FM / sgd

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/sgd/

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/sgd

Name	Size	Mode	Actions
adadelta_op.h	5852	0644	edit dl rm
adagrad_fused.h	16632	0644	edit dl rm
adagrad_op.h	17993	0644	edit dl rm
adam_op.h	26165	0644	edit dl rm
clip_tensor_op.h	1853	0644	edit dl rm
decay_adagrad_op.h	3215	0644	edit dl rm
fp16_momentum_sgd_op.h	2292	0644	edit dl rm
fp32_momentum_sgd_op.h	2015	0644	edit dl rm
ftrl_op.h	2219	0644	edit dl rm
gftrl_op.h	1028	0644	edit dl rm
iter_op.h	3379	0644	edit dl rm
lars_op.h	2479	0644	edit dl rm
learning_rate_adaption_op.h	1974	0644	edit dl rm
learning_rate_functors.h	15135	0644	edit dl rm
learning_rate_op.h	12696	0644	edit dl rm
math_lp.h	655	0644	edit dl rm
momentum_sgd_op.h	6035	0644	edit dl rm
rmsprop_op.h	1980	0644	edit dl rm
rowwise_adagrad_fused.h	29595	0644	edit dl rm
rowwise_counter.h	2083	0644	edit dl rm
storm_op.h	5915	0644	edit dl rm
weight_scale_op.h	2552	0644	edit dl rm
wngrad_op.h	7272	0644	edit dl rm
yellowfin_op.h	10284	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/caffe2/sgd/rmsprop_op.h (1980B)

#pragma once

#include "caffe2/core/common_omp.h"
#include "caffe2/core/operator.h"

namespace caffe2 {

template <typename Context>
void rmsprop_update(
    int N,
    const float* g,
    const float* ms,
    const float* mom,
    float* ng,
    float* nms,
    float* nmom,
    float decay,
    float momentum,
    float epsilon,
    const float* lr,
    Context* context);

template <typename T, class Context>
class RmsPropOp final : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  RmsPropOp(const OperatorDef& operator_def, Workspace* ws)
      : Operator<Context>(operator_def, ws),
        decay_(this->template GetSingleArgument<float>("decay", 0.9f)),
        momentum_(this->template GetSingleArgument<float>("momentum", 0.0f)),
        epsilon_(this->template GetSingleArgument<float>("epsilon", 1e-5f)) {}
  bool RunOnDevice() override {
    CAFFE_ENFORCE(Input(LR).numel() == 1);
    CAFFE_ENFORCE(Input(GRAD).numel() == Input(MEAN_SQUARES).numel());
    CAFFE_ENFORCE(Input(GRAD).numel() == Input(OUTPUT_MOMENTUM).numel());
    Output(OUTPUT_GRAD)->ResizeLike(Input(GRAD));
    Output(OUTPUT_GRAD)->ResizeLike(Input(GRAD));
    Output(OUTPUT_MEAN_SQUARES)->ResizeLike(Input(MEAN_SQUARES));
    Output(OUTPUT_MOMENTUM)->ResizeLike(Input(MOMENTUM));
    rmsprop_update<Context>(
        Input(GRAD).numel(),
        Input(GRAD).template data<T>(),
        Input(MEAN_SQUARES).template data<T>(),
        Input(MOMENTUM).template data<T>(),
        Output(OUTPUT_GRAD)->template mutable_data<T>(),
        Output(OUTPUT_MEAN_SQUARES)->template mutable_data<T>(),
        Output(OUTPUT_MOMENTUM)->template mutable_data<T>(),
        decay_,
        momentum_,
        epsilon_,
        Input(LR).template data<T>(),
        &context_);
    return true;
  }

protected:
  T decay_{0.9};
  T momentum_{0.0};
  T epsilon_{1e-8};
  INPUT_TAGS(GRAD, MEAN_SQUARES, MOMENTUM, LR);
  OUTPUT_TAGS(OUTPUT_GRAD, OUTPUT_MEAN_SQUARES, OUTPUT_MOMENTUM);
};
}

#pragma once #include "caffe2/core/common_omp.h" #include "caffe2/core/operator.h" namespace caffe2 { template void rmsprop_update( int N, const float* g, const float* ms, const float* mom, float* ng, float* nms, float* nmom, float decay, float momentum, float epsilon, const float* lr, Context* context); template class RmsPropOp final : public Operator { public: USE_OPERATOR_CONTEXT_FUNCTIONS; RmsPropOp(const OperatorDef& operator_def, Workspace* ws) : Operator(operator_def, ws), decay_(this->template GetSingleArgument("decay", 0.9f)), momentum_(this->template GetSingleArgument("momentum", 0.0f)), epsilon_(this->template GetSingleArgument("epsilon", 1e-5f)) {} bool RunOnDevice() override { CAFFE_ENFORCE(Input(LR).numel() == 1); CAFFE_ENFORCE(Input(GRAD).numel() == Input(MEAN_SQUARES).numel()); CAFFE_ENFORCE(Input(GRAD).numel() == Input(OUTPUT_MOMENTUM).numel()); Output(OUTPUT_GRAD)->ResizeLike(Input(GRAD)); Output(OUTPUT_GRAD)->ResizeLike(Input(GRAD)); Output(OUTPUT_MEAN_SQUARES)->ResizeLike(Input(MEAN_SQUARES)); Output(OUTPUT_MOMENTUM)->ResizeLike(Input(MOMENTUM)); rmsprop_update( Input(GRAD).numel(), Input(GRAD).template data(), Input(MEAN_SQUARES).template data(), Input(MOMENTUM).template data(), Output(OUTPUT_GRAD)->template mutable_data(), Output(OUTPUT_MEAN_SQUARES)->template mutable_data(), Output(OUTPUT_MOMENTUM)->template mutable_data(), decay_, momentum_, epsilon_, Input(LR).template data(), &context_); return true; } protected: T decay_{0.9}; T momentum_{0.0}; T epsilon_{1e-8}; INPUT_TAGS(GRAD, MEAN_SQUARES, MOMENTUM, LR); OUTPUT_TAGS(OUTPUT_GRAD, OUTPUT_MEAN_SQUARES, OUTPUT_MOMENTUM); }; }