Caffe源码（四）：base_conv_layer 分析

目录

简单介绍

base_conv_layer.cpp 中定义了 BaseConvolutionLayer 类的一些成员函数，而BaseConvolutionLayer 是 ConvolutionLayer 的父类，ConvolutionLayer中用到的一些函数都在这里定义，所以在看conv_layer前需要看此源代码。

主要函数

1. LayerSetUp 函数：

template <typename Dtype>
void BaseConvolutionLayer::LayerSetUp(const vector*>& bottom,
      const vector*>& top) {
  CHECK_EQ(4, bottom[0]->num_axes()) << "Input must have 4 axes, "
      << "corresponding to (num, channels, height, width)";
  // Configure the kernel size, padding, stride, and inputs.
  ConvolutionParameter conv_param = this->layer_param_.convolution_param();
  CHECK(!conv_param.has_kernel_size() !=
      !(conv_param.has_kernel_h() && conv_param.has_kernel_w()))
      << "Filter size is kernel_size OR kernel_h and kernel_w; not both";
  CHECK(conv_param.has_kernel_size() ||
      (conv_param.has_kernel_h() && conv_param.has_kernel_w()))
      << "For non-square filters both kernel_h and kernel_w are required.";
  CHECK((!conv_param.has_pad() && conv_param.has_pad_h()
      && conv_param.has_pad_w())
      || (!conv_param.has_pad_h() && !conv_param.has_pad_w()))
      << "pad is pad OR pad_h and pad_w are required.";
  CHECK((!conv_param.has_stride() && conv_param.has_stride_h()
      && conv_param.has_stride_w())
      || (!conv_param.has_stride_h() && !conv_param.has_stride_w()))
      << "Stride is stride OR stride_h and stride_w are required.";
  if (conv_param.has_kernel_size()) {
    kernel_h_ = kernel_w_ = conv_param.kernel_size();
  } else {
    kernel_h_ = conv_param.kernel_h();
    kernel_w_ = conv_param.kernel_w();
  }//用户自定义kernel size 的两种方式
  CHECK_GT(kernel_h_, 0) << "Filter dimensions cannot be zero.";
  CHECK_GT(kernel_w_, 0) << "Filter dimensions cannot be zero.";
  if (!conv_param.has_pad_h()) {
    pad_h_ = pad_w_ = conv_param.pad();
  } else {
    pad_h_ = conv_param.pad_h();
    pad_w_ = conv_param.pad_w();
  }
  if (!conv_param.has_stride_h()) {
    stride_h_ = stride_w_ = conv_param.stride();
  } else {
    stride_h_ = conv_param.stride_h();
    stride_w_ = conv_param.stride_w();
  }
  // Special case: im2col is the identity for 1x1 convolution with stride 1
  // and no padding, so flag for skipping the buffer and transformation.
  is_1x1_ = kernel_w_ == 1 && kernel_h_ == 1
      && stride_h_ == 1 && stride_w_ == 1 && pad_h_ == 0 && pad_w_ == 0;
  // Configure output channels and groups.
  channels_ = bottom[0]->channels();
  num_output_ = this->layer_param_.convolution_param().num_output();
  CHECK_GT(num_output_, 0);
  group_ = this->layer_param_.convolution_param().group();
  CHECK_EQ(channels_ % group_, 0);
  CHECK_EQ(num_output_ % group_, 0)
      << "Number of output should be multiples of group.";
      //channel 和 输出 feature map 个数必须为group的整数倍，每个group中只用本group的featrue map
  if (reverse_dimensions()) {
    conv_out_channels_ = channels_;
    conv_in_channels_ = num_output_;
  } else {
    conv_out_channels_ = num_output_;//用户指定输出feature map的数量
    conv_in_channels_ = channels_;
  }
  // Handle the parameters: weights and biases.
  // - blobs_[0] holds the filter weights
  // - blobs_[1] holds the biases (optional)
  bias_term_ = this->layer_param_.convolution_param().bias_term();//默认为 true
  if (this->blobs_.size() > 0) {
    LOG(INFO) << "Skipping parameter initialization";
  } else {
    if (bias_term_) {
      this->blobs_.resize(2);
    } else {
      this->blobs_.resize(1);
    }
    // Initialize and fill the weights:
    // output channels x input channels per-group x kernel height x kernel width
    this->blobs_[0].reset(new Blob(
        conv_out_channels_, conv_in_channels_ / group_, kernel_h_, kernel_w_));
    shared_ptr > weight_filler(GetFiller(
        this->layer_param_.convolution_param().weight_filler()));
    weight_filler->Fill(this->blobs_[0].get()); //用weight_filler初始化
    // If necessary, initialize and fill the biases.
    if (bias_term_) {
      vector<int> bias_shape(1, num_output_);
      this->blobs_[1].reset(new Blob(bias_shape));
      shared_ptr > bias_filler(GetFiller(
          this->layer_param_.convolution_param().bias_filler()));
      bias_filler->Fill(this->blobs_[1].get());
    }
  }
  // Propagate gradients to the parameters (as directed by backward pass).
  this->param_propagate_down_.resize(this->blobs_.size(), true);
}

