def extract(cls, node):
mode = onnx_attr(node,
'mode',
's',
default='constant',
dst_type=lambda x: x.decode())
pads = onnx_attr(node,
'pads',
'ints',
dst_type=lambda x: np.array(x, dtype=np.int64))
value = onnx_attr(node, 'value', 'f', default=0.)
assert pads is not None
# MO Pad op and ONNX Pad op have different format for pads values
# MO Pad has Dx2 where D is the total number of dimensions
# ONNX Pad pads flat layout, so
# need to reshape and transpose
pads = np.transpose(pads.reshape([2, -1]))
Pad.update_node_stat(node, {
'mode': mode,
'pads': pads,
'fill_value': value
})
return cls.enabled
def test_two_inputs_value_infer(self):
in_value = np.random.rand(*self.node_attrs['data_in']['shape']).astype(
np.float32)
graph = build_graph(
self.node_attrs,
self.edge_attrs + [('pads_begin', 'pad'), ('pads_end', 'pad')],
{'data_in': {
'value': in_value
}},
nodes_with_edges_only=True,
)
pads = np.insert(
self.node_attrs['pads_end']['value'],
np.arange(len(self.node_attrs['pads_begin']['value'])),
self.node_attrs['pads_begin']['value'])
pads = np.reshape(pads,
(len(self.node_attrs['pads_begin']['value']), 2))
ref_value = np.pad(in_value, pads, constant_values=0, mode='constant')
pad_node = Node(graph, 'pad')
Pad.infer(pad_node)
self.assertTrue(
np.array_equal(
Node(graph, 'data_out').shape,
np.array([1, 3, 100 + 1 + 3, 200 + 2 + 4])))
self.assertTrue(
np.array_equal(Node(graph, 'data_out').value, ref_value))
def find_and_replace_pattern(self, graph: Graph):
for attr_pad in graph.get_op_nodes(op='AttributedPad'):
# save the original node name to use it in the new Pad op instance
original_name = attr_pad.soft_get('name', attr_pad.id)
new_pad = Pad(graph, {
'mode': attr_pad.soft_get('mode', None),
}).create_node()
rename_nodes([(attr_pad, original_name + '/to_be_removed'),
(new_pad, original_name)])
attr_pad.in_port(0).get_connection().set_destination(
new_pad.in_port(0))
new_pad.in_port(1).connect(
Const(graph, {
'value': attr_pad.pads[:, 0]
}).create_node().out_port(0))
new_pad.in_port(2).connect(
Const(graph, {
'value': attr_pad.pads[:, 1]
}).create_node().out_port(0))
if attr_pad.soft_get('mode') == 'constant':
new_pad.in_port(3).connect(
Const(graph, {
'value': attr_pad.fill_value
}).create_node().out_port(0))
attr_pad.out_port(0).get_connection().set_source(
new_pad.out_port(0))
graph.remove_node(attr_pad.id)
def test_two_inputs_dynamic_value_infer(self):
in_value = shape_array([dynamic_dimension_value, 3]).reshape(
(1, 1, 1, 2))
graph = build_graph(
self.node_attrs,
self.edge_attrs + [('pads_begin', 'pad'), ('pads_end', 'pad')],
{'data_in': {
'value': in_value,
'shape': in_value.shape
}},
nodes_with_edges_only=True,
)
out_shape = (1, 1, 5, 8)
mask = np.zeros(out_shape, dtype=np.bool)
mask[0][0][1][2] = True
ref_value = np.ma.masked_array(np.zeros(out_shape, dtype=np.int64),
mask=mask,
dtype=np.int64)
ref_value[0][0][1][3] = 3
pad_node = Node(graph, 'pad')
Pad.infer(pad_node)
output_value = Node(graph, 'data_out').value
self.assertTrue(
np.array_equal(Node(graph, 'data_out').shape, ref_value.shape))
self.assertTrue(strict_compare_tensors(output_value, ref_value))
self.assertTrue(isinstance(output_value, np.ma.masked_array))
self.assertTrue(output_value[0][0][1][2] is dynamic_dimension)
def find_and_replace_pattern(self, graph: Graph):
for node in graph.get_op_nodes(op='SpaceToBatch') + graph.get_op_nodes(op='BatchToSpace'):
node.add_input_port(3, skip_if_exist=True)
# convert TF representation of the pads/crops as [N, 2] to IE representation: [N] and [N]
transposed_pads = create_op_with_const_inputs(graph, Transpose, {1: int64_array([1, 0])})
node.in_port(2).get_connection().set_destination(transposed_pads.in_port(0))
split_pads = create_op_with_const_inputs(graph, Split, {1: int64_array(0)}, {'num_splits': 2})
transposed_pads.out_port(0).connect(split_pads.in_port(0))
for port_ind in range(2):
node.in_port(port_ind + 2).connect(split_pads.out_port(port_ind))
node.in_port(port_ind + 2).get_connection().insert_node(
create_op_with_const_inputs(graph, Squeeze, {1: int64_array([0])}))
# add zeros/ones to related inputs to align it with data input
in0_rank = Rank(graph, {'name': node.name + '/rank_0'}).create_node()
