caffe2/operators/distance_op.cc

#include "caffe2/operators/distance_op.h"

namespace caffe2 {

template<>
bool SquaredL2DistanceOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(0);
  auto& Y = Input(1);
  auto* distance = Output(0);
  CAFFE_ENFORCE_EQ(X.ndim(), Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE_EQ(X.dim32(i), Y.dim32(i));
  }
  int N = X.ndim() > 0 ? X.dim32(0) : 1;
  distance->Resize(N);
  int D = N > 0 ? X.size() / N : 0;
  float* distance_data = distance->mutable_data<float>();
  const float* X_data = X.data<float>();
  const float* Y_data = Y.data<float>();
  for (int i = 0; i < N; ++i) {
    float Xscale, Yscale, cross;
    math::Dot<float, CPUContext>(
        D, X_data + i * D, X_data + i * D, &Xscale, &context_);
    math::Dot<float, CPUContext>(
        D, Y_data + i * D, Y_data + i * D, &Yscale, &context_);
    math::Dot<float, CPUContext>(
        D, X_data + i * D, Y_data + i * D, &cross, &context_);
    distance_data[i] = (Xscale + Yscale) * 0.5 - cross;
  }
  return true;
}

template <>
bool L1DistanceOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(0);
  auto& Y = Input(1);
  auto* distance = Output(0);
  CAFFE_ENFORCE_EQ(X.ndim(), Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE_EQ(X.dim32(i), Y.dim32(i));
  }
  int N = X.ndim() > 0 ? X.dim32(0) : 1;
  distance->Resize(N);
  int D = N > 0 ? X.size() / N : 0;

  const float* X_data = X.data<float>();
  const float* Y_data = Y.data<float>();

  for (int i = 0; i < N; ++i) {
    (distance->mutable_data<float>())[i] =
        (ConstEigenVectorMap<float>(X_data + i * D, D).array() -
         ConstEigenVectorMap<float>(Y_data + i * D, D).array())
            .abs()
            .sum();
  }
  return true;
}

template <>
bool L1DistanceGradientOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(0);
  auto& Y = Input(1);
  auto& dDistance = Input(2);
  auto* dX = Output(0);
  auto* dY = Output(1);
  CAFFE_ENFORCE_EQ(X.ndim(), Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE_EQ(X.dim32(i), Y.dim32(i));
  }
  int N = X.ndim() > 0 ? X.dim32(0) : 1;
  int D = N > 0 ? X.size() / N : 0;
  CAFFE_ENFORCE(X.ndim() == Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE(X.dim32(i) == Y.dim32(i));
  }
  CAFFE_ENFORCE(dDistance.ndim() == 1);
  CAFFE_ENFORCE(dDistance.dim32(0) == N);
  dX->ResizeLike(X);
  dY->ResizeLike(Y);

  for (int i = 0; i < N; ++i) {
    auto offset = i * D;
    for (int j = 0; j < D; ++j) {
      const float temp =
          (X.data<float>())[offset + j] - (Y.data<float>())[offset + j];
      const float kEps = 1e-12f;
      if (temp < -kEps) {
        dX->mutable_data<float>()[offset + j] = -(dDistance.data<float>())[i];
        dY->mutable_data<float>()[offset + j] = (dDistance.data<float>())[i];
      } else if (temp > kEps) {
        dX->mutable_data<float>()[offset + j] = (dDistance.data<float>())[i];
        dY->mutable_data<float>()[offset + j] = -(dDistance.data<float>())[i];
      } else {
        dX->mutable_data<float>()[offset + j] = 0;
        dY->mutable_data<float>()[offset + j] = 0;
      }
    }
  }
  return true;
}

