ngraph/test/runtime/pass/opset1_upgrade.cpp

   1 //*****************************************************************************
   2 // Copyright 2017-2020 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 #include "opset1_upgrade.hpp"
  17
  18 #include <functional>
  19 #include <iterator>
  20 #include <limits>
  21 #include <numeric>
  22
  23 #include "ngraph/builder/autobroadcast.hpp"
  24 #include "ngraph/builder/reshape.hpp"
  25 #include "ngraph/graph_util.hpp"
  26 #include "ngraph/op/util/op_types.hpp"
  27 #include "ngraph/ops.hpp"
  28 #include "ngraph/provenance.hpp"
  29 #include "op/avg_pool.hpp"
  30 #include "op/convolution.hpp"
  31 #include "op/group_conv.hpp"
  32
  33 using namespace std;
  34 using namespace ngraph;
  35
  36 namespace
  37 {
  38     template <typename OpV0, typename OpV1>
  39     shared_ptr<Node> op_cast_binary_elementwise_node(const shared_ptr<OpV0>& node)
  40     {
  41         const auto autob = node->get_autob();
  42         auto replacement_node =
  43             make_shared<OpV1>(node->input_value(0), node->input_value(1), autob);
  44         replace_node(node, replacement_node);
  45         return replacement_node;
  46     }
  47
  48     // Default is that we didn nothing
  49     shared_ptr<Node> op_cast(shared_ptr<Node> node) { return nullptr; }
  50     shared_ptr<Node> op_cast(shared_ptr<op::Add> node)
  51     {
  52         return op_cast_binary_elementwise_node<op::v0::Add, op::v1::Add>(node);
  53     }
  54
  55     shared_ptr<Node> op_cast(shared_ptr<op::Broadcast> node)
  56     {
  57         auto replacement_node = ngraph::builder::opset1::make_broadcast(
  58             node->input_value(0), node->get_broadcast_shape(), node->get_broadcast_axes());
  59         replace_node(node, replacement_node.get_node_shared_ptr());
  60         return replacement_node.get_node_shared_ptr();
  61     }
  62
  63     shared_ptr<Node> op_cast(shared_ptr<op::BroadcastLike> node) { return nullptr; }
  64     shared_ptr<Node> op_cast(shared_ptr<op::v0::Convolution> node)
  65     {
  66         auto strides = node->get_window_movement_strides();
  67         auto dilations = node->get_window_dilation_strides();
  68         auto pads_begin = node->get_padding_below();
  69         auto pads_end = node->get_padding_above();
  70         auto data_dilation_strides = node->get_data_dilation_strides();
  71         auto auto_pad = node->get_pad_type();
  72
  73         bool is_dds_valid = all_of(data_dilation_strides.begin(),
  74                                    data_dilation_strides.end(),
  75                                    [](size_t value) { return value == 1; });
  76
  77         NGRAPH_CHECK(is_dds_valid,
  78                      "Unable to convert Convolution:0 to Convolution:1 with data dilation strides "
  79                      "other than `1`. Node: ",
  80                      *node);
  81
  82         auto replacement_node = make_shared<op::v1::Convolution>(node->input_value(0),
  83                                                                  node->input_value(1),
  84                                                                  strides,
  85                                                                  pads_begin,
  86                                                                  pads_end,
  87                                                                  dilations,
  88                                                                  auto_pad);
  89         replace_node(node, replacement_node);
  90         return replacement_node;
  91     }
  92
  93     shared_ptr<Node> op_cast(shared_ptr<op::v0::ConvolutionBackpropData> node)
  94     {
  95         auto data_batch_shape = node->get_data_batch_shape();
  96         auto strides = node->get_window_movement_strides_forward();
  97         auto dilations = node->get_window_dilation_strides_forward();
  98         auto pads_begin = node->get_padding_below_forward();
  99         auto pads_end = node->get_padding_above_forward();
