model-optimizer/extensions/middle/Reduce_test.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 unittest
  18
  19 import numpy as np
  20
  21 from extensions.middle.Reduce import ReduceReplacer
  22 from mo.middle.passes.eliminate_test import build_graph
  23 from mo.middle.passes.fusing.fuse_linear_ops_test import compare_graphs
  24
  25 # The dictionary with nodes attributes used to build various graphs. A key is the name of the node and the value is the
  26 # dictionary with node attributes.
  27 nodes_attributes = {
  28     # Placeholder layers
  29     'placeholder_1': {'shape': None, 'type': 'Placeholder', 'kind': 'op', 'op': 'Placeholder'},
  30     'placeholder_1_data': {'value': None, 'shape': None, 'kind': 'data', 'data_type': None},
  31     'placeholder_2_data': {'value': None, 'shape': None, 'kind': 'data', 'data_type': None},
  32     'placeholder_3_data': {'value': None, 'shape': None, 'kind': 'data', 'data_type': None},
  33     'placeholder_4_data': {'value': None, 'shape': None, 'kind': 'data', 'data_type': None},
  34
  35     # Reshape layers
  36     'reduce_1': {'type': 'Reduce', 'kind': 'op', 'op': 'Reduce'},
  37     'reduce_1_data': {'value': None, 'shape': None, 'kind': 'data'},
  38
  39     # Reshape layers
  40     'reshape_1': {'type': 'Reshape', 'kind': 'op', 'op': 'Reshape'},
  41     'reshape_1_data': {'value': None, 'shape': None, 'kind': 'data'},
  42
  43     'reshape_2': {'type': 'Reshape', 'kind': 'op', 'op': 'Reshape'},
  44     'reshape_2_data': {'value': None, 'shape': None, 'kind': 'data'},
  45
  46     # Pooling
  47     'pooling': {'type': 'Pooling', 'kind': 'op', 'op': 'Pooling'},
  48     'pooling_data': {'value': None, 'shape': None, 'kind': 'data'},
  49
  50     # Power
  51     'power': {'type': 'Power', 'kind': 'op', 'op': 'Power'},
  52     'power_data': {'value': None, 'shape': None, 'kind': 'data'},
  53
  54     # Concat
  55     'concat': {'type': 'Concat', 'kind': 'op', 'op': 'Concat'},
  56 }
  57
  58
  59 class ReduceReplacerTest(unittest.TestCase):
  60     def test1(self):
  61         #   Original graph
  62         #   data(1,64,1)-->Reduce(axis=1,keep_dims=True)-->data(1,1,1)
  63         #
  64         #   Reference graph
  65         #   data(1,61,1)->Reshape(1,1,64,1)->Pool(1,1,1,1)->Reshape(1,1,1)
  66         #
  67         graph = build_graph(nodes_attributes,
  68                             [('placeholder_1_data', 'reduce_1'),
  69                              ('reduce_1', 'reduce_1_data'),
  70                              ('reduce_1_data', 'concat'),
  71                              ],
  72                             {'placeholder_1_data': {'shape': np.array([1, 64, 1])},
  73                              'reduce_1': {'axis': np.array([1]), 'keep_dims': True, 'reduce_type': 'Mean'},
  74                              'reduce_1_data': {'shape': np.array([1, 1, 1])},
  75                              }, nodes_with_edges_only=True)
  76
  77         graph.graph['layout'] = 'NCHW'
  78
  79         graph_ref = build_graph(nodes_attributes,
  80                                 [('placeholder_1_data', 'reshape_1'),
  81                                  ('reshape_1', 'reshape_1_data'),
  82                                  ('reshape_1_data', 'pooling'),
  83                                  ('pooling', 'pooling_data'),
  84                                  ('pooling_data', 'reshape_2'),
  85                                  ('reshape_2', 'reshape_2_data'),
  86                                  ('reshape_2_data', 'concat'),
  87                                  ],
  88                                 {'placeholder_1_data': {'shape': np.array([1, 64, 1])},
  89                                  'reshape_1': {'dim': np.array([1, 1, 64, 1])},
  90                                  'reshape_1_data': {'shape': np.array([1, 1, 64, 1])},
  91                                  'pooling': {'window': np.array([1, 1, 64, 1])},
  92                                  'pooling_data': {'shape': np.array([1, 1, 1, 1])},
  93                                  'reshape_2': {'dim': np.array([1, 1, 1])},
  94                                  'reshape_2_data': {'shape': np.array([1, 1, 1])},
  95                                  }, nodes_with_edges_only=True)
  96
  97         pattern = ReduceReplacer()
