caffe源码

版权声明:本文为博主原创文章,未经博主允许不得转载。 https://blog.csdn.net/lantuxin/article/details/82743618

1、各层实现的简单框架

参考网站:https://github.com/BVLC/caffe/wiki/Simple-Example:-Sin-Layer

目前的理解:每一层的代码框架都可以单独编写,并且可以测试自己写的层的代码能否通过。

我这里以tanh_layer为例。

1)tanh_layer.hpp

定义了ReLULayer的类,链接了protobuf,声明了前向传播、反向传播函数(cpu/gpu)

#ifndef CAFFE_TANH_LAYER_HPP_
#define CAFFE_TANH_LAYER_HPP_

#include <vector>

#include "caffe/blob.hpp"
#include "caffe/layer.hpp"
#include "caffe/proto/caffe.pb.h"

#include "caffe/layers/neuron_layer.hpp"

namespace caffe {

/**
 * @brief TanH hyperbolic tangent non-linearity @f$
 *         y = \frac{\exp(2x) - 1}{\exp(2x) + 1}
 *     @f$, popular in auto-encoders.
 *
 * Note that the gradient vanishes as the values move away from 0.
 * The ReLULayer is often a better choice for this reason.
 */
template <typename Dtype>
class TanHLayer : public NeuronLayer<Dtype> {
 public:
  explicit TanHLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}

  virtual inline const char* type() const { return "TanH"; } //link the protobuf code to your C++ code具体怎么链接的还没搞懂。

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *        y = \frac{\exp(2x) - 1}{\exp(2x) + 1}
   *      @f$
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the sigmoid inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x}
   *            = \frac{\partial E}{\partial y}
   *              \left(1 - \left[\frac{\exp(2x) - 1}{exp(2x) + 1} \right]^2 \right)
   *            = \frac{\partial E}{\partial y} (1 - y^2)
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
};

}  // namespace caffe

#endif  // CAFFE_TANH_LAYER_HPP_

2)tanh_layer实现

①cpu实现方法:tanh_layer.cpp

实现了前向传播、反向传播函数,其中主要是需要学会使用cpu_data(),mutable_cpu_data(),cpu_diff()

// TanH neuron activation function layer.
// Adapted from ReLU layer code written by Yangqing Jia

#include <vector>

#include "caffe/layers/tanh_layer.hpp"

namespace caffe {

template <typename Dtype>
void TanHLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  const Dtype* bottom_data = bottom[0]->cpu_data();
  Dtype* top_data = top[0]->mutable_cpu_data();
  const int count = bottom[0]->count();
  for (int i = 0; i < count; ++i) {
    top_data[i] = tanh(bottom_data[i]);
  }
}

template <typename Dtype>
void TanHLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
    const vector<bool>& propagate_down,
    const vector<Blob<Dtype>*>& bottom) {
  if (propagate_down[0]) {
    const Dtype* top_data = top[0]->cpu_data();
    const Dtype* top_diff = top[0]->cpu_diff();
    Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
    const int count = bottom[0]->count();
    Dtype tanhx;
    for (int i = 0; i < count; ++i) {
      tanhx = top_data[i];
      bottom_diff[i] = top_diff[i] * (1 - tanhx * tanhx);
    }
  }
}

#ifdef CPU_ONLY
STUB_GPU(TanHLayer);
#endif

INSTANTIATE_CLASS(TanHLayer);

}  // namespace caffe

②gpu实现方法:tanh_layer.cu

gpu边长语法格式与c++类似,但还有一些特定的调用方法,如__global__的使用。

// TanH neuron activation function layer.
// Adapted from ReLU layer code written by Yangqing Jia

#include <vector>

#include "caffe/layers/tanh_layer.hpp"

namespace caffe {

template <typename Dtype>
__global__ void TanHForward(const int n, const Dtype* in, Dtype* out) {
  CUDA_KERNEL_LOOP(index, n) {
    out[index] = tanh(in[index]);
  }
}

template <typename Dtype>
void TanHLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  const Dtype* bottom_data = bottom[0]->gpu_data();
  Dtype* top_data = top[0]->mutable_gpu_data();
  const int count = bottom[0]->count();
  // NOLINT_NEXT_LINE(whitespace/operators)
  TanHForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
      count, bottom_data, top_data);
  CUDA_POST_KERNEL_CHECK;
}

template <typename Dtype>
__global__ void TanHBackward(const int n, const Dtype* in_diff,
    const Dtype* out_data, Dtype* out_diff) {
  CUDA_KERNEL_LOOP(index, n) {
    Dtype tanhx = out_data[index];
    out_diff[index] = in_diff[index] * (1 - tanhx * tanhx);
  }
}

template <typename Dtype>
void TanHLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
    const vector<bool>& propagate_down,
    const vector<Blob<Dtype>*>& bottom) {
  if (propagate_down[0]) {
    const Dtype* top_data = top[0]->gpu_data();
    const Dtype* top_diff = top[0]->gpu_diff();
    Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
    const int count = bottom[0]->count();
    // NOLINT_NEXT_LINE(whitespace/operators)
    TanHBackward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
        count, top_diff, top_data, bottom_diff);
    CUDA_POST_KERNEL_CHECK;
  }
}

