pad_op.cc

#include "caffe2/operators/pad_op.h"

#include <algorithm>

namespace caffe2 {

PadMode StringToPadMode(const string& mode) {
  if (mode == "constant") {
    return PadMode::CONSTANT;
  } else if (mode == "reflect") {
    return PadMode::REFLECT;
  } else if (mode == "edge") {
    return PadMode::EDGE;
  } else {
    CAFFE_THROW("Unknown padding mode: " + mode);
  }
}

using std::min;
using std::max;

template <>
bool PadImageOp<float, CPUContext>::RunOnDeviceWithOrderNCHW() {
  auto& X = Input(0);
  auto* Y = Output(0);
  int channels = X.dim32(1);
  int height = X.dim32(2);
  int width = X.dim32(3);
  ConvPoolOpBase::SetOutputSize(X, Y, channels);

  const float* Xdata = X.data<float>();
  float* Ydata = Y->mutable_data<float>();
  // The main loop
  int padded_height = Y->dim32(2);
  int padded_width = Y->dim32(3);

  switch (mode_) {
    case PadMode::CONSTANT:
      for (int n = 0; n < X.dim32(0); ++n) {
        for (int c = 0; c < channels; ++c) {
          for (int ph = 0; ph < padded_height; ++ph) {
            for (int pw = 0; pw < padded_width; ++pw) {
              int h = ph - pad_t();
              int w = pw - pad_l();
              Ydata[ph * padded_width + pw] =
                  (h < 0 || w < 0 || h >= height || w >= width)
                  ? value_
                  : Xdata[h * width + w];
            }
          }
          // Do offset.
          Xdata += height * width;
          Ydata += padded_height * padded_width;
        }
      }
      break;
    case PadMode::REFLECT:
      if (pad_r() >= 0 && pad_t() >= 0 && pad_l() >= 0 && pad_b() >= 0) {
        for (int n = 0; n < X.dim32(0); ++n) {
          for (int c = 0; c < channels; ++c) {
            // Handle the valid region:
            // i.e. Y[n][c][pad_t:pad_t+h][pad_l:pad_l+w]
            auto* Ystart = Ydata + pad_t() * padded_width + pad_l();
            math::CopyMatrix<CPUContext>(
                sizeof(float),
                height,
                width,
                Xdata,
                width,
                Ystart,
                padded_width,
                &context_);

// Fixup areas where we need to reflect
#define X(ph, pw)                 \
  int h = ph - pad_t();           \
  int w = pw - pad_l();           \
  h = max(h, -h);                 \
  h = min(h, 2 * height - h - 2); \
  w = max(w, -w);                 \
  w = min(w, 2 * width - w - 2);  \
  Ydata[ph * padded_width + pw] = Xdata[h * width + w]

            // Top part
            for (int ph = 0; ph < pad_t(); ++ph) {
              for (int pw = 0; pw < padded_width; ++pw) {
                X(ph, pw);
              }
            }

            // Bottom part
            for (int ph = padded_height - pad_b(); ph < padded_height; ++ph) {
              for (int pw = 0; pw < padded_width; ++pw) {
                X(ph, pw);
              }
            }

            // Interior
            for (int ph = pad_t(); ph < padded_height - pad_b(); ++ph) {
              // Left
              for (int pw = 0; pw < pad_l(); ++pw) {
                X(ph, pw);
              }
              // Right
              for (int pw = padded_width - pad_r(); pw < padded_width; ++pw) {
                X(ph, pw);
              }
            }
#undef X

