2 Copyright (c) 2018-2019 Intel Corporation
4 Licensed under the Apache License, Version 2.0 (the "License");
5 you may not use this file except in compliance with the License.
6 You may obtain a copy of the License at
8 http://www.apache.org/licenses/LICENSE-2.0
10 Unless required by applicable law or agreed to in writing, software
11 distributed under the License is distributed on an "AS IS" BASIS,
12 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 See the License for the specific language governing permissions and
14 limitations under the License.
17 from copy import deepcopy
21 from mo.front.common.layout import get_features_dim, shape_for_layout
22 from mo.graph.graph import Node, Graph
23 from mo.middle.passes.fusing.helpers import get_value_id
24 from mo.middle.replacement import MiddleReplacementPattern
25 from mo.ops.op import Op
26 from mo.ops.reshape import Reshape
29 class Eltwise1DInputReshape(MiddleReplacementPattern):
31 Inserts Reshape before 1-D input to Eltwise if another input of Eltwise is multi-dimensional tensor with the
32 same feature size as 1-D input
34 Replacer is useful in cases of layout change in MO (for example NHWC-> NCHW translation of TensorFlow models)
37 Eltwise Mul operation in TF multiplies Tensors by feature dimension with shapes [1,375,500,24] and [24].
38 After layout change in MO Eltwise Mul have input shapes [1,24,375,500] and [24]. It is a problem (500!=24).
39 We have to insert Reshape layer for Tensor with [24] shape to correspond the feature dimension of
40 Tensor [1,24,375,500] shape
42 change of graph.graph['layout'] may cause an issue
43 change in re-layout function: convert_nhwc_to_nchw(graph) may cause an issue
48 return [EltwiseInputReshape]
50 def find_and_replace_pattern(self, graph: Graph):
51 layout = graph.graph['layout']
52 for n in list(graph.nodes()):
53 if 'type' in graph.node[n] and graph.node[n]['type'] == 'Eltwise' and get_value_id(Node(graph, n)) is None:
54 eltwise_op_node = Node(graph, n)
55 out_shape = eltwise_op_node.out_node().shape
56 if 4 <= len(out_shape) <= 5:
57 out_features = out_shape[get_features_dim(layout, len(out_shape))]
58 for port, node in eltwise_op_node.in_nodes().items():
59 if len(node.shape) != len(out_shape) and len(node.shape) == 1 and out_features == node.shape[0]:
60 in_atts = deepcopy(graph.get_edge_data(node.id, n)[0])
61 graph.remove_edge(node.id, n)
62 new_shape = shape_for_layout(layout, batch=1, features=out_features, height=1, width=1,
63 depth=1 if len(out_shape) == 5 else None)
64 reshape_data_op = Reshape(graph, attrs={'dim': new_shape, 'name': node.id + '/Broadcast'})
65 reshape_data_node = reshape_data_op.create_node_with_data([node])
66 graph.add_edge(reshape_data_node.id, eltwise_op_node.id, **in_atts)
69 class EltwiseInputReshape(MiddleReplacementPattern):
73 from extensions.middle.pass_separator import MiddleStart
76 def find_and_replace_pattern(self, graph: Graph):
77 data_nodes = [Node(graph, node) for node in graph.node if Node(graph, node).kind == 'data']
78 for node in data_nodes:
79 # Get all requested shapes for current node
80 # This mapping will contain pairs like {shape:[list of consumers nodes]}
82 for consumer in node.out_nodes():
83 edge_attrs = graph.get_edge_data(node.id, consumer.id)[0]
84 if 'new_shape' in edge_attrs:
85 if np.array_equal(edge_attrs['new_shape'], node.shape):
87 new_shape = tuple([x for x in edge_attrs['new_shape']])
88 if not new_shape in mapping:
89 mapping.update({new_shape: [consumer]})
91 mapping[new_shape].append(consumer)
93 if node.has_valid('value'):
94 # Check that requested shape are the same
95 # In case if they are different, we duplicate them
96 for shape_key in mapping.keys():
97 shape = list(shape_key)
98 new_value = np.reshape(node.value, shape)
99 node_copy = Op.create_input_data_node(graph, node.id + '/copy', value=np.array(new_value))
100 for consumer in mapping[shape_key]:
101 edge_attrs = graph.get_edge_data(node.id, consumer.id)[0]
102 del edge_attrs['new_shape']
104 # Remove edge from previous data node and connect new data node with its consumer
105 graph.remove_edge(node.id, consumer.id)
106 graph.add_edge(node_copy.id, consumer.id, **edge_attrs)
108 # Insert Reshape layer between data node and consumer
109 for shape_key in mapping.keys():
110 shape = list(shape_key)
111 reshape = Reshape(graph, attrs={'dim': shape, 'name': 'EltwiseReshapeNormalization'})
112 reshape_data = reshape.create_node_with_data(inputs=[node])
114 # Iterate over consumers and reconnect them to Reshape layer output
115 for consumer in mapping[shape_key]:
116 edge_attrs = graph.get_edge_data(node.id, consumer.id)[0]
117 del edge_attrs['new_shape']
119 # Reconnect edge from original data node to Reshape output datanode
120 graph.remove_edge(node.id, consumer.id)
121 graph.add_edge(reshape_data.id, consumer.id, **edge_attrs)