverse_match['interpolate'])
transpose_1 = Node(graph, inverse_match['transpose_1'])
transpose_2 = Node(graph, inverse_match['transpose_2'])
# Check for data layout and transposes orders
if graph.graph['layout'] != 'NCHW' or np.array_equal(transpose_1.in_port(1).data.get_value(), [0, 2, 3, 1]) or \
np.array_equal(transpose_2.in_port(1).data.get_value(), [0, 3, 1, 2]):
return
transpose_1.in_port(0).get_connection().set_destination(
interpolate.in_port(0))
transpose_2.out_port(0).get_connection().set_source(
interpolate.out_port(0))
graph.remove_nodes_from([transpose_1.id, transpose_2.id])
def replace_sub_graph(self, graph: Graph, match: dict):
slice_like = match['slice_like']
const = slice_like.in_nodes()[0]
crop_shape = slice_like.in_nodes()[1]
variants_dict = {
'mul_scalar1x': 0.1,
'mul_scalar2x': 0.2,
'mul_scalar1y': 0.1,
'mul_scalar2y': 0.2
}
for matches in find_pattern_matches(graph,
self.variants_pattern['nodes'],
self.variants_pattern['edges'],
None, None):
for k, v in matches.items():
if v in variants_dict.keys():
variants_dict[v] = Node(graph, k).in_nodes()[1].value[0]
variants = np.array([
variants_dict['mul_scalar1x'], variants_dict['mul_scalar1y'],
variants_dict['mul_scalar2x'], variants_dict['mul_scalar2y']
] * int(const.value.size / 4)).reshape(const.value.shape)
priorbox_variants = Const(
graph,
dict(value=variants,
symbol_dict={'name':
const.id + '/priorbox_variants'})).create_node()
variants_slice_like = Crop(graph, dict(axis=slice_like.axis, offset=slice_like.offset, dim=slice_like.dim, axes=slice_like.axes,
symbol_dict={'name': slice_like.id + '/variants_slice_like'})) \
.create_node()
variants_slice_like.in_port(0).connect(priorbox_variants.out_port(0))
variants_slice_like.in_port(1).connect(crop_shape.out_port(0))
concat = match['reshape3'].out_port(0).get_destination().node
assert concat.op == 'Concat'
concat_nodes_count = len(concat.in_nodes())
concat.add_input_port(concat_nodes_count)
concat.in_port(concat_nodes_count).get_connection().set_source(
variants_slice_like.out_port(0))
def replace_pattern(graph: Graph, match: dict):
node = match['op']
if node.name == 'iteration_number_out':
return
# calculate length of context when state of inference becomes meaningful
inputs = []
for n in graph.get_op_nodes(**{'op': 'Parameter'}):
inputs.append(n)
in_nodes = []
for inp in inputs:
for ins in inp.out_port(0).get_destinations():
in_nodes.append(ins.node.name)
context_len = 1
try:
subgraph = invert_sub_graph_between_nodes(
graph, [node.in_port(0).get_source().node.name], in_nodes)
except Error:
return
for n in subgraph:
n_node = Node(graph, n)
if n_node.kind == 'op' and n_node.op == 'Splice':
context_len += len(n_node.context) - 1
if context_len == 1:
return
in_node_port = node.in_port(0).get_source()
in_node_shape = node.in_port(0).data.get_shape()
node.in_port(0).disconnect()
# add Select before saving state to avoid saving garbage
select_node = Select(graph, {
'name': 'select_' + node.name
}).create_node()
zero_else = Const(graph, {
'name': 'zero_else',
'value': np.zeros(in_node_shape)
}).create_node()
select_node.in_port(1).connect(in_node_port)
select_node.in_port(2).connect(zero_else.out_port(0))
# check if we have already appropriate iteration counter
existing_counters = find_pattern_matches(
graph,
nodes=[('mem_in', dict(op='ReadValue')),
('mem_in_data', dict(shape=int64_array([context_len]))),
