inference-engine/src/inference_engine/shape_infer/built-in/ie_bin_conv_shape_infer.hpp

   1 // Copyright (C) 2019 Intel Corporation
   2 // SPDX-License-Identifier: Apache-2.0
   3 //
   4
   5 #pragma once
   6
   7 #include <description_buffer.hpp>
   8 #include "ie_built_in_impl.hpp"
   9 #include <map>
  10 #include <memory>
  11 #include <string>
  12 #include <vector>
  13 #include <debug.h>
  14 #include <cmath>
  15 #include <ie_format_parser.h>
  16
  17 namespace InferenceEngine {
  18 namespace ShapeInfer {
  19
  20 /**
  21  *@brief Implementation of Shape inference for BinaryConvolution layer
  22  */
  23 class BinConvShapeProp : public BuiltInShapeInferImpl {
  24 public:
  25     explicit BinConvShapeProp(const std::string& type) : BuiltInShapeInferImpl(type) {}
  26
  27     void inferShapesImpl(const std::vector<Blob::CPtr>& inBlobs,
  28                          const std::map<std::string, std::string>& params,
  29                          const std::map<std::string, Blob::Ptr>& blobs,
  30                          std::vector<SizeVector>& outShapes) override {
  31         LayerParams lp{};
  32         BinaryConvolutionLayer binConvLayer(lp);
  33         binConvLayer.params = params;
  34         binConvLayer.type = _type;
  35         validate(&binConvLayer, inBlobs, params, blobs);
  36
  37         auto dims = inShapes[0];
  38         auto computeSpatialShape = [&](size_t inDim, int axis) {
  39             size_t kernel = 0;
  40             if (binConvLayer._dilation[axis])
  41                 kernel = (binConvLayer._kernel[axis] - 1) * binConvLayer._dilation[axis] + 1;
  42             else
  43                 kernel = binConvLayer._kernel[axis];
  44             size_t stride = binConvLayer._stride[axis];
  45             size_t pad = binConvLayer._padding[axis];
  46
  47             float outDim;
  48             std::string padType = binConvLayer._auto_pad;
  49             if (padType == "valid") {
  50                 outDim = std::ceil((inDim - kernel + 1.f) / stride);
  51             } else if (padType == "same_upper") {
  52                 outDim = std::ceil(1.f * inDim / stride);
  53             } else if (padType == "same_lower") {
  54                 outDim = std::floor(1.f * inDim / stride);
  55             } else {
  56                 int padEnd = binConvLayer._pads_end[axis];
  57                 outDim = std::floor(1.f * (inDim + pad + padEnd - kernel) / stride) + 1.f;
  58             }
  59
  60             if (outDim < 0)
  61                 THROW_IE_EXCEPTION << "New shapes " << details::dumpVec(dims) << " make output shape negative";
  62
  63             return static_cast<size_t>(outDim);
  64         };
  65
  66         size_t inputN = dims[0];
  67         size_t OC = binConvLayer._out_depth;
  68         SizeVector shapes;
  69         shapes.push_back(inputN);
  70         shapes.push_back(OC);
  71         if (dims.size() == 5)
  72             shapes.push_back(computeSpatialShape(dims[dims.size() - 3], Z_AXIS));
  73         shapes.push_back(computeSpatialShape(dims[dims.size() - 2], Y_AXIS));
  74         shapes.push_back(computeSpatialShape(dims[dims.size() - 1], X_AXIS));
  75         outShapes.push_back(shapes);
  76     }
  77 };
  78
  79 }  // namespace ShapeInfer
  80 }  // namespace InferenceEngine