template <>
bool CosineSimilarityOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(X_IN);
  auto& Y = Input(Y_IN);
  auto* result = Output(COS_OUT);
  CAFFE_ENFORCE_EQ(X.ndim(), Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE_EQ(X.dim32(i), Y.dim32(i));
  }
  const int N = X.ndim() > 0 ? X.dim32(0) : 1;
  const int D = X.size_from_dim(1);
  result->Resize(N);
  float* result_data = result->mutable_data<float>();
  const float* X_data = X.data<float>();
  const float* Y_data = Y.data<float>();
  float X2, Y2;
  const float kEps = 1e-12f;
  for (int i = 0; i < N; ++i) { // TODO: multithreading
    auto offset = i * D;
    math::Dot<float, CPUContext>(
        D, X_data + offset, X_data + offset, &X2, &context_);
    math::Dot<float, CPUContext>(
        D, Y_data + offset, Y_data + offset, &Y2, &context_);
    math::Dot<float, CPUContext>(
        D, X_data + offset, Y_data + offset, result_data + i, &context_);
    result_data[i] /= std::sqrt(std::max(X2, kEps) * std::max(Y2, kEps));
  }
  return true;
}

template <>
bool CosineSimilarityGradientOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(X_IN);
  auto& Y = Input(Y_IN);
  auto& dCos = Input(DER_COS_IN);
  auto* dX = Output(DER_X_OUT);
  auto* dY = Output(DER_Y_OUT);
  const int N = X.ndim() > 0 ? X.dim32(0) : 1;
  const int D = X.size_from_dim(1);
  CAFFE_ENFORCE(X.ndim() == Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE(X.dim32(i) == Y.dim32(i));
  }
  CAFFE_ENFORCE(dCos.ndim() == 1);
  CAFFE_ENFORCE(dCos.dim32(0) == N);
  dX->ResizeLike(X);
  dY->ResizeLike(Y);

  const auto* X_data = X.template data<float>();
  const auto* Y_data = Y.template data<float>();
  const auto* dCos_data = dCos.template data<float>();
  auto* dX_data = dX->template mutable_data<float>();
  auto* dY_data = dY->template mutable_data<float>();
  float XN, YN, XY;
  const float kEps = 1e-12f;
  for (int i = 0; i < N; ++i) { // TODO: multithreading
    auto offset = i * D;

    // TODO: cache these result from the forward pass
    // ||x||
    math::Dot<float, CPUContext>(
        D, X_data + offset, X_data + offset, &XN, &context_);
    XN = std::sqrt(std::max(XN, kEps));
    // ||y||
    math::Dot<float, CPUContext>(
        D, Y_data + offset, Y_data + offset, &YN, &context_);
    YN = std::sqrt(std::max(YN, kEps));
    // ||x|| * || y ||
    float XYN = XN * YN;
    // x^Ty
    math::Dot<float, CPUContext>(
        D, X_data + offset, Y_data + offset, &XY, &context_);

    math::Scale<float, CPUContext>(
        D, dCos_data[i] / XYN, Y_data + offset, dX_data + offset, &context_);
    math::Axpy(
        D,
        -dCos_data[i] * XY / (XN * XN * XYN),
        X_data + offset,
        dX_data + offset,
        &context_);

    math::Scale<float, CPUContext>(
        D, dCos_data[i] / XYN, X_data + offset, dY_data + offset, &context_);
    math::Axpy(
        D,
        -dCos_data[i] * XY / (YN * YN * XYN),
        Y_data + offset,
        dY_data + offset,
        &context_);
  }

  return true;
}

template <>
bool DotProductOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(X_IN);
  auto& Y = Input(Y_IN);
  auto* result = Output(DOT_OUT);
  CAFFE_ENFORCE_EQ(X.ndim(), Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE_EQ(X.dim32(i), Y.dim32(i), "dimension at ", i);
  }
  int N, D;
  if (X.size() > 0) {
    N = X.ndim() > 0 ? X.dim32(0) : 1;
    D = X.size() / N;
  } else {
    N = 0;
    D = 0;
  }
  result->Resize(N);
  float* result_data = result->template mutable_data<float>();
  const float* X_data = X.template data<float>();
  const float* Y_data = Y.template data<float>();
  for (int i = 0; i < N; ++i) { // TODO: multithreading
    auto offset = i * D;
    math::Dot<float, CPUContext>(
        D, X_data + offset, Y_data + offset, result_data + i, &context_);
  }
  return true;
}