  98         pattern.find_and_replace_pattern(graph)
  99
 100         (flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
 101         self.assertTrue(flag, resp)
 102
 103     def test2(self):
 104         #   Original graph
 105         #   data(1,3,64,64)-->Reduce(axis=2,keep_dims=True)-->data(1,3,1,64)
 106         #
 107         #   Reference graph
 108         #   data(1,3,64,64)->Reshape->Pool(1,3,1,64)->Reshape(1,3,1,64)
 109         #
 110         graph = build_graph(nodes_attributes,
 111                             [('placeholder_1_data', 'reduce_1'),
 112                              ('reduce_1', 'reduce_1_data'),
 113                              ('reduce_1_data', 'concat'),
 114                              ],
 115                             {'placeholder_1_data': {'shape': np.array([1, 3, 64, 64])},
 116                              'reduce_1': {'axis': np.array([2]), 'keep_dims': True, 'reduce_type': 'Mean'},
 117                              'reduce_1_data': {'shape': np.array([1, 3, 1, 64])},
 118                              }, nodes_with_edges_only=True)
 119
 120         graph.graph['layout'] = 'NCHW'
 121
 122         graph_ref = build_graph(nodes_attributes,
 123                                 [('placeholder_1_data', 'reshape_1'),
 124                                  ('reshape_1', 'reshape_1_data'),
 125                                  ('reshape_1_data', 'pooling'),
 126                                  ('pooling', 'pooling_data'),
 127                                  ('pooling_data', 'reshape_2'),
 128                                  ('reshape_2', 'reshape_2_data'),
 129                                  ('reshape_2_data', 'concat'),
 130                                  ],
 131                                 {'placeholder_1_data': {'shape': np.array([1, 3, 64, 64])},
 132                                  'reshape_1': {'dim': np.array([1, 3, 64, 64])},
 133                                  'reshape_1_data': {'shape': np.array([1, 3, 64, 64])},
 134                                  'pooling': {'window': np.array([1, 1, 64, 1])},
 135                                  'pooling_data': {'shape': np.array([1, 3, 1, 64])},
 136                                  'reshape_2': {'dim': np.array([1, 3, 1, 64])},
 137                                  'reshape_2_data': {'shape': np.array([1, 3, 1, 64])},
 138                                  }, nodes_with_edges_only=True)
 139
 140         pattern = ReduceReplacer()
 141         pattern.find_and_replace_pattern(graph)
 142
 143         (flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
 144         self.assertTrue(flag, resp)
 145
 146     def test3(self):
 147         #   Original graph
 148         #   data(1,3,64,64)-->Reduce(axis=[2,3],keep_dims=True)-->data(1,3,1,1)
 149         #
 150         #   Reference graph
 151         #   data(1,3,64,64)->Reshape->Pool(1,3,1,1)->Reshape(1,3,1,1)
 152         #
 153         graph = build_graph(nodes_attributes,
 154                             [('placeholder_1_data', 'reduce_1'),
 155                              ('reduce_1', 'reduce_1_data'),
 156                              ('reduce_1_data', 'concat'),
 157                              ],
 158                             {'placeholder_1_data': {'shape': np.array([1, 3, 64, 64])},
 159                              'reduce_1': {'axis': np.array([2, 3]), 'keep_dims': True, 'reduce_type': 'Mean'},
 160                              'reduce_1_data': {'shape': np.array([1, 3, 1, 1])},
 161                              }, nodes_with_edges_only=True)
 162
 163         graph.graph['layout'] = 'NCHW'
 164
 165         graph_ref = build_graph(nodes_attributes,
 166                                 [('placeholder_1_data', 'reshape_1'),
 167                                  ('reshape_1', 'reshape_1_data'),
 168                                  ('reshape_1_data', 'pooling'),
 169                                  ('pooling', 'pooling_data'),
 170                                  ('pooling_data', 'reshape_2'),
 171                                  ('reshape_2', 'reshape_2_data'),
 172                                  ('reshape_2_data', 'concat'),
 173                                  ],
 174                                 {'placeholder_1_data': {'shape': np.array([1, 3, 64, 64])},