            // Do offset.
            Xdata += height * width;
            Ydata += padded_height * padded_width;
          }
        }
      } else {
        for (int n = 0; n < X.dim32(0); ++n) {
          for (int c = 0; c < channels; ++c) {
            for (int ph = 0; ph < padded_height; ++ph) {
              for (int pw = 0; pw < padded_width; ++pw) {
                int h = ph - pad_t();
                int w = pw - pad_l();
                // max(h, -h) does reflection over 0
                h = max(h, -h);
                // min(h, 2 * height - h - 2) does reflection over height.
                h = min(h, 2 * height - h - 2);
                w = max(w, -w);
                w = min(w, 2 * width - w - 2);
                Ydata[ph * padded_width + pw] = Xdata[h * width + w];
              }
            }
            // Do offset.
            Xdata += height * width;
            Ydata += padded_height * padded_width;
          }
        }
      }
      break;
    case PadMode::EDGE:
      for (int n = 0; n < X.dim32(0); ++n) {
        for (int c = 0; c < channels; ++c) {
          for (int ph = 0; ph < padded_height; ++ph) {
            for (int pw = 0; pw < padded_width; ++pw) {
              // Bounds to the right range.
              int h = min(height - 1, max(ph - pad_t(), 0));
              int w = min(width - 1, max(pw - pad_l(), 0));
              Ydata[ph * padded_width + pw] = Xdata[h * width + w];
            }
          }
          // Do offset.
          Xdata += height * width;
          Ydata += padded_height * padded_width;
        }
      }
      break;
  }
  return true;
}

template <>
bool PadImageOp<float, CPUContext>::RunOnDeviceWithOrderNHWC() {
  auto& X = Input(0);
  auto* Y = Output(0);
  int height = X.dim32(1);
  int width = X.dim32(2);
  int channels = X.dim32(3);
  ConvPoolOpBase::SetOutputSize(X, Y, channels);
  const float* Xdata = X.data<float>();
  float* Ydata = Y->mutable_data<float>();

  // The main loop
  int padded_height = Y->dim32(1);
  int padded_width = Y->dim32(2);

  switch (mode_) {
    case PadMode::CONSTANT:
      for (int n = 0; n < X.dim32(0); ++n) {
        for (int ph = 0; ph < padded_height; ++ph) {
          for (int pw = 0; pw < padded_width; ++pw) {
            int h = ph - pad_t();
            int w = pw - pad_l();
            const int pad_index = (ph * padded_width + pw) * channels;
            if (h < 0 || w < 0 || h >= height || w >= width) {
              for (int c = 0; c < channels; ++c) {
                Ydata[pad_index + c] = value_;
              }
            } else {
              const int input_index = (h * width + w) * channels;
              for (int c = 0; c < channels; ++c) {
                Ydata[pad_index + c] = Xdata[input_index + c];
              }
            }
          }
        }
        // Do offset.
        Xdata += X.size() / X.dim32(0);
        Ydata += Y->size() / Y->dim32(0);
      }
      break;
    case PadMode::REFLECT:
      for (int n = 0; n < X.dim32(0); ++n) {
        for (int ph = 0; ph < padded_height; ++ph) {
          for (int pw = 0; pw < padded_width; ++pw) {
            const int pad_index = (ph * padded_width + pw) * channels;
            int h = ph - pad_t();
            int w = pw - pad_l();
            // max(h, -h) does reflection over 0
            h = max(h, -h);
            // min(h, 2 * height - h - 2) does reflection over height.
            h = min(h, 2 * height - h - 2);
            w = max(w, -w);
            w = min(w, 2 * width - w - 2);
            const int input_index = (h * width + w) * channels;
            for (int c = 0; c < channels; ++c) {
              Ydata[pad_index + c] = Xdata[input_index + c];
            }
          }
        }
        // Do offset.
        Xdata += X.size() / X.dim32(0);
        Ydata += Y->size() / Y->dim32(0);
      }
      break;
    case PadMode::EDGE:
      for (int n = 0; n < X.dim32(0); ++n) {
        for (int ph = 0; ph < padded_height; ++ph) {
          for (int pw = 0; pw < padded_width; ++pw) {
            const int pad_index = (ph * padded_width + pw) * channels;
            int h = min(height - 1, max(ph - pad_t(), 0));
            int w = min(width - 1, max(pw - pad_l(), 0));
            const int input_index = (h * width + w) * channels;
            for (int c = 0; c < channels; ++c) {
              Ydata[pad_index + c] = Xdata[input_index + c];
            }
          }
        }
        // Do offset.
        Xdata += X.size() / X.dim32(0);
        Ydata += Y->size() / Y->dim32(0);
      }
      break;
  }
  return true;
}

