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/ftrl_op.h (2219B)

#pragma once

#include "caffe2/core/operator.h"

namespace caffe2 {

template <typename T>
struct FtrlParams {
  explicit FtrlParams(OperatorBase* op)
      : alphaInv(1.0 / op->GetSingleArgument<float>("alpha", 0.005f)),
        beta(op->GetSingleArgument<float>("beta", 1.0f)),
        lambda1(op->GetSingleArgument<float>("lambda1", 0.001f)),
        lambda2(op->GetSingleArgument<float>("lambda2", 0.001f)) {}
  T alphaInv;
  T beta;
  T lambda1;
  T lambda2;
};

// TODO(dzhulgakov): implement GPU version if necessary
template <typename T, class Context>
class FtrlOp final : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  FtrlOp(const OperatorDef& operator_def, Workspace* ws)
      : Operator<Context>(operator_def, ws), params_(this) {
    CAFFE_ENFORCE(
        !HasArgument("alpha") || ALPHA >= InputSize(),
        "Cannot specify alpha by both input and argument");
  }
  bool RunOnDevice() override;

protected:
  FtrlParams<T> params_;
  INPUT_TAGS(VAR, N_Z, GRAD, ALPHA);
  OUTPUT_TAGS(OUTPUT_VAR, OUTPUT_N_Z);
};

template <typename T>
class SparseFtrlOp final : public Operator<CPUContext> {
 public:
  SparseFtrlOp(const OperatorDef& operator_def, Workspace* ws)
      : Operator<CPUContext>(operator_def, ws), params_(this) {
    CAFFE_ENFORCE(
        !HasArgument("alpha") || ALPHA >= InputSize(),
        "Cannot specify alpha by both input and argument");
  }

bool RunOnDevice() override {
    // run time learning rate override
    if (ALPHA < InputSize()) {
      CAFFE_ENFORCE_EQ(Input(ALPHA).numel(), 1, "alpha should be real-valued");
      params_.alphaInv = 1.0 / *(Input(ALPHA).template data<T>());
    }
    // Use run-time polymorphism
    auto& indices = Input(INDICES);
    if (indices.template IsType<int32_t>()) {
      DoRun<int32_t>();
    } else if (indices.template IsType<int64_t>()) {
      DoRun<int64_t>();
    } else {
      LOG(FATAL) << "Unsupported type of INDICES in SparseFtrlOp: "
                 << indices.dtype().name();
    }
    return true;
  }

protected:
  FtrlParams<T> params_;
  INPUT_TAGS(VAR, N_Z, INDICES, GRAD, ALPHA);
  OUTPUT_TAGS(OUTPUT_VAR, OUTPUT_N_Z);

private:
  template <typename SIndex>
  void DoRun();
};

}

#pragma once #include "caffe2/core/operator.h" namespace caffe2 { template struct FtrlParams { explicit FtrlParams(OperatorBase* op) : alphaInv(1.0 / op->GetSingleArgument("alpha", 0.005f)), beta(op->GetSingleArgument("beta", 1.0f)), lambda1(op->GetSingleArgument("lambda1", 0.001f)), lambda2(op->GetSingleArgument("lambda2", 0.001f)) {} T alphaInv; T beta; T lambda1; T lambda2; }; // TODO(dzhulgakov): implement GPU version if necessary template class FtrlOp final : public Operator { public: USE_OPERATOR_CONTEXT_FUNCTIONS; FtrlOp(const OperatorDef& operator_def, Workspace* ws) : Operator(operator_def, ws), params_(this) { CAFFE_ENFORCE( !HasArgument("alpha") || ALPHA >= InputSize(), "Cannot specify alpha by both input and argument"); } bool RunOnDevice() override; protected: FtrlParams params_; INPUT_TAGS(VAR, N_Z, GRAD, ALPHA); OUTPUT_TAGS(OUTPUT_VAR, OUTPUT_N_Z); }; template class SparseFtrlOp final : public Operator { public: SparseFtrlOp(const OperatorDef& operator_def, Workspace* ws) : Operator(operator_def, ws), params_(this) { CAFFE_ENFORCE( !HasArgument("alpha") || ALPHA >= InputSize(), "Cannot specify alpha by both input and argument"); } bool RunOnDevice() override { // run time learning rate override if (ALPHA < InputSize()) { CAFFE_ENFORCE_EQ(Input(ALPHA).numel(), 1, "alpha should be real-valued"); params_.alphaInv = 1.0 / *(Input(ALPHA).template data()); } // Use run-time polymorphism auto& indices = Input(INDICES); if (indices.template IsType()) { DoRun(); } else if (indices.template IsType()) { DoRun(); } else { LOG(FATAL) << "Unsupported type of INDICES in SparseFtrlOp: " << indices.dtype().name(); } return true; } protected: FtrlParams params_; INPUT_TAGS(VAR, N_Z, INDICES, GRAD, ALPHA); OUTPUT_TAGS(OUTPUT_VAR, OUTPUT_N_Z); private: template void DoRun(); }; }