Files
openvino/model-optimizer/extensions/front/broadcast_with_range.py

86 lines
3.7 KiB
Python

"""
Copyright (C) 2018-2020 Intel Corporation
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
import numpy as np
from extensions.ops.gather import Gather
from extensions.ops.range import Range
from mo.front.common.partial_infer.utils import int64_array
from mo.front.common.replacement import FrontReplacementSubgraph
from mo.front.tf.graph_utils import create_op_with_const_inputs, create_op_node_with_second_input
from mo.graph.graph import Graph, rename_nodes, Node
from mo.ops.unsqueeze import Unsqueeze
class ExpandRangeConstant(FrontReplacementSubgraph):
"""
Searches for Constant operations filled with range values starting from 0 and replaces it with Range operation
Faced in ONNX BERT -- replacing it makes model reshape-able by sequence length
WARNING: true BIDIRECTIONAL mode of Broadcast could cause issues
(the probability is small, so we decided to keep the optimization)
value_input[1, X] (value=range(0,X)) shape_input[Y, 1]
\ /
Broadcast(mode='bidirectional') [Y, X]
"""
enabled = True
def find_and_replace_pattern(self, graph: Graph):
for node in graph.get_op_nodes(type='Broadcast'):
value = node.in_port(0).get_source().node
if value.soft_get('type') == 'Const':
self.replace(node, value)
@staticmethod
def replace(node: Node, const: Node):
graph = node.graph
shape = const.shape
const_name = const.soft_get('name', const.id)
non_one_dims = np.argwhere(shape != 1).flatten()
one_dims = np.argwhere(shape == 1).flatten()
if not (non_one_dims.size == 1 and 5 < np.prod(shape) < 500):
# (5;500) range is deduced to affect less models
return
value = const.value
if not np.array_equal(np.arange(0, np.prod(shape), 1).reshape(shape), value):
return
positive_idx = non_one_dims.item(0)
negative_idx = positive_idx - len(shape)
gather = create_op_with_const_inputs(graph, Gather, {1: int64_array(negative_idx), 2: int64_array(0)},
{'name': node.soft_get('name', node.id) + '/BroadcastingDim'})
range_node = create_op_with_const_inputs(graph, Range,
{0: np.array(0, dtype=value.dtype),
2: np.array(1, dtype=value.dtype)},
{'name': const_name + '/Range', 'dtype': value.dtype})
node.in_port(1).get_connection().add_destination(gather.in_port(0))
gather.out_port(0).connect(range_node.in_port(1))
node.in_port(0).get_connection().set_source(range_node.out_port(0))
if one_dims.size:
unsqueeze = create_op_node_with_second_input(graph, Unsqueeze, one_dims,
{'name': const_name + '/KeepShape'})
range_node.out_port(0).get_connection().insert_node(unsqueeze)
rename_nodes([(const, const_name + '/ToBeDeleted'), (unsqueeze, const_name)])
else:
rename_nodes([(const, const_name + '/ToBeDeleted'), (range_node, const_name)])