2.Reshape 函数：

template <typename Dtype>
void BaseConvolutionLayer::Reshape(const vector*>& bottom,
      const vector*>& top) {
  CHECK_EQ(4, bottom[0]->num_axes()) << "Input must have 4 axes, "
      << "corresponding to (num, channels, height, width)";//blob是四维数组
  num_ = bottom[0]->num();
  height_ = bottom[0]->height();
  width_ = bottom[0]->width();
  CHECK_EQ(bottom[0]->channels(), channels_) << "Input size incompatible with"
    " convolution kernel.";
  // TODO: generalize to handle inputs of different shapes.
  for (int bottom_id = 1; bottom_id < bottom.size(); ++bottom_id) {
    CHECK_EQ(num_, bottom[bottom_id]->num()) << "Inputs must have same num.";
    CHECK_EQ(channels_, bottom[bottom_id]->channels())
        << "Inputs must have same channels.";
    CHECK_EQ(height_, bottom[bottom_id]->height())
        << "Inputs must have same height.";
    CHECK_EQ(width_, bottom[bottom_id]->width())
        << "Inputs must have same width.";
  }//有多少个bottom 就有多少个top输出，要求每个bottom有相同的shape，因为用的是同一组filter
  // Shape the tops.
  compute_output_shape();//在conv_layer中定义，计算输出feature map 的shape
  for (int top_id = 0; top_id < top.size(); ++top_id) {
    top[top_id]->Reshape(num_, num_output_, height_out_, width_out_);
  }
  if (reverse_dimensions()) {
    conv_in_height_ = height_out_;//根据pad情况计算所得输出top 的height，具体如何计算在conv_layer的compute_output_shape()中定义
    conv_in_width_ = width_out_;
    conv_out_spatial_dim_ = height_ * width_;
  } else {
    conv_in_height_ = height_; //输入bottom 的height
    conv_in_width_ = width_;   //输入bottom 的width
    conv_out_spatial_dim_ = height_out_ * width_out_;
  }
  kernel_dim_ = conv_in_channels_ * kernel_h_ * kernel_w_;//对应一个输出的feature map
  weight_offset_ = conv_out_channels_ * kernel_dim_ / group_ / group_;
  col_offset_ = kernel_dim_ * conv_out_spatial_dim_ / group_;
  output_offset_ = conv_out_channels_ * conv_out_spatial_dim_ / group_;
  // The im2col result buffer will only hold one image at a time to avoid
  // overly large memory usage. In the special case of 1x1 convolution
  // it goes lazily unused to save memory.
  if (reverse_dimensions()) {
    col_buffer_.Reshape(1, kernel_dim_, height_, width_);
  } else {
    col_buffer_.Reshape(1, kernel_dim_, height_out_, width_out_);
  }
  // Set up the all ones "bias multiplier" for adding biases by BLAS
  if (bias_term_) {
    vector<int> bias_multiplier_shape(1, height_out_ * width_out_);
    bias_multiplier_.Reshape(bias_multiplier_shape);
    caffe_set(bias_multiplier_.count(), Dtype(1),
        bias_multiplier_.mutable_cpu_data());
  }
}

3.forward_cpu_gemm 函数：

template type>
void BaseConvolutionLayertype>::forward_cpu_gemm(const Dtype* input,
    const Dtype* weights, Dtype* output, bool skip_im2col) {
  const Dtype* col_buff = input;
  if (!is_1x1_) {
    if (!skip_im2col) {
      conv_im2col_cpu(input, col_buffer_.mutable_cpu_data());
    }
    col_buff = col_buffer_.cpu_data();
  }
  for (int g = 0; g < group_; ++g) {
    caffe_