 100         auto data_dilation_strides = node->get_data_dilation_strides_forward();
 101
 102         bool is_dds_valid = all_of(data_dilation_strides.begin(),
 103                                    data_dilation_strides.end(),
 104                                    [](size_t value) { return value == 1; });
 105
 106         NGRAPH_CHECK(is_dds_valid,
 107                      "Unable to convert ConvolutionBackpropData:0 to ConvolutionBackpropData:1 "
 108                      "with data dilation strides "
 109                      "other than `1`. Node: ",
 110                      *node);
 111
 112         auto replacement_node = make_shared<op::v1::ConvolutionBackpropData>(
 113             node->input_value(1), // data
 114             node->input_value(0), // filters
 115             op::Constant::create(
 116                 element::i64,
 117                 Shape{data_batch_shape.size() - 2},
 118                 vector<size_t>(data_batch_shape.begin() + 2, data_batch_shape.end())),
 119             strides,
 120             pads_begin,
 121             pads_end,
 122             dilations);
 123         replace_node(node, replacement_node);
 124         return replacement_node;
 125     }
 126
 127     shared_ptr<Node> op_cast(shared_ptr<op::Divide> node)
 128     {
 129         const auto autob = node->get_autob();
 130         const bool pydiv = node->is_pythondiv();
 131         auto replacement_node =
 132             make_shared<op::v1::Divide>(node->input_value(0), node->input_value(1), pydiv, autob);
 133         replace_node(node, replacement_node);
 134         return replacement_node;
 135     }
 136
 137     shared_ptr<Node> op_cast(shared_ptr<op::Reshape> node)
 138     {
 139         shared_ptr<Node> replacement_node =
 140             builder::opset1::reshape(node->input_value(0), node->get_reshape_output_shape());
 141         replace_node(node, replacement_node);
 142         return replacement_node;
 143     }
 144
 145     shared_ptr<Node> op_cast(shared_ptr<op::Equal> node)
 146     {
 147         return op_cast_binary_elementwise_node<op::v0::Equal, op::v1::Equal>(node);
 148     }
 149
 150     shared_ptr<Node> op_cast(shared_ptr<op::Gather> node)
 151     {
 152         int64_t axis = node->get_axis();
 153
 154         auto axis_node = make_shared<op::Constant>(element::i64, Shape{}, vector<int64_t>{axis});
 155         auto replacement_node =
 156             make_shared<op::v1::Gather>(node->input_value(0), node->input_value(1), axis_node);
 157         replace_node(node, replacement_node);
 158         return replacement_node;
 159     }
 160
 161     shared_ptr<Node> op_cast(shared_ptr<op::Greater> node)
 162     {
 163         return op_cast_binary_elementwise_node<op::v0::Greater, op::v1::Greater>(node);
 164     }
 165
 166     shared_ptr<Node> op_cast(shared_ptr<op::GreaterEq> node)
 167     {
 168         return op_cast_binary_elementwise_node<op::v0::GreaterEq, op::v1::GreaterEqual>(node);
 169     }
 170
 171     shared_ptr<Node> op_cast(shared_ptr<op::v0::GroupConvolution> node)
 172     {
 173         auto strides = node->get_window_movement_strides();
 174         auto dilations = node->get_window_dilation_strides();
 175         auto pads_begin = node->get_padding_below();
 176         auto pads_end = node->get_padding_above();
 177         auto data_dilation_strides = node->get_data_dilation_strides();
 178         auto auto_pad = node->get_pad_type();
 179
 180         bool is_dds_valid = all_of(data_dilation_strides.begin(),
 181                                    data_dilation_strides.end(),
 182                                    [](size_t value) { return value == 1; });
 183
 184         NGRAPH_CHECK(is_dds_valid,
 185                      "Unable to convert GroupConvolution:0 to GroupConvolution:1"
 186                      "with data dilation strides other than `1`. Node: ",
 187                      *node);
 188
 189         shared_ptr<Node> replacement_node;
 190         if (node->has_groups_in_filters())
 191         {