template <>
bool PadImageGradientOp<float, CPUContext>::RunOnDeviceWithOrderNCHW() {
  auto& dY = Input(0);
  auto* dX = Output(0);
  dX->Resize(
      dY.dim32(0),
      dY.dim32(1),
      dY.dim32(2) - pad_t() - pad_b(),
      dY.dim32(3) - pad_l() - pad_r());
  int padded_height = dY.dim32(2);
  int padded_width = dY.dim32(3);
  int channels = dX->dim32(1);
  int height = dX->dim32(2);
  int width = dX->dim32(3);

  const float* dYdata = dY.data<float>();
  float* dXdata = dX->mutable_data<float>();
  math::Set<float, CPUContext>(dX->size(), 0, dXdata, &context_);
  // The main loop
  switch (mode_) {
    case PadMode::CONSTANT:
      for (int n = 0; n < dY.dim32(0); ++n) {
        for (int c = 0; c < channels; ++c) {
          for (int ph = 0; ph < padded_height; ++ph) {
            for (int pw = 0; pw < padded_width; ++pw) {
              int h = ph - pad_t();
              int w = pw - pad_l();
              if (!(h < 0 || w < 0 || h >= height || w >= width)) {
                dXdata[h * width + w] += dYdata[ph * padded_width + pw];
              }
            }
          }
          // Do offset.
          dXdata += height * width;
          dYdata += padded_height * padded_width;
        }
      }
      break;
    case PadMode::REFLECT:
      for (int n = 0; n < dY.dim32(0); ++n) {
        for (int c = 0; c < channels; ++c) {
          for (int ph = 0; ph < padded_height; ++ph) {
            for (int pw = 0; pw < padded_width; ++pw) {
              int h = ph - pad_t();
              int w = pw - pad_l();
              // max(h, -h) does reflection over 0
              h = max(h, -h);
              // min(h, 2 * height - h - 2) does reflection over height.
              h = min(h, 2 * height - h - 2);
              w = max(w, -w);
              w = min(w, 2 * width - w - 2);
              dXdata[h * width + w] += dYdata[ph * padded_width + pw];
            }
          }
          // Do offset.
          dXdata += height * width;
          dYdata += padded_height * padded_width;
        }
      }
      break;
    case PadMode::EDGE:
      for (int n = 0; n < dY.dim32(0); ++n) {
        for (int c = 0; c < channels; ++c) {
          for (int ph = 0; ph < padded_height; ++ph) {
            for (int pw = 0; pw < padded_width; ++pw) {
              int h = min(height - 1, max(ph - pad_t(), 0));
              int w = min(width - 1, max(pw - pad_l(), 0));
              dXdata[h * width + w] += dYdata[ph * padded_width + pw];
            }
          }
          // Do offset.
          dXdata += height * width;
          dYdata += padded_height * padded_width;
        }
      }
      break;
  }
  return true;
}

template <>
bool PadImageGradientOp<float, CPUContext>::RunOnDeviceWithOrderNHWC() {
  auto& dY = Input(0);
  auto* dX = Output(0);
  dX->Resize(
      dY.dim32(0),
      dY.dim32(1) - pad_t() - pad_b(),
      dY.dim32(2) - pad_l() - pad_r(),
      dY.dim32(3));
  int padded_height = dY.dim32(1);
  int padded_width = dY.dim32(2);
  int channels = dY.dim32(3);
  int height = dX->dim32(1);
  int width = dX->dim32(2);

  const float* dYdata = dY.data<float>();
  float* dXdata = dX->mutable_data<float>();
  math::Set<float, CPUContext>(dX->size(), 0, dXdata, &context_);