 175                                  'reshape_1': {'dim': np.array([1, 3, 64 * 64, 1])},
 176                                  'reshape_1_data': {'shape': np.array([1, 3, 64 * 64, 1])},
 177                                  'pooling': {'window': np.array([1, 1, 64 * 64, 1])},
 178                                  'pooling_data': {'shape': np.array([1, 3, 1, 1])},
 179                                  'reshape_2': {'dim': np.array([1, 3, 1, 1])},
 180                                  'reshape_2_data': {'shape': np.array([1, 3, 1, 1])},
 181                                  }, nodes_with_edges_only=True)
 182
 183         pattern = ReduceReplacer()
 184         pattern.find_and_replace_pattern(graph)
 185
 186         (flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
 187         self.assertTrue(flag, resp)
 188
 189     def test4(self):
 190         #   Original graph
 191         #   data(2,3,64,64)-->Reduce(axis=[1,2,3],keep_dims=False)-->data(2)
 192         #
 193         #   Reference graph
 194         #   data(2,3,64,64)->Reshape(2,1,3*64*64,1)->Pool(2,1,1,1)->Reshape(2)
 195         #
 196         graph = build_graph(nodes_attributes,
 197                             [('placeholder_1_data', 'reduce_1'),
 198                              ('reduce_1', 'reduce_1_data'),
 199                              ('reduce_1_data', 'concat'),
 200                              ],
 201                             {'placeholder_1_data': {'shape': np.array([2, 3, 64, 64])},
 202                              'reduce_1': {'axis': np.array([1, 2, 3]), 'keep_dims': False, 'reduce_type': 'Mean'},
 203                              'reduce_1_data': {'shape': np.array([2])},
 204                              }, nodes_with_edges_only=True)
 205
 206         graph.graph['layout'] = 'NCHW'
 207
 208         graph_ref = build_graph(nodes_attributes,
 209                                 [('placeholder_1_data', 'reshape_1'),
 210                                  ('reshape_1', 'reshape_1_data'),
 211                                  ('reshape_1_data', 'pooling'),
 212                                  ('pooling', 'pooling_data'),
 213                                  ('pooling_data', 'reshape_2'),
 214                                  ('reshape_2', 'reshape_2_data'),
 215                                  ('reshape_2_data', 'concat'),
 216                                  ],
 217                                 {'placeholder_1_data': {'shape': np.array([2, 3, 64, 64])},
 218                                  'reshape_1': {'dim': np.array([2, 1, 3 * 64 * 64, 1])},
 219                                  'reshape_1_data': {'shape': np.array([2, 1, 3 * 64 * 64, 1])},
 220                                  'pooling': {'window': np.array([1, 1, 3 * 64 * 64, 1])},
 221                                  'pooling_data': {'shape': np.array([2, 1, 1, 1])},
 222                                  'reshape_2': {'dim': np.array([2])},
 223                                  'reshape_2_data': {'shape': np.array([2])},
 224                                  }, nodes_with_edges_only=True)
 225
 226         pattern = ReduceReplacer()
 227         pattern.find_and_replace_pattern(graph)
 228
 229         (flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
 230         self.assertTrue(flag, resp)
 231
 232     def test5(self):
 233         #   Original graph
 234         #   data(1, 16, 64, 64, 64, 4)-->Reduce(axis=[5],keep_dims=False)-->data(1, 16, 64, 64, 64)
 235         #
 236         #   Reference graph
 237         #   data(1, 16, 64, 64, 64, 4)->Reshape(1*16*64*64, 64, 4, 1)->Pool(1, 1, 4, 1)->Reshape(1, 16, 64, 64, 64)
 238         #
 239         graph = build_graph(nodes_attributes,
 240                             [('placeholder_1_data', 'reduce_1'),
 241                              ('reduce_1', 'reduce_1_data'),
 242                              ('reduce_1_data', 'concat'),
 243                              ],
 244                             {'placeholder_1_data': {'shape': np.array([1, 16, 64, 64, 64, 4])},
 245                              'reduce_1': {'axis': np.array([5]), 'keep_dims': False, 'reduce_type': 'max'},
 246                              'reduce_1_data': {'shape': np.array([1, 16, 64, 64, 64])},
 247                              }, nodes_with_edges_only=True)
 248
 249         graph.graph['layout'] = 'NCHW'