('crop_mem_in',
dict(op='Crop',
axis=int64_array([1]),
offset=int64_array([1]),
dim=int64_array([context_len - 1]))),
('crop_mem_in_data', dict()),
('concat', dict(op='Concat', axis=1)),
('concat_data', dict()), ('const_1', dict(op='Const')),
('const_1_data', dict()), ('mem_out', dict(op='Assign')),
('crop_out',
dict(op='Crop',
axis=int64_array([1]),
offset=int64_array([0]),
dim=int64_array([1]))), ('crop_out_data', dict()),
('select', dict(op='Select'))],
edges=[('mem_in', 'mem_in_data'), ('mem_in_data', 'crop_mem_in'),
('crop_mem_in', 'crop_mem_in_data'),
('crop_mem_in_data', 'concat', {
'in': 0
}), ('const_1', 'const_1_data'),
('const_1_data', 'concat', {
'in': 1
}), ('concat', 'concat_data'), ('concat_data', 'mem_out'),
('concat_data', 'crop_out'), ('crop_out', 'crop_out_data'),
('crop_out_data', 'select')])
counter_match = next(existing_counters, None)
if counter_match is not None:
ones = Node(graph, inverse_dict(counter_match)['const_1'])
input_port = Node(
graph,
inverse_dict(counter_match)['crop_out']).out_port(0)
else:
init_value_mem_out = create_zero_value_with_batch_from_input(
in_node_port, context_len, np.int32)
mem_out = ReadValue(
graph, {
'name': 'iteration_number',
'variable_id': 'iteration_' + node.name
}).create_node()
mem_out.in_port(0).connect(init_value_mem_out.out_port(0))
cut_first = Crop(
graph, {
'name': 'cut_first',
'axis': int64_array([1]),
'offset': int64_array([1]),
'dim': int64_array([context_len - 1])
}).create_node()
cut_first.in_port(0).connect(mem_out.out_port(0))
ones = Const(graph, {
'name': 'ones',
'value': np.ones([1, 1], dtype=np.int32)
}).create_node()
concat = Concat(graph, {
'name': 'concat_ones',
'in_ports_count': 2,
'axis': 1
}).create_node()
concat.in_port(0).connect(cut_first.out_port(0))
concat.in_port(1).connect(ones.out_port(0))
mem_in = Assign(
graph, {
'name': 'iteration_number_out',
'variable_id': 'iteration_' + node.name
}).create_node()
mem_in.in_port(0).connect(concat.out_port(0))
res = Result(graph, {}).create_node()
mem_in.out_port(0).connect(res.in_port(0))
cut_last = Crop(
graph, {
'name': 'cut_last',
'axis': int64_array([1]),
'offset': int64_array([0]),
'dim': int64_array([1])
}).create_node()
cut_last.in_port(0).connect(concat.out_port(0))
input_port = cut_last.out_port(0)
# Check if data from memory is 1
# if it is True, we have correct data and should proceed with saving it to memory
# else we have not gathered context and have garbage here, shouldn't change initial state of memory
cast_in = Equal(graph, {
'name': input_port.node.name + '/cast_to_bool'
}).create_node()
cast_in.in_port(0).connect(ones.out_port(0))
cast_in.in_port(1).connect(input_port)
select_node.in_port(0).connect(cast_in.out_port(0))
select_node.out_port(0).connect(node.in_port(0))
select_node.out_port(0).data.set_shape(in_node_shape)
def replace_pattern(graph: Graph, match: dict):
node = match['op']
if node.name == 'iteration_number_out':
return
# calculate length of context when state of inference becomes meaningful
inputs = []
for n in graph.get_op_nodes(**{'op': 'Parameter'}):
inputs.append(n)
in_nodes = []
for inp in inputs:
for ins in inp.out_port(0).get_destinations():
in_nodes.append(ins.node.name)
context_len = 1
try:
subgraph = invert_sub_graph_between_nodes(
graph, [node.in_port(0).get_source().node.name], in_nodes)
except Error:
return