OpSchema::Cost CostInferenceForDotProduct(
    const OperatorDef& def,
    const vector<TensorShape>& in) {
  struct OpSchema::Cost c = PointwiseCostInference<1>(def, in);
  c.params_bytes = 0;
  return c;
}

template <>
bool DotProductGradientOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(X_IN);
  auto& Y = Input(Y_IN);
  auto& dDot = Input(DER_DOT_IN);
  auto* dX = Output(DER_X_OUT);
  auto* dY = Output(DER_Y_OUT);
  int N, D;
  if (X.size() > 0) {
    N = X.ndim() > 0 ? X.dim32(0) : 1;
    D = X.size() / N;
  } else {
    N = 0;
    D = 0;
  }
  CAFFE_ENFORCE(X.ndim() == Y.ndim());
  for (int i = 0; i < X.ndim(); ++i) {
    CAFFE_ENFORCE(X.dim32(i) == Y.dim32(i));
  }
  CAFFE_ENFORCE(dDot.ndim() == 1);
  CAFFE_ENFORCE(dDot.dim32(0) == N);
  dX->ResizeLike(X);
  dY->ResizeLike(Y);

  const auto* X_data = X.template data<float>();
  const auto* Y_data = Y.template data<float>();
  const auto* dDot_data = dDot.template data<float>();
  auto* dX_data = dX->template mutable_data<float>();
  auto* dY_data = dY->template mutable_data<float>();
  for (int i = 0; i < N; ++i) { // TODO: multithreading
    auto offset = i * D;
    math::Scale<float, CPUContext>(
        D, dDot_data[i], X_data + offset, dY_data + offset, &context_);
    math::Scale<float, CPUContext>(
        D, dDot_data[i], Y_data + offset, dX_data + offset, &context_);
  }
  return true;
}

template <>
bool DotProductWithPaddingOp<float, CPUContext>::RunOnDevice() {
  auto& X = Input(X_IN);
  auto& Y = Input(Y_IN);
  auto* result = Output(DOT_OUT);
  CAFFE_ENFORCE_EQ(X.ndim(), Y.ndim());
  CAFFE_ENFORCE_EQ(X.dim32(0), Y.dim32(0));

  int N, D, DX, DY, restD;
  if (X.size() > 0) {
    N = X.ndim() > 0 ? X.dim32(0) : 1;
    DX = X.size() / N;
    DY = Y.size() / N;
  } else {
    N = 0;
    DX = 0;
    DY = 0;
  }

  D = std::min(DX, DY);
  restD = std::max(DX, DY) - D;
  result->Resize(N);
  float* result_data = result->mutable_data<float>();
  const float* X_data = X.data<float>();
  const float* Y_data = Y.data<float>();
  for (int i = 0; i < N; ++i) { // TODO: multithreading
    auto offsetX = i * DX, offsetY = i * DY;
    if (replicate_) {
      // L_ for longer vector and S_ for shorter vector
      const float *L_data, *S_data;
      int DL, DS;
      if (DX > DY) {
        L_data = X_data + offsetX;
        S_data = Y_data + offsetY;
        DL = DX;
        DS = DY;
      } else {
        L_data = Y_data + offsetY;
        S_data = X_data + offsetX;
        DL = DY;
        DS = DX;
      }
      float sum = 0.0;
      float tmp = 0.0;
      for (int j = 0; j < DL / DS; j++) {
        math::Dot<float, CPUContext>(
            DS, L_data + j * DS, S_data, &tmp, &context_);
        sum += tmp;
      }
      *(result_data + i) = sum;
    } else {
      math::Dot<float, CPUContext>(
          D, X_data + offsetX, Y_data + offsetY, result_data + i, &context_);
    }

    if (!replicate_ && DX != DY) {
      const float* rest_data;
      float rest_sum = 0;
      if (DX > DY) {
        rest_data = X_data + offsetX + D;
      } else {
        rest_data = Y_data + offsetY + D;
      }
      math::Sum<float, CPUContext>(restD, rest_data, &rest_sum, &context_);
      result_data[i] += rest_sum * pad_value_;
    }
  }
  return true;
}

