inference-engine/thirdparty/clDNN/src/lstm_gemm.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_gemm_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_gemm_type_id()
  26 {
  27     static primitive_type_base<lstm_gemm> instance;
  28     return &instance;
  29 }
  30
  31
  32 layout lstm_gemm_inst::calc_output_layout(lstm_gemm_node const& node)
  33 {
  34     assert((bool)node.get_primitive()->output_data_type == false
  35            && "Output data type forcing is not supported for lstm_gemm_node!");
  36     auto desc = node.get_primitive();
  37     auto input_layout = node.input().get_output_layout();
  38     auto weights_layout = node.weights().get_output_layout();
  39
  40     //   input{bfyx}     = [b: batch, f: sequence,   x: input_size,      y: 1]
  41     //   weights{bfyx}   = [b: 1,     f: direction,  x: 4 * hidden_size, y: input_size ]
  42     //   recurrent{bfyx} = [b: 1,     f: direction,  x: 4 * hidden_size, y: hidden_size ]
  43     //   biases{bfyx}    = [b: 1,     f:1 ,          x: direction,       y:  4 * hidden_size ]
  44     //   hidden{bfyx}    = [b: batch, f:  direction, x: 1 ,              y: hidden_size ] optional
  45     //   tempGEMM{bfyx}  = [b: batch, f: direction,  x: 4*hidden_size,   y: 1] output
  46     auto result = layout(input_layout.data_type, input_layout.format, tensor(input_layout.size.batch[0], weights_layout.size.feature[0], weights_layout.size.spatial[1], 1));
  47     return result;
  48 }
  49
  50 std::string lstm_gemm_inst::to_string(lstm_gemm_node const& node)
  51 {
  52     auto desc         = node.get_primitive();
  53     auto node_info    = node.desc_to_json();
  54     auto weights_id   = desc->weights;
  55     auto recurrent_id = desc->recurrent;
  56     auto bias_id      = desc->bias != "" ? desc->bias : "no bias";
  57     auto hidden_id    = desc->hidden != "" ? desc->hidden : "no inital hidden";
  58
  59     std::stringstream primitive_description;
  60
  61     json_composite lstm_gemm_info;
  62     lstm_gemm_info.add("weights id", weights_id);
  63     lstm_gemm_info.add("recurrent id", recurrent_id);
  64     lstm_gemm_info.add("bias id", bias_id);
  65     lstm_gemm_info.add("hidden id", hidden_id);
  66     node_info->add("lstm gemm info", lstm_gemm_info);
  67     node_info->dump(primitive_description);
  68
  69     return primitive_description.str();
  70 }
  71
  72 lstm_gemm_inst::typed_primitive_inst(network_impl& network, lstm_gemm_node const& node)
  73     :parent(network, node)
  74 {
  75     auto input_layout = node.input().get_output_layout();
  76     CLDNN_ERROR_NOT_PROPER_FORMAT(node.id(), "input format", input_layout.format.value, "expected format", format::bfyx, format::fyxb);
  77 }
  78 }