in1_rank = Shape(graph, {'name': node.name + '/rank_1'}).create_node()
diff_size = Sub(graph, {'name': node.name + '/sub_0'}).create_node()
diff = Sub(graph, {'name': node.name + '/sub_1'}).create_node()
const_begin = Const(graph, {'value': int64_array([1])}).create_node()
const_pad_val = Const(graph, {'value': int64_array([1])}).create_node()
block_shape = Pad(graph, {'name': node.name + '/aligned_block_shape', 'mode': 'constant'}).create_node()
# in case of SpaceToBatch begin = pads_begin, end = pads_end
# in case of BatchToSpace begin = crops_begin, end = crops_end
new_begin_name = '/aligned_pads_begin'
new_end_name = '/aligned_pads_end'
if node.type == 'BatchToSpace':
new_begin_name = '/aligned_crops_begin'
new_end_name = '/aligned_crops_end'
begin = Pad(graph, {'name': node.name + new_begin_name, 'mode': 'constant'}).create_node()
end = Pad(graph, {'name': node.name + new_end_name, 'mode': 'constant'}).create_node()
in0_rank_1d = create_op_node_with_second_input(graph, Unsqueeze, int64_array([0]),
{'name': node.name + '/1d_rank_of_0'}, in0_rank)
in1_rank_1d = create_op_node_with_second_input(graph, Unsqueeze, int64_array([0]),
{'name': node.name + '/1d_rank_of_1'}, in1_rank)
node.in_port(0).get_source().connect(in0_rank.in_port(0))
node.in_port(1).get_source().connect(in1_rank.in_port(0))
in0_rank_1d.out_port(0).connect(diff_size.in_port(0))
in1_rank_1d.out_port(0).connect(diff_size.in_port(1))
diff_size.out_port(0).connect(diff.in_port(0))
const_begin.out_port(0).connect(diff.in_port(1))
const_pad_val.out_port(0).connect(block_shape.in_port(3))
inputs_array = [block_shape, begin, end]
for idx, input_to_node in enumerate(inputs_array):
node.in_port(idx + 1).get_connection().set_destination(input_to_node.in_port(0))
const_begin.out_port(0).connect(input_to_node.in_port(1))
diff.out_port(0).connect(input_to_node.in_port(2))
input_to_node.out_port(0).connect(node.in_port(idx + 1))
def test_one_input_and_no_pads(self):
graph = build_graph(
self.node_attrs,
self.edge_attrs,
nodes_with_edges_only=True,
)
pad_node = Node(graph, 'pad')
with self.assertRaisesRegex(AssertionError,
".*pads attribute is missing.*"):
Pad.infer(pad_node)
def test_not_enough_inputs(self):
graph = build_graph(
self.node_attrs,
self.edge_attrs + [('pads_begin', 'pad')],
nodes_with_edges_only=True,
)
pad_node = Node(graph, 'pad')
with self.assertRaisesRegex(AssertionError,
".*must have 3 or 4 inputs.*"):
Pad.infer(pad_node)
def find_and_replace_pattern(self, graph: Graph):
for tfpad in graph.get_op_nodes(op='TFPad'):
# save the original node name to use it in the new Pad op instance
original_name = tfpad.soft_get('name', tfpad.id)
tfpad['name'] = original_name + '/to_be_removed'
new_pad = Pad(graph, {'mode': tfpad.soft_get('mode', None), }).create_node()
rename_node(new_pad, original_name)
tfpad.in_port(0).get_connection().set_destination(new_pad.in_port(0))
if tfpad.soft_get('mode') == 'constant':
# the input with fill value is an optional third input in TF
if not tfpad.in_port(2).disconnected():
tfpad.in_port(2).get_connection().set_destination(new_pad.in_port(3))
# convert TF representation of the pads as [N, 2] to MO representation: [N] and [N]
transposed_pads = create_op_with_const_inputs(graph, Transpose, {1: int64_array([1, 0])})
tfpad.in_port(1).get_connection().set_destination(transposed_pads.in_port(0))
split_pads = create_op_with_const_inputs(graph, Split, {1: int64_array(0)}, {'num_splits': 2})
transposed_pads.out_port(0).connect(split_pads.in_port(0))
for port_ind in range(2):
split_pads.add_output_port(port_ind, skip_if_exist=True)
new_pad.in_port(port_ind + 1).connect(split_pads.out_port(port_ind))
new_pad.in_port(port_ind + 1).get_connection().insert_node(
create_op_with_const_inputs(graph, Squeeze, {1: int64_array([0])}))
tfpad.out_port(0).get_connection().set_source(new_pad.out_port(0))
graph.remove_node(tfpad.id)
def test_two_inputs(self):
graph = build_graph(
self.node_attrs,
self.edge_attrs + [('pads_begin', 'pad'), ('pads_end', 'pad')],
nodes_with_edges_only=True,
)
pad_node = Node(graph, 'pad')
Pad.infer(pad_node)
self.assertTrue(
np.array_equal(
Node(graph, 'data_out').shape,
np.array([1, 3, 100 + 1 + 3, 200 + 2 + 4])))
def extract(node):