INSTANTIATE_LAYER_GPU_FUNCS(TanHLayer);


}  // namespace caffe

3)test_tanh_layer.cpp

 

#include <algorithm>
#include <vector>

#include "gtest/gtest.h"

#include "caffe/blob.hpp"
#include "caffe/common.hpp"
#include "caffe/filler.hpp"
#include "caffe/layers/tanh_layer.hpp"

#include "caffe/test/test_caffe_main.hpp"
#include "caffe/test/test_gradient_check_util.hpp"

namespace caffe {

double tanh_naive(double x) {
  if (x < -40) {
    // avoid negative overflow
    return -1;
  } else if (x > 40) {
    // avoid positive overflow
    return 1;
  } else {
    // exact expression for tanh, which is unstable for large x
    double exp2x = exp(2 * x);
    return (exp2x - 1.0) / (exp2x + 1.0);
  }
}

template <typename TypeParam>
class TanHLayerTest : public MultiDeviceTest<TypeParam> {
  typedef typename TypeParam::Dtype Dtype;

 protected:
  TanHLayerTest()
      : blob_bottom_(new Blob<Dtype>(2, 3, 4, 5)),
        blob_top_(new Blob<Dtype>()) {
    Caffe::set_random_seed(1701);
    FillerParameter filler_param;
    blob_bottom_vec_.push_back(blob_bottom_);
    blob_top_vec_.push_back(blob_top_);
  }
  virtual ~TanHLayerTest() { delete blob_bottom_; delete blob_top_; }

  void TestForward(Dtype filler_std) {
    FillerParameter filler_param;
    filler_param.set_std(filler_std);
    GaussianFiller<Dtype> filler(filler_param);
    filler.Fill(this->blob_bottom_);

    LayerParameter layer_param;
    TanHLayer<Dtype> layer(layer_param);
    layer.SetUp(this->blob_bottom_vec_, this->blob_top_vec_);
    layer.Forward(this->blob_bottom_vec_, this->blob_top_vec_);
    // Now, check values
    const Dtype* bottom_data = this->blob_bottom_->cpu_data();
    const Dtype* top_data = this->blob_top_->cpu_data();
    const Dtype min_precision = 1e-5;
    for (int i = 0; i < this->blob_bottom_->count(); ++i) {
      Dtype expected_value = tanh_naive(bottom_data[i]);
      Dtype precision = std::max(
        Dtype(std::abs(expected_value * Dtype(1e-4))), min_precision);
      EXPECT_NEAR(expected_value, top_data[i], precision);
    }
  }

  void TestBackward(Dtype filler_std) {
    FillerParameter filler_param;
    filler_param.set_std(filler_std);
    GaussianFiller<Dtype> filler(filler_param);
    filler.Fill(this->blob_bottom_);

    LayerParameter layer_param;
    TanHLayer<Dtype> layer(layer_param);
    GradientChecker<Dtype> checker(1e-2, 1e-2, 1701);
    checker.CheckGradientEltwise(&layer, this->blob_bottom_vec_,
        this->blob_top_vec_);
  }

  Blob<Dtype>* const blob_bottom_;
  Blob<Dtype>* const blob_top_;
  vector<Blob<Dtype>*> blob_bottom_vec_;
  vector<Blob<Dtype>*> blob_top_vec_;
};

TYPED_TEST_CASE(TanHLayerTest, TestDtypesAndDevices);

TYPED_TEST(TanHLayerTest, TestTanH) {
  this->TestForward(1.0);
}

TYPED_TEST(TanHLayerTest, TestTanHOverflow) {
  // this will fail if tanh overflow is not properly handled
  this->TestForward(10000.0);
}

TYPED_TEST(TanHLayerTest, TestTanHGradient) {
  this->TestBackward(1.0);
}

}  // namespace caffe

最后就可以测试自己自己写的层是否能用。

> make
> make test
> make runtest GTEST_FILTER='TanHLayerTest/*'

 

阅读更多
想对作者说点什么?

博主推荐

换一批

没有更多推荐了,返回首页