 192             replacement_node = make_shared<op::v1::GroupConvolution>(node->input_value(0),
 193                                                                      node->input_value(1),
 194                                                                      strides,
 195                                                                      pads_begin,
 196                                                                      pads_end,
 197                                                                      dilations,
 198                                                                      auto_pad);
 199         }
 200         else
 201         {
 202             NGRAPH_CHECK(node->get_input_partial_shape(1).is_static(),
 203                          "Unable to convert GroupConvolution:0 to GroupConvolution:1"
 204                          "with dynamic filters shape. Node: ",
 205                          *node);
 206
 207             auto filters_shape = node->get_input_shape(1);
 208             auto groups = node->get_groups();
 209             filters_shape[0] /= groups;
 210             filters_shape.insert(filters_shape.begin(), groups);
 211
 212             auto reshaped_filters = builder::opset1::reshape(node->input_value(1), filters_shape);
 213
 214             replacement_node = make_shared<op::v1::GroupConvolution>(node->input_value(0),
 215                                                                      reshaped_filters,
 216                                                                      strides,
 217                                                                      pads_begin,
 218                                                                      pads_end,
 219                                                                      dilations,
 220                                                                      auto_pad);
 221         }
 222         replace_node(node, replacement_node);
 223         return replacement_node;
 224     }
 225
 226     shared_ptr<Node> op_cast(shared_ptr<op::v0::GroupConvolutionBackpropData> node)
 227     {
 228         const auto strides = node->get_window_movement_strides();
 229         const auto dilations = node->get_window_dilation_strides();
 230         const auto pads_begin = node->get_padding_below();
 231         const auto pads_end = node->get_padding_above();
 232
 233         const auto data_batch_pshape = node->get_input_partial_shape(0);
 234         const auto filters_pshape = node->get_input_partial_shape(1);
 235
 236         NGRAPH_CHECK(data_batch_pshape.is_static(),
 237                      "Unable to convert GroupConvolutionBackpropData:0 to "
 238                      "GroupConvolutionBackpropData:1 with dynamic data_batch shape. Node: ",
 239                      *node);
 240         NGRAPH_CHECK(filters_pshape.is_static(),
 241                      "Unable to convert GroupConvolutionBackpropData:0 to "
 242                      "GroupConvolutionBackpropData:1 with dynamic filters shape. Node: ",
 243                      *node);
 244
 245         auto data_batch_shape = data_batch_pshape.to_shape();
 246         // Remove N, C from output shape to preserve only spatial dimentions.
 247         data_batch_shape.erase(std::begin(data_batch_shape),
 248                                std::next(std::begin(data_batch_shape), 2));
 249         auto filters_shape = filters_pshape.to_shape();
 250         auto groups = node->get_groups();
 251
 252         filters_shape[0] /= groups;
 253         filters_shape.insert(filters_shape.begin(), groups);
 254         auto reshaped_filters = builder::opset1::reshape(node->input_value(1), filters_shape);
 255
 256         auto replacement_node = make_shared<op::v1::GroupConvolutionBackpropData>(
 257             node->input_value(2),
 258             reshaped_filters,
 259             op::Constant::create(element::i64, Shape{data_batch_shape.size()}, data_batch_shape),
 260             strides,
 261             pads_begin,
 262             pads_end,