attrs = get_mxnet_layer_attrs(node.symbol_dict)
pads = np.array(list(attrs.tuple('pad_width', int, None)))
pads = pads.reshape([-1, 2])
value = attrs.float('constant_value', 0.0)
node_attrs = {
'pads': pads,
'mode': attrs.str('mode', None),
'fill_value': value,
}
Pad.update_node_stat(node, node_attrs)
return __class__.enabled
def test_two_inputs_and_pads(self):
graph = build_graph(
self.node_attrs,
self.edge_attrs + [('data_pads', 'pad')],
{
'pad': {
'pads':
np.array([[0, 0], [0, 0], [1, 3], [2, 4]], dtype=np.int64)
}
},
nodes_with_edges_only=True,
)
pad_node = Node(graph, 'pad')
with self.assertRaisesRegex(
AssertionError, ".*unexpected additional input argument.*"):
Pad.infer(pad_node)
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 replace_op(self, graph: Graph, node: Node):
# save the original node name to use it in the new Pad op instance
original_name = node.soft_get('name', node.id)
rename_node(node, original_name + '/TBR')
new_pad = Pad(graph, {
'mode': node.soft_get('mode', None)
}).create_node()
rename_node(new_pad, original_name)
node.in_port(0).get_connection().set_destination(new_pad.in_port(0))
if node.soft_get('mode') == 'constant':
# the input with fill value is an optional third input in ONNX
if not node.in_port(2).disconnected():
node.in_port(2).get_connection().set_destination(
new_pad.in_port(3))
else:
new_pad.in_port(3).connect(
Const(graph, {
'value': 0.0
}).create_node().out_port(0))
# convert ONNX representation of the pads as [2 * N] to MO representation: [N] and [N]
split_pads = create_op_with_const_inputs(graph, Split,
{1: int64_array(0)},
{'num_splits': 2})
node.in_port(1).get_connection().set_destination(split_pads.in_port(0))
split_pads.out_port(0).connect(new_pad.in_port(1))
split_pads.out_port(1).connect(new_pad.in_port(2))
return [new_pad.id]
def test_one_input(self):
graph = build_graph(
self.node_attrs,
self.edge_attrs,
{
'pad': {
'pads':
np.array([[0, 0], [0, 0], [1, 3], [2, 4]], dtype=np.int64)
}
},
nodes_with_edges_only=True,
)
pad_node = Node(graph, 'pad')
Pad.infer(pad_node)
self.assertTrue(
np.array_equal(
Node(graph, 'data_out').shape,
np.array([1, 3, 100 + 1 + 3, 200 + 2 + 4])))
def replace_op(self, graph: Graph, node: Node):
# Note that PyTorch paddings are reversed
pads_begin = node.module.pad[::2][::-1]
pads_end = node.module.pad[1::2][::-1]
pads_begin = Const(graph, {
'name': node.name + '/pads_begin',
'value': pads_begin
}).create_node()
pads_end = Const(graph, {
'name': node.name + '/pads_end',
'value': pads_end
}).create_node()
pad_value = Const(graph, {
'name': node.name + '/pad_value',
'value': 0.0
}).create_node()
pad = Pad(graph, dict(name=node.name)).create_node(
[node.in_node(0), pads_begin, pads_end, pad_value])
return [pad.id]
def extract(node):
Pad.update_node_stat(node, {'mode': node.pb.attr['mode'].s.decode('utf-8').lower()})
return __class__.enabled
def extract(node):
Pad.update_node_stat(node)
return __class__.enabled
def extract(cls, node):
Pad.update_node_stat(node)
return cls.enabled
def find_and_replace_pattern(self, graph: Graph):
for attr_pad in graph.get_op_nodes(op='AttributedPad'):
# save the original node name to use it in the new Pad op instance
original_name = attr_pad.soft_get('name', attr_pad.id)
new_pad = Pad(graph, {
'mode': attr_pad.soft_get('mode', None),
}).create_node()
rename_nodes([(attr_pad, original_name + '/to_be_removed'),
(new_pad, original_name)])
attr_pad.in_port(0).get_connection().set_destination(
new_pad.in_port(0))
new_pad.in_port(1).connect(
Const(graph, {
'value': attr_pad.pads[:, 0]
}).create_node().out_port(0))
new_pad.in_port(2).connect(
Const(graph, {
'value': attr_pad.pads[:, 1]
}).create_node().out_port(0))
if attr_pad.soft_get('mode') == 'constant':
# create Constant node of proper data type (equal to the data type of the Pad first input)
convert_pad_value = create_op_with_const_inputs(
graph, ConvertLike, {0: attr_pad.fill_value},
{'name': original_name + '/pad_value_convert'})
convert_pad_value.in_port(1).connect(
new_pad.in_port(0).get_source())
new_pad.in_port(3).connect(convert_pad_value.out_port(0))
attr_pad.out_port(0).get_connection().set_source(
new_pad.out_port(0))
graph.remove_node(attr_pad.id)