 250
 251         graph_ref = build_graph(nodes_attributes,
 252                                 [('placeholder_1_data', 'reshape_1'),
 253                                  ('reshape_1', 'reshape_1_data'),
 254                                  ('reshape_1_data', 'pooling'),
 255                                  ('pooling', 'pooling_data'),
 256                                  ('pooling_data', 'reshape_2'),
 257                                  ('reshape_2', 'reshape_2_data'),
 258                                  ('reshape_2_data', 'concat'),
 259                                  ],
 260                                 {'placeholder_1_data': {'shape': np.array([1, 16, 64, 64, 64, 4])},
 261                                  'reshape_1': {'dim': np.array([65536, 64, 4, 1])},
 262                                  'reshape_1_data': {'shape': np.array([65536, 64, 4, 1])},
 263                                  'pooling': {'window': np.array([1, 1, 4, 1])},
 264                                  'pooling_data': {'shape': np.array([65536, 64, 1, 1])},
 265                                  'reshape_2': {'dim': np.array([1, 16, 64, 64, 64])},
 266                                  'reshape_2_data': {'shape': np.array([1, 16, 64, 64, 64])},
 267                                  }, nodes_with_edges_only=True)
 268
 269         pattern = ReduceReplacer()
 270         pattern.find_and_replace_pattern(graph)
 271
 272         (flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
 273         self.assertTrue(flag, resp)
 274
 275     def test6(self):
 276         #   Original graph
 277         #   data(1,64,1)-->Reduce(axis=-2,keep_dims=True, reduce_type=Sum)-->data(1,1,1)
 278         #
 279         #   Reference graph
 280         #   data(1,61,1)->Reshape(1,1,64,1)->Pool(1,1,1,1)->Reshape(1,1,1)->Power(scale=64)
 281         #
 282         graph = build_graph(nodes_attributes,
 283                             [('placeholder_1_data', 'reduce_1'),
 284                              ('reduce_1', 'reduce_1_data'),
 285                              ('reduce_1_data', 'concat'),
 286                              ],
 287                             {'placeholder_1_data': {'shape': np.array([1, 64, 1])},
 288                              'reduce_1': {'axis': np.array([-2]), 'keep_dims': True, 'reduce_type': 'Sum'},
 289                              'reduce_1_data': {'shape': np.array([1, 1, 1])},
 290                              }, nodes_with_edges_only=True)
 291
 292         graph.graph['layout'] = 'NCHW'
 293
 294         graph_ref = build_graph(nodes_attributes,
 295                                 [('placeholder_1_data', 'reshape_1'),
 296                                  ('reshape_1', 'reshape_1_data'),
 297                                  ('reshape_1_data', 'pooling'),
 298                                  ('pooling', 'pooling_data'),
 299                                  ('pooling_data', 'reshape_2'),
 300                                  ('reshape_2', 'reshape_2_data'),
 301                                  ('reshape_2_data', 'power'),
 302                                  ('power', 'power_data'),
 303                                  ('power_data', 'concat'),
 304                                  ],
 305                                 {'placeholder_1_data': {'shape': np.array([1, 64, 1])},
 306                                  'reshape_1': {'dim': np.array([1, 1, 64, 1])},
 307                                  'reshape_1_data': {'shape': np.array([1, 1, 64, 1])},
 308                                  'pooling': {'window': np.array([1, 1, 64, 1])},
 309                                  'pooling_data': {'shape': np.array([1, 1, 1, 1])},
 310                                  'reshape_2': {'dim': np.array([1, 1, 1])},
 311                                  'reshape_2_data': {'shape': np.array([1, 1, 1])},
 312                                  'power': {'scale': 64.0},
 313                                  'power_data': {'shape': np.array([1, 1, 1])},
 314                                  }, nodes_with_edges_only=True)
 315
 316         pattern = ReduceReplacer()
 317         pattern.find_and_replace_pattern(graph)
 318
 319         (flag, resp) = compare_graphs(graph, graph_ref, 'concat', check_op_attrs=True)
 320         self.assertTrue(flag, resp)