 263             dilations);
 264         replace_node(node, replacement_node);
 265         return replacement_node;
 266     }
 267
 268     shared_ptr<Node> op_cast(shared_ptr<op::Less> node)
 269     {
 270         return op_cast_binary_elementwise_node<op::v0::Less, op::v1::Less>(node);
 271     }
 272
 273     shared_ptr<Node> op_cast(shared_ptr<op::LessEq> node)
 274     {
 275         return op_cast_binary_elementwise_node<op::v0::LessEq, op::v1::LessEqual>(node);
 276     }
 277
 278     shared_ptr<Node> op_cast(shared_ptr<op::Max> node)
 279     {
 280         bool keep_dims = false;
 281         auto replacement_node =
 282             make_shared<op::v1::ReduceMax>(node->input_value(0), node->input_value(1), keep_dims);
 283         replace_node(node, replacement_node);
 284         return replacement_node;
 285     }
 286
 287     shared_ptr<Node> op_cast(shared_ptr<op::Maximum> node)
 288     {
 289         return op_cast_binary_elementwise_node<op::v0::Maximum, op::v1::Maximum>(node);
 290     }
 291
 292     shared_ptr<Node> op_cast(shared_ptr<op::Min> node)
 293     {
 294         bool keep_dims = false;
 295         auto replacement_node =
 296             make_shared<op::v1::ReduceMin>(node->input_value(0), node->input_value(1), keep_dims);
 297         replace_node(node, replacement_node);
 298         return replacement_node;
 299     }
 300
 301     shared_ptr<Node> op_cast(shared_ptr<op::Minimum> node)
 302     {
 303         return op_cast_binary_elementwise_node<op::v0::Minimum, op::v1::Minimum>(node);
 304     }
 305
 306     shared_ptr<Node> op_cast(shared_ptr<op::Multiply> node)
 307     {
 308         return op_cast_binary_elementwise_node<op::v0::Multiply, op::v1::Multiply>(node);
 309     }
 310
 311     shared_ptr<Node> op_cast(shared_ptr<op::Not> node)
 312     {
 313         auto replacement_node = make_shared<op::v1::LogicalNot>(node->input_value(0));
 314         replace_node(node, replacement_node);
 315         return replacement_node;
 316     }
 317
 318     shared_ptr<Node> op_cast(shared_ptr<op::NotEqual> node)
 319     {
 320         return op_cast_binary_elementwise_node<op::v0::NotEqual, op::v1::NotEqual>(node);
 321     }
 322
 323     shared_ptr<Node> op_cast(shared_ptr<op::OneHot> node)
 324     {
 325         const auto indices = node->input_value(0).get_node_shared_ptr();
 326         const auto one_hot_axis = node->get_one_hot_axis();
 327
 328         const auto output_pshape = node->get_output_partial_shape(0);
 329         NGRAPH_CHECK(output_pshape[one_hot_axis].is_static(),
 330                      "OneHot:v0 one hot axis dimension must be static ",
 331                      *node);
 332         const auto depth = output_pshape[one_hot_axis].get_length();
 333         const auto depth_node = op::Constant::create(element::i64, Shape{}, {depth});
 334
 335         const auto on_value = op::Constant::create(element::i64, Shape{}, {1});
 336         const auto off_value = op::Constant::create(element::i64, Shape{}, {0});
 337
 338         auto replacement_node =
 339             make_shared<op::v1::OneHot>(indices, depth_node, on_value, off_value, one_hot_axis);
 340         replace_node(node, replacement_node);
 341         return replacement_node;
 342     }
 343
 344     shared_ptr<Node> op_cast(shared_ptr<op::Or> node)
 345     {
 346         return op_cast_binary_elementwise_node<op::v0::Or, op::v1::LogicalOr>(node);
 347     }
 348
 349     shared_ptr<Node> op_cast(shared_ptr<op::Power> node)
 350     {
 351         return op_cast_binary_elementwise_node<op::v0::Power, op::v1::Power>(node);
 352     }
 353
 354     shared_ptr<Node> op_cast(shared_ptr<op::Product> node)
 355     {
 356         bool keep_dims = false;
 357         auto replacement_node =
 358             make_shared<op::v1::ReduceProd>(node->input_value(0), node->input_value(1), keep_dims);
 359         replace_node(node, replacement_node);
 360         return replacement_node;
 361     }
 362
 363     shared_ptr<Node> op_cast(shared_ptr<op::Reverse> node)