for n in subgraph:
n_node = Node(graph, n)
if n_node.kind == 'op' and n_node.op == 'Splice':
context_len += len(n_node.context) - 1
if context_len == 1:
return
in_node_port = node.in_port(0).get_source()
in_node_shape = node.in_port(0).data.get_shape()
node.in_port(0).disconnect()
# add Select before saving state to avoid saving garbage
select_node = Select(graph, {
'name': 'select_' + node.name
}).create_node()
zero_else = Const(graph, {
'name': 'zero_else',
'value': np.zeros(in_node_shape)
}).create_node()
select_node.in_port(1).connect(in_node_port)
select_node.in_port(2).connect(zero_else.out_port(0))
# check if we have already appropriate iteration counter
existing_counters = find_pattern_matches(
graph,
nodes=[('mem_in',
dict(op='Memory',
index=1,
shape=int64_array([context_len]))),
('mem_in_data', dict()),
('crop_mem_in',
dict(op='Crop',
axis=int64_array([1]),
offset=int64_array([1]),
dim=int64_array([context_len - 1]))),
('crop_mem_in_data', dict()),
('concat', dict(op='Concat', axis=1)),
('concat_data', dict()), ('const_1', dict(op='Const')),
('const_1_data', dict()),
('mem_out',
dict(op='Memory',
index=0,
shape=int64_array([context_len]))),
('crop_out',
dict(op='Crop',
axis=int64_array([1]),
offset=int64_array([0]),
dim=int64_array([1]))), ('crop_out_data', dict()),
('select', dict(op='Select'))],
edges=[('mem_in', 'mem_in_data'), ('mem_in_data', 'crop_mem_in'),
('crop_mem_in', 'crop_mem_in_data'),
('crop_mem_in_data', 'concat', {
'in': 0
}), ('const_1', 'const_1_data'),
('const_1_data', 'concat', {
'in': 1
}), ('concat', 'concat_data'), ('concat_data', 'mem_out'),
('concat_data', 'crop_out'), ('crop_out', 'crop_out_data'),
('crop_out_data', 'select')])
counter_match = next(existing_counters, None)
if counter_match is not None:
input_port = Node(
graph,
inverse_dict(counter_match)['crop_out']).out_port(0)
else:
mem_out = Memory(
graph, {
'name': 'iteration_number',
'size': 2,
'index': 1,
'id': 'iteration_' + node.name,
'shape': int64_array([context_len]),
'dst_type': np.int32
}).create_node()
cut_first = Crop(
graph, {
'name': 'cut_first',
'axis': int64_array([1]),
'offset': int64_array([1]),
'dim': int64_array([context_len - 1])
}).create_node()
cut_first.in_port(0).connect(mem_out.out_port(0))
ones = Const(graph, {
'name': 'ones',
'value': np.ones([1, 1], dtype=np.int32)
}).create_node()
concat = Concat(graph, {
'name': 'concat_ones',
'in_ports_count': 2,
'axis': 1
}).create_node()
concat.in_port(0).connect(cut_first.out_port(0))
concat.in_port(1).connect(ones.out_port(0))
mem_in = Memory(
graph, {
'name': 'iteration_number_out',
'size': 2,
'index': 0,
'id': 'iteration_' + node.name,
'shape': int64_array([context_len])
}).create_node()
mem_in.in_port(0).connect(concat.out_port(0))
res = Result(graph, {}).create_node()
mem_in.out_port(0).connect(res.in_port(0))
cut_last = Crop(
graph, {
'name': 'cut_last',
'axis': int64_array([1]),
'offset': int64_array([0]),
'dim': int64_array([1])
}).create_node()
cut_last.in_port(0).connect(concat.out_port(0))
input_port = cut_last.out_port(0)
select_node.in_port(0).connect(input_port)
select_node.out_port(0).connect(node.in_port(0))
select_node.out_port(0).data.set_shape(in_node_shape)
def insert_select(graph: Graph, node: Node):
context_len = node.frame_time + 1
if context_len == 1:
return
in_node_port = node.in_port(0).get_source()
