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/weight_scale_op.h (2552B)

/**
 * Copyright (c) 2016-present, Facebook, Inc.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#pragma once

#include "caffe2/core/operator.h"

#include <stdlib.h>
#include <time.h>

namespace caffe2 {

template <typename T, class Context>
void weight_scale_update(
    int N,
    const T* w,
    const T scale,
    int64_t iter,
    int64_t stepsize,
    int64_t update_upper_bound,
    T* nw,
    Context* context) {
  const auto w_size = N * sizeof(float);
  if (iter % stepsize != 0 || iter >= update_upper_bound) {
    memcpy(nw, w, w_size);
    return;
  }
  // perform the weight scaling
  caffe2::math::Scale<T, T, Context>(N, scale, w, nw, context);
}

template <class Context>
class WeightScaleOp final : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  WeightScaleOp(const OperatorDef& operator_def, Workspace* ws)
      : Operator<Context>(operator_def, ws),
        stepsize_(OperatorBase::GetSingleArgument<int64_t>(
            "stepsize",
            std::numeric_limits<int64_t>::max())),
        update_upper_bound_(OperatorBase::GetSingleArgument<int64_t>(
            "upper_bound_iter",
            std::numeric_limits<int64_t>::max())),
        scale_(this->template GetSingleArgument<float>("scale", 1.0f)) {}

bool RunOnDevice() override {
    Output(OUTPUT_WEIGHTS)->ResizeLike(Input(WEIGHTS));
    return DispatchHelper<TensorTypes<float>>::call(this, Input(WEIGHTS));
  }

template <typename T>
  bool DoRunWithType() {
    const auto iter =
        OperatorBase::Input<Tensor>(ITER, CPU).template data<int64_t>()[0] + 1;

weight_scale_update<T, Context>(
        Input(WEIGHTS).size(),
        Input(WEIGHTS).template data<T>(),
        scale_,
        iter,
        stepsize_,
        update_upper_bound_,
        Output(OUTPUT_WEIGHTS)->template mutable_data<T>(),
        &context_);
    return true;
  }

protected:
  int64_t stepsize_;
  int64_t update_upper_bound_;
  float scale_;
  INPUT_TAGS(WEIGHTS, ITER);
  OUTPUT_TAGS(OUTPUT_WEIGHTS);
};

} // namespace caffe2

cmd: