model-optimizer/extensions/middle/NormalizeFullyConnected.py

   1 """
   2  Copyright (c) 2018-2019 Intel Corporation
   3
   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
   7
   8       http://www.apache.org/licenses/LICENSE-2.0
   9
  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.
  15 """
  16
  17 import numpy as np
  18
  19 from mo.graph.graph import Graph
  20 from mo.middle.replacement import MiddleReplacementPattern
  21 from mo.ops.op import Op
  22 from mo.ops.reshape import Reshape
  23
  24
  25 class NormalizeFullyConnected(MiddleReplacementPattern):
  26     enabled = True
  27     graph_condition = [lambda graph: graph.graph['fw'] == 'onnx']
  28
  29     def run_after(self):
  30         from extensions.middle.GemmToFullyConnected import GemmToFullyConnected
  31         return [GemmToFullyConnected]
  32
  33     def run_before(self):
  34         from extensions.middle.pass_separator import MiddleFinish
  35         return [MiddleFinish]
  36
  37     def pattern(self):
  38         return dict(
  39             nodes=[
  40                 ('fc', dict(kind='op', type='FullyConnected')),
  41                 ('fc_output', dict(kind='data'))],
  42             edges=[('fc', 'fc_output')],
  43         )
  44
  45     def replace_pattern(self, graph: Graph, match: dict):
  46         """
  47             This pass normalize FC layer
  48             Example:
  49
  50             (2,16,512)-->FC->(2,16,101)    =>    (2,16,512)-->Reshape-->(32,512)-->FC-->(32,101)-->Reshape-->(2,16,101)
  51
  52         """
  53         fc = match['fc']
  54         fc_weights = fc.in_node(1)
  55         fc_output = match['fc_output']
  56         fc_input = fc.in_node()
  57
  58         input_shape = fc.in_node().shape
  59         if len(input_shape) <= 2 or np.prod(fc_input.shape[1:]) == fc_weights.shape[fc_weights.input_channel_dim]:
  60             return
  61
  62         # Insert Reshape to normalize input for FC layer that should be in [N,C] layout
  63         first_reshape_shape = np.array([np.prod(input_shape[0:-1]), input_shape[-1]], dtype=np.int64)
  64         second_reshape_shape = np.array([*input_shape[0:-1], fc['out-size']], dtype=np.int64)
  65         fc_out_shape = np.array([np.prod(input_shape[0:-1]), fc['out-size']], dtype=np.int64)
  66         first_reshape = Reshape(graph, {'dim': np.array(first_reshape_shape)})
  67         second_reshape = Reshape(graph, {'dim': np.array(second_reshape_shape)})
  68
  69         input_edge_attrs = graph.get_edge_data(fc_input.id, fc.id)[0]
  70         output_edge_attrs = graph.get_edge_data(fc.id, fc_output.id)[0]
  71
  72         graph.remove_edge(fc_input.id, fc.id)
  73         graph.remove_edge(fc.id, fc_output.id)
  74
  75         # Insert Reshapes before and after FullyConnected layer
  76         reshape_data = first_reshape.create_node_with_data(inputs=[fc_input])
  77         graph.add_edge(reshape_data.id, fc.id, **input_edge_attrs)
  78
  79         new_fc_output = Op.create_data_node(graph, fc, {'shape': fc_out_shape}, edge_attrs=output_edge_attrs)
  80
  81         second_reshape.create_node_with_data(inputs=[new_fc_output], data_nodes=fc_output)