in_node_shape = node.in_port(0).data.get_shape()
node.in_port(0).disconnect()
# add Select before saving state to avoid saving garbage
select_node = Select(graph, {'name': 'select_' + node.name}).create_node()
zero_else = create_const_with_batch_from_input(in_node_port, in_node_shape[1])
select_node.in_port(1).connect(in_node_port)
select_node.in_port(2).connect(zero_else.out_port(0))
# check if we have already appropriate iteration counter
existing_counters = find_pattern_matches(graph, nodes=[('mem_in', dict(op='ReadValue')),
('mem_in_data', dict(shape=int64_array([context_len]))),
('crop_mem_in', dict(op='Crop', axis=int64_array([1]),
offset=int64_array([1]),
dim=int64_array([context_len - 1]))),
('crop_mem_in_data', dict()),
('concat', dict(op='Concat', axis=1)),
('concat_data', dict()),
('const_1', dict(op='Const')),
('const_1_data', dict()),
('mem_out', dict(op='Assign')),
('crop_out', dict(op='Crop', axis=int64_array([1]),
offset=int64_array([0]),
dim=int64_array([1]))),
('crop_out_data', dict()),
('select', dict(op='Select'))
],
edges=[('mem_in', 'mem_in_data'), ('mem_in_data', 'crop_mem_in'),
('crop_mem_in', 'crop_mem_in_data'),
('crop_mem_in_data', 'concat', {'in': 0}),
('const_1', 'const_1_data'),
('const_1_data', 'concat', {'in': 1}),
('concat', 'concat_data'), ('concat_data', 'mem_out'),
('concat_data', 'crop_out'), ('crop_out', 'crop_out_data'),
('crop_out_data', 'select')])
counter_match = next(existing_counters, None)
if counter_match is not None:
ones = Node(graph, inverse_dict(counter_match)['const_1'])
input_port = Node(graph, inverse_dict(counter_match)['crop_out']).out_port(0)
else:
init_value_mem_out = create_const_with_batch_from_input(in_node_port, context_len, precision=np.int32)
mem_out = ReadValue(graph, {'name': 'iteration_number',
'variable_id': 'iteration_' + node.name}).create_node()
mem_out.in_port(0).connect(init_value_mem_out.out_port(0))
cut_first = Crop(graph, {'name': 'cut_first', 'axis': int64_array([1]),
'offset': int64_array([1]), 'dim': int64_array([context_len - 1])}).create_node()
cut_first.in_port(0).connect(mem_out.out_port(0))
ones = create_const_with_batch_from_input(in_node_port, 1, 1, np.int32)
concat = Concat(graph, {'name': 'concat_ones', 'in_ports_count': 2, 'axis': 1}).create_node()
concat.in_port(0).connect(cut_first.out_port(0))
concat.in_port(1).connect(ones.out_port(0))
mem_in = Assign(graph, {'name': 'iteration_number_out',
'variable_id': 'iteration_' + node.name}).create_node()
mem_in.in_port(0).connect(concat.out_port(0))
res = Result(graph, {}).create_node()
mem_in.out_port(0).connect(res.in_port(0))
cut_last = Crop(graph, {'name': 'cut_last', 'axis': int64_array([1]),
'offset': int64_array([0]), 'dim': int64_array([1])}).create_node()
cut_last.in_port(0).connect(concat.out_port(0))
input_port = cut_last.out_port(0)
# Check if data from memory is 1
# if it is True, we have correct data and should proceed with saving it to memory
# else we have not gathered context and have garbage here, shouldn't change initial state of memory
cast_in = Equal(graph, {'name': input_port.node.name + '/cast_to_bool'}).create_node()
cast_in.in_port(0).connect(ones.out_port(0))
cast_in.in_port(1).connect(input_port)
select_node.in_port(0).connect(cast_in.out_port(0))
select_node.out_port(0).connect(node.in_port(0))
select_node.out_port(0).data.set_shape(in_node_shape)