 364     {
 365         // creates a Constant node from the v0::Reverse reversed_axes attribute
 366         // and uses it as the second input of v1::Reverse
 367         const auto reversed_axes = node->get_reversed_axes();
 368
 369         const auto reversed_axes_constant = op::Constant::create(
 370             element::i64, Shape{reversed_axes.size()}, reversed_axes.to_vector());
 371
 372         const auto replacement_node = make_shared<op::v1::Reverse>(
 373             node->input_value(0), reversed_axes_constant, op::v1::Reverse::Mode::INDEX);
 374
 375         replace_node(node, replacement_node);
 376         return replacement_node;
 377     }
 378
 379     shared_ptr<Node> op_cast(shared_ptr<op::Select> node)
 380     {
 381         auto replacement_node = make_shared<op::v1::Select>(node->input_value(0),
 382                                                             node->input_value(1),
 383                                                             node->input_value(2),
 384                                                             op::AutoBroadcastSpec());
 385         replace_node(node, replacement_node);
 386         return replacement_node;
 387     }
 388
 389     shared_ptr<Node> op_cast(shared_ptr<op::Softmax> node)
 390     {
 391         NGRAPH_CHECK(op::is_constant(node->input_value(1).get_node()),
 392                      "axes parameter is expected to be a static constant");
 393
 394         AxisSet axes = node->get_axes();
 395
 396         NGRAPH_CHECK(
 397             axes.size() == 1,
 398             "Unable to convert Softmax:0 to Softmax:1 with zero or more than one axis. Node: ",
 399             *node);
 400
 401         auto replacement_node =
 402             make_shared<op::v1::Softmax>(node->input_value(0), axes.to_vector()[0]);
 403         replace_node(node, replacement_node);
 404         return replacement_node;
 405     }
 406
 407     shared_ptr<Node> op_cast(shared_ptr<op::Slice> node)
 408     {
 409         const auto data = node->input_value(0);
 410         const auto begin = op::Constant::create(
 411             element::i64, Shape{node->get_lower_bounds().size()}, node->get_lower_bounds());
 412         const auto end = op::Constant::create(
 413             element::i64, Shape{node->get_upper_bounds().size()}, node->get_upper_bounds());
 414         const auto strides = op::Constant::create(
 415             element::i64, Shape{node->get_strides().size()}, node->get_strides());
 416         int64_t input_size = node->get_lower_bounds().size();
 417
 418         auto replacement_node = make_shared<op::v1::StridedSlice>(data,
 419                                                                   begin,
 420                                                                   end,
 421                                                                   strides,
 422                                                                   vector<int64_t>(input_size, 0),
 423                                                                   vector<int64_t>(input_size, 0));
 424
 425         replace_node(node, replacement_node);
 426         return replacement_node;
 427     }
 428
 429     shared_ptr<Node> op_cast(shared_ptr<op::Split> node)
 430     {
 431         const auto& splits_vec = node->get_splits();
 432         const auto first_elem = splits_vec.front();
 433
 434         const bool split_evenly =
 435             std::all_of(splits_vec.begin(), splits_vec.end(), [first_elem](const size_t split) {
 436                 return split == first_elem;
 437             });
 438
 439         std::shared_ptr<Node> replacement_node;
 440         if (split_evenly)
 441         {
 442             replacement_node = make_shared<op::v1::Split>(
 443                 node->input_value(0), node->input_value(1), splits_vec.front());
 444         }
 445         else
 446         {
 447             const auto split_lengths =
 448                 ngraph::op::Constant::create(element::u64, Shape{splits_vec.size()}, splits_vec);
 449