// L2
REGISTER_CPU_OPERATOR(SquaredL2Distance,
                      SquaredL2DistanceOp<float, CPUContext>);
REGISTER_CPU_OPERATOR(SquaredL2DistanceGradient,
                      SquaredL2DistanceGradientOp<float, CPUContext>);

OPERATOR_SCHEMA(SquaredL2Distance)
    .NumInputs(2)
    .NumOutputs(1)
    .IdenticalTypeAndShapeOfInputDim(0, 0)
    .SetDoc(R"DOC(
Given two input float tensors X, Y, and produces one output float tensor
of the L2 difference between X and Y that is computed as ||(X - Y)^2 / 2||.
)DOC")
    .Input(0, "X", "1D or 2D input tensor")
    .Input(1, "Y", "1D or 2D input tensor (must have the same shape as X)")
    .Output(0, "Z", "1D output tensor");

OPERATOR_SCHEMA(SquaredL2DistanceGradient).NumInputs(3).NumOutputs(2);

class GetSquaredL2DistanceGradient : public GradientMakerBase {
  using GradientMakerBase::GradientMakerBase;
  vector<OperatorDef> GetGradientDefs() override {
    return SingleGradientDef(
        "SquaredL2DistanceGradient", "",
        vector<string>{I(0), I(1), GO(0)},
        vector<string>{GI(0), GI(1)});
  }
};
REGISTER_GRADIENT(SquaredL2Distance, GetSquaredL2DistanceGradient);

// L1
REGISTER_CPU_OPERATOR(L1Distance, L1DistanceOp<float, CPUContext>);
REGISTER_CPU_OPERATOR(
    L1DistanceGradient,
    L1DistanceGradientOp<float, CPUContext>);

OPERATOR_SCHEMA(L1Distance)
    .NumInputs(2)
    .NumOutputs(1)
    .IdenticalTypeAndShapeOfInputDim(0, 0)
    .SetDoc(R"DOC(
Given two input float tensors X, Y, and produces one output float tensor
of the L1 difference between X and Y, computed as L1(x,y) = sum over |x-y|
)DOC")
    .Input(0, "X", "1D or 2D input tensor")
    .Input(1, "Y", "1D or 2D input tensor (must have the same shape as X)")
    .Output(0, "Z", "1D output tensor");

OPERATOR_SCHEMA(L1DistanceGradient).NumInputs(3).NumOutputs(2);

class GetL1DistanceGradient : public GradientMakerBase {
  using GradientMakerBase::GradientMakerBase;
  vector<OperatorDef> GetGradientDefs() override {
    return SingleGradientDef(
        "L1DistanceGradient",
        "",
        vector<string>{I(0), I(1), GO(0)},
        vector<string>{GI(0), GI(1)});
  }
};

REGISTER_GRADIENT(L1Distance, GetL1DistanceGradient);

// Dot Product
REGISTER_CPU_OPERATOR(DotProduct, DotProductOp<float, CPUContext>);
REGISTER_CPU_OPERATOR(
    DotProductGradient,
    DotProductGradientOp<float, CPUContext>);

OPERATOR_SCHEMA(DotProduct)
    .NumInputs(2)
    .NumOutputs(1)
    .IdenticalTypeAndShapeOfInputDim(0, 0)
    .SetDoc(R"DOC(
Given two input float tensors X, Y, and produces one output float tensor
of the dot product between X and Y.
)DOC")
    .Input(0, "X", "1D or 2D input tensor")
    .Input(1, "Y", "1D or 2D input tensor (must have the same shape as X)")
    .Output(0, "Z", "1D output tensor")
    .CostInferenceFunction(
        OpSchema::CostInferenceFunctionType(CostInferenceForDotProduct));

OPERATOR_SCHEMA(DotProductGradient).NumInputs(3).NumOutputs(2);

class GetDotProductGradient : public GradientMakerBase {
  using GradientMakerBase::GradientMakerBase;
  vector<OperatorDef> GetGradientDefs() override {
    return SingleGradientDef(
        "DotProductGradient",
        "",
        vector<string>{I(0), I(1), GO(0)},
        vector<string>{GI(0), GI(1)});
  }
};
REGISTER_GRADIENT(DotProduct, GetDotProductGradient);

