inference-engine/thirdparty/clDNN/src/lstm.cpp

   1 /*
   2 // Copyright (c) 2016 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 ///////////////////////////////////////////////////////////////////////////////////////////////////
  18 #include "lstm_inst.h"
  19 #include "primitive_type_base.h"
  20 #include "error_handler.h"
  21 #include "json_object.h"
  22
  23 namespace cldnn
  24 {
  25 primitive_type_id lstm_type_id()
  26 {
  27     static primitive_type_base<lstm> instance;
  28     return &instance;
  29 }
  30
  31
  32 layout lstm_inst::calc_output_layout(lstm_node const& node)
  33 {
  34     assert((bool)node.get_primitive()->output_data_type == false
  35            && "Output data type forcing is not supported for lstm_node!");
  36     auto input_layout = node.input().get_output_layout();
  37     auto hidden_layout = node.inital_hidden().get_output_layout();
  38
  39     // input     = [ batch,  sequence,       direction,      input_size ]
  40     // weights   = [     1, direction, 4 * hidden_size,      input_size ]
  41     // recurrent = [     1, direction, 4 * hidden_size,     hidden_size ]
  42     // biases    = [     1,         1,       direction, 4 * hidden_size ]
  43     // hidden    = [ batch,         1,       direction,     hidden_size ]
  44     // cell      = [ batch,         1,       direction,     hidden_size ]
  45     // output    = [ batch,  sequence,       direction,     hidden_size ]
  46         auto result = layout(input_layout.data_type, format::bfyx,
  47                   tensor(hidden_layout.size.feature[0], input_layout.size.feature[0],
  48                          hidden_layout.size.spatial[0], hidden_layout.size.spatial[1]));
  49     return result;
  50 }
  51
  52 std::string lstm_inst::to_string(lstm_node const& node)
  53 {
  54     auto desc         = node.get_primitive();
  55     auto node_info    = node.desc_to_json();
  56     auto weights_id   = desc->weights;
  57     auto recurrent_id = desc->recurrent;
  58     auto bias_id      = desc->bias != "" ? desc->bias : "no bias";
  59     auto peepholes_id = desc->peepholes != "" ? desc->peepholes : "no peepholes";
  60     auto initial_hidden_id = desc->initial_hidden != "" ? desc->initial_hidden : "no inital hidden";
  61     auto initial_cell_id = desc->initial_cell != "" ? desc->initial_cell : "no initial cell";
  62
  63     std::stringstream primitive_description;
  64
  65     json_composite lstm_info;
  66     lstm_info.add("weights id", weights_id);
  67     lstm_info.add("recurrent id", recurrent_id);
  68     lstm_info.add("bias id", bias_id);
  69     lstm_info.add("peepholes id", peepholes_id);
  70     lstm_info.add("initial_hidden id", initial_hidden_id);
  71     lstm_info.add("initial_cell id", initial_cell_id);
  72     node_info->add("lstm info", lstm_info);
  73     node_info->dump(primitive_description);
  74
  75     return primitive_description.str();
  76 }
  77
  78 lstm_inst::typed_primitive_inst(network_impl& network, lstm_node const& node)
  79     :parent(network, node)
  80 {
  81     auto input_layout = node.input().get_output_layout();
  82     CLDNN_ERROR_NOT_PROPER_FORMAT(node.id(), "input format", input_layout.format.value, "expected format", format::bfyx);
  83 }
  84
  85 }