 450             replacement_node = make_shared<op::v1::VariadicSplit>(
 451                 node->input_value(0), node->input_value(1), split_lengths);
 452         }
 453
 454         replace_node(node, replacement_node);
 455         return replacement_node;
 456     }
 457
 458     shared_ptr<Node> op_cast(shared_ptr<op::Subtract> node)
 459     {
 460         return op_cast_binary_elementwise_node<op::v0::Subtract, op::v1::Subtract>(node);
 461     }
 462
 463     shared_ptr<Node> op_cast(shared_ptr<op::Sum> node)
 464     {
 465         bool keep_dims = false;
 466         auto replacement_node =
 467             make_shared<op::v1::ReduceSum>(node->input_value(0), node->input_value(1), keep_dims);
 468         replace_node(node, replacement_node);
 469         return replacement_node;
 470     }
 471
 472     shared_ptr<Node> op_cast(shared_ptr<op::TopK> node)
 473     {
 474         NGRAPH_CHECK(op::is_constant(node->input_value(1).get_node()),
 475                      "parameter k is expected to be a static constant");
 476         NGRAPH_CHECK(op::is_constant(node->input_value(2).get_node()),
 477                      "parameter top_k_axis is expected to be a static constant");
 478
 479         const auto k = node->get_k();
 480         const auto axis = node->get_top_k_axis();
 481
 482         std::string sort;
 483         switch (node->get_sort())
 484         {
 485         case op::TopK::SortType::SORT_INDICES: sort = "index"; break;
 486         case op::TopK::SortType::SORT_VALUES: sort = "value"; break;
 487         case op::TopK::SortType::NONE: sort = "none"; break;
 488         }
 489
 490         std::string mode;
 491         if (node->get_compute_max())
 492         {
 493             mode = "max";
 494         }
 495         else
 496         {
 497             mode = "min";
 498         }
 499
 500         const auto k_constant = op::Constant::create(element::i64, Shape{}, {k});
 501         auto replacement_node =
 502             make_shared<op::v1::TopK>(node->input_value(0), k_constant, axis, mode, sort);
 503
 504         // indices output will be 0, values 1
 505         vector<int64_t> output_order{1, 0};
 506         replace_node(node, replacement_node, output_order);
 507         return replacement_node;
 508     }
 509
 510     shared_ptr<Node> op_cast(shared_ptr<op::Xor> node)
 511     {
 512         auto replacement_node = make_shared<op::v1::LogicalXor>(
 513             node->input_value(0), node->input_value(1), node->get_autob());
 514         replace_node(node, replacement_node);
 515         return replacement_node;
 516     }
 517
 518     using DispatchMap = map<NodeTypeInfo, std::function<bool(shared_ptr<Node> node)>>;
 519
 520     template <typename T>
 521     bool op_cast_thunk(shared_ptr<Node> node)
 522     {
 523         auto upgraded_node = op_cast(as_type_ptr<T>(node));
 524         if (upgraded_node)
 525         {
 526             if (ngraph::get_provenance_enabled())
 527             {
 528                 const std::string provenance_tag =
 529                     "<Opset1_Upgrade (v0 " + std::string(node->get_type_name()) + ")>";
 530                 upgraded_node->add_provenance_tags_above(node->input_values(), {provenance_tag});
 531             }
 532             return true;
 533         }
 534         return false;
 535     }
 536
 537     DispatchMap& get_dispatch_map()
 538     {
 539         static DispatchMap dispatch_map{
 540 #define NGRAPH_OP(NAME, NAMESPACE) {NAMESPACE::NAME::type_info, op_cast_thunk<NAMESPACE::NAME>},
 541 #include "opset0_tbl.hpp"
 542 #undef NGRAPH_OP
 543         };
 544         return dispatch_map;
 545     }
 546 } // namespace
 547
 548 bool pass::Opset1Upgrade::run_on_node(shared_ptr<Node> node)
 549 {
 550     bool modified = false;
 551     auto& dispatch_map = get_dispatch_map();
 552     auto it = dispatch_map.find(node->get_type_info());
 553     if (it != dispatch_map.end())
 554     {
 555         modified = it->second(node);
 556     }
 557     return modified;
 558 }