// Cosine Similarity
REGISTER_CPU_OPERATOR(CosineSimilarity, CosineSimilarityOp<float, CPUContext>);
REGISTER_CPU_OPERATOR(
    CosineSimilarityGradient,
    CosineSimilarityGradientOp<float, CPUContext>);

OPERATOR_SCHEMA(CosineSimilarity)
    .NumInputs(2)
    .NumOutputs(1)
    .IdenticalTypeAndShapeOfInputDim(0, 0)
    .SetDoc(R"DOC(
Given two input float tensors X, Y, and produces one output float tensor
of the cosine similarity between X and Y.
)DOC")
    .Input(0, "X", "1D or 2D input tensor")
    .Input(1, "Y", "1D or 2D input tensor (must have the same shape as X)")
    .Output(0, "Z", "1D output tensor");

OPERATOR_SCHEMA(CosineSimilarityGradient).NumInputs(3).NumOutputs(2);

class GetCosineSimilarityGradient : public GradientMakerBase {
  using GradientMakerBase::GradientMakerBase;
  vector<OperatorDef> GetGradientDefs() override {
    return SingleGradientDef(
        "CosineSimilarityGradient",
        "",
        vector<string>{I(0), I(1), GO(0)},
        vector<string>{GI(0), GI(1)});
  }
};
REGISTER_GRADIENT(CosineSimilarity, GetCosineSimilarityGradient);

// Dot Product allows padding
REGISTER_CPU_OPERATOR(
    DotProductWithPadding,
    DotProductWithPaddingOp<float, CPUContext>);
REGISTER_CPU_OPERATOR(
    DotProductWithPaddingGradient,
    DotProductWithPaddingGradientOp<float, CPUContext>);

OPERATOR_SCHEMA(DotProductWithPadding)
    .NumInputs(2)
    .NumOutputs(1)
    .SetDoc(R"DOC(
Given two input float tensors X, Y with different shapes and produces one
output float tensor of the dot product between X and Y. We currently support
two kinds of strategies to achieve this. Before doing normal dot_product 1)
pad the smaller tensor (using pad_value) to the same shape as the other one.
2) replicate the smaller tensor to the same shape as the other one. Note the
first dimension of X, Y must be equal. Only the second dimension of X or Y
can be padded.
)DOC")
    .Input(0, "X", "1D or 2D input tensor")
    .Input(1, "Y", "1D or 2D input tensor")
    .Output(0, "Z", "1D output tensor")
    .IdenticalTypeAndShapeOfInputDim(0, 0)
    .Arg("pad_value", "the padding value for tensors with smaller dimension")
    .Arg("replicate", "whether to replicate the smaller tensor or not");

OPERATOR_SCHEMA(DotProductWithPaddingGradient).NumInputs(3).NumOutputs(2);

class GetDotProductWithPaddingGradient : public GradientMakerBase {
  using GradientMakerBase::GradientMakerBase;
  vector<OperatorDef> GetGradientDefs() override {
    float pad_value = 0;
    bool replicate = false;
    if (ArgumentHelper::HasArgument(Def(), "pad_value")) {
      pad_value = GetArgument(Def(), "pad_value").f();
    }
    if (ArgumentHelper::HasArgument(Def(), "replicate")) {
      replicate = GetArgument(Def(), "replicate").i();
    }

    const auto dot_arg =
        vector<Argument>{MakeArgument<float>("pad_value", pad_value),
                         MakeArgument<bool>("replicate", replicate)};

    return SingleGradientDef(
        "DotProductWithPaddingGradient",
        "",
        vector<string>{I(0), I(1), GO(0)},
        vector<string>{GI(0), GI(1)},
        dot_arg);
  }
};
REGISTER_GRADIENT(DotProductWithPadding, GetDotProductWithPaddingGradient);
}  // namespace caffe2