#ifndef CAFFE2_OPERATORS_SINUSOID_POSITION_ENCODING_OP_H_
|
#define CAFFE2_OPERATORS_SINUSOID_POSITION_ENCODING_OP_H_
|
|
#ifdef _MSC_VER
|
#define _USE_MATH_DEFINES
|
#endif // _MSC_VER
|
#include <cmath>
|
|
#include "caffe2/core/operator.h"
|
|
#include "Eigen/Core"
|
#include "caffe2/utils/eigen_utils.h"
|
|
namespace caffe2 {
|
|
template <class Context>
|
class SinusoidPositionEncodingOp : public Operator<Context> {
|
public:
|
template <class... Args>
|
explicit SinusoidPositionEncodingOp(Args&&... args)
|
: Operator<Context>(std::forward<Args>(args)...),
|
embedding_size_(
|
this->template GetSingleArgument<int>("embedding_size", 100)),
|
alpha_(this->template GetSingleArgument<float>("alpha", 10000)),
|
amplitude_(this->template GetSingleArgument<float>("amplitude", 1)) {}
|
USE_OPERATOR_CONTEXT_FUNCTIONS;
|
|
bool RunOnDevice() override {
|
return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(
|
this, this->template Input<Tensor>(0, CPU));
|
}
|
|
template <typename Index>
|
bool DoRunWithType() {
|
auto& positions = Input(0);
|
|
CAFFE_ENFORCE_EQ(positions.dim(), 2, "POSITIONS should be a 2-D tensor");
|
|
auto shape = positions.sizes().vec();
|
shape.push_back(embedding_size_);
|
auto* output = Output(0, shape, at::dtype<float>());
|
|
int M = shape[0];
|
int K = shape[1];
|
const Index* idxs = positions.template data<Index>();
|
float* out = output->template mutable_data<float>();
|
|
float log_alpha = std::log(alpha_);
|
float max_alpha_pow =
|
((float)embedding_size_ - 1.0f) / (float)embedding_size_;
|
|
for (int i = 0; i < M; ++i) {
|
float pos = (float)idxs[i * K];
|
|
// Compute the embedding for position i, example 0 first
|
float* row = &out[i * K * embedding_size_];
|
Eigen::Map<Eigen::VectorXf> row_map(row, embedding_size_, 1);
|
auto row_array = row_map.array();
|
|
float log_pos = std::log(pos);
|
row_array.setLinSpaced(
|
embedding_size_, log_pos, log_pos - log_alpha * max_alpha_pow);
|
row_array = row_array.exp().eval();
|
// row_array[k] == pos / alpha^(k / embedding_size)
|
|
// Phase shift so that alternating elements are cosines
|
for (int k = 1; k < embedding_size_; k += 2) {
|
row[k] += (float)M_PI_2;
|
}
|
row_array = amplitude_ * row_array.sin().eval();
|
|
// Copy the embedding to position i in the other examples
|
for (int j = 1; j < K; ++j) {
|
int base = i * K * embedding_size_;
|
std::copy(
|
&out[base],
|
&out[base + embedding_size_],
|
&out[base + j * embedding_size_]);
|
}
|
}
|
return true;
|
}
|
|
protected:
|
int embedding_size_;
|
float alpha_;
|
float amplitude_;
|
};
|
|
} // namespace caffe2
|
|
#endif // CAFFE2_OPERATORS_SINUSOID_POSITION_ENCODING_OP_H_
|