  switch (mode_) {
    case PadMode::CONSTANT:
      for (int n = 0; n < dY.dim32(0); ++n) {
        for (int ph = 0; ph < padded_height; ++ph) {
          for (int pw = 0; pw < padded_width; ++pw) {
            int h = ph - pad_t();
            int w = pw - pad_l();
            const int pad_index = (ph * padded_width + pw) * channels;
            if (!(h < 0 || w < 0 || h >= height || w >= width)) {
              const int input_index = (h * width + w) * channels;
              for (int c = 0; c < channels; ++c) {
                dXdata[input_index + c] += dYdata[pad_index + c];
              }
            }
          }
        }
        // Do offset.
        dXdata += dX->size() / dX->dim32(0);
        dYdata += dY.size() / dY.dim32(0);
      }
      break;
    case PadMode::REFLECT:
      for (int n = 0; n < dY.dim32(0); ++n) {
        for (int ph = 0; ph < padded_height; ++ph) {
          for (int pw = 0; pw < padded_width; ++pw) {
            const int pad_index = (ph * padded_width + pw) * channels;
            int h = ph - pad_t();
            int w = pw - pad_l();
            // max(h, -h) does reflection over 0
            h = max(h, -h);
            // min(h, 2 * height - h - 2) does reflection over height.
            h = min(h, 2 * height - h - 2);
            w = max(w, -w);
            w = min(w, 2 * width - w - 2);
            const int input_index = (h * width + w) * channels;
            for (int c = 0; c < channels; ++c) {
              dXdata[input_index + c] += dYdata[pad_index + c];
            }
          }
        }
        // Do offset.
        dXdata += dX->size() / dX->dim32(0);
        dYdata += dY.size() / dY.dim32(0);
      }
      break;
    case PadMode::EDGE:
      for (int n = 0; n < dY.dim32(0); ++n) {
        for (int ph = 0; ph < padded_height; ++ph) {
          for (int pw = 0; pw < padded_width; ++pw) {
            const int pad_index = (ph * padded_width + pw) * channels;
            // Bounds to the right range.
            int h = min(height - 1, max(ph - pad_t(), 0));
            int w = min(width - 1, max(pw - pad_l(), 0));
            const int input_index = (h * width + w) * channels;
            for (int c = 0; c < channels; ++c) {
              dXdata[input_index + c] += dYdata[pad_index + c];
            }
          }
        }
        // Do offset.
        dXdata += dX->size() / dX->dim32(0);
        dYdata += dY.size() / dY.dim32(0);
      }
      break;
  }
  return true;
}

template <>
std::vector<TensorShape> PadImageOp<float, CPUContext>::PadTensorInference(
    const OperatorDef& def,
    const vector<TensorShape>& in) {
  return ConvPoolOpBase::TensorInferenceForPool(def, in);
}

REGISTER_CPU_OPERATOR(PadImage, PadImageOp<float, CPUContext>);
REGISTER_CPU_OPERATOR(PadImageGradient, PadImageGradientOp<float, CPUContext>);

OPERATOR_SCHEMA(PadImage)
    .NumInputs(1)
    .NumOutputs(1)
    .TensorInferenceFunction(PadImageOp<float, CPUContext>::PadTensorInference)
    .SetDoc(R"DOC(
PadImage pads values around the boundary of an image according to the pad
values and stride sizes defined by the ConvPoolOpBase operator.
  )DOC")
    .Input(
        0,
        "X",
        "Input data tensor from the previous operator; dimensions "
        "depend on whether the NCHW or NHWC operators are being used. For example, "
        "in the former, the input has size (N x C x H x W), where N is the batch "
        "size, C is the number of channels, and H and W are the height and the width "
        "of the data. The corresponding permutation of dimensions is used in the "
        "latter case. ")
    .Output(
        0,
        "Y",
        "Output data tensor from padding the H and W dimensions on "
        "the tensor. Dimensions will vary based on various pad and stride "
        "sizes.");

OPERATOR_SCHEMA(PadImageGradient).NumInputs(1).NumOutputs(1);

class GetPadImageGradient : public GradientMakerBase {
  using GradientMakerBase::GradientMakerBase;
  vector<OperatorDef> GetGradientDefs() override {
    return SingleGradientDef(
        "PadImageGradient", "", vector<string>{GO(0)}, vector<string>{GI(0)});
  }
};
REGISTER_GRADIENT(PadImage, GetPadImageGradient);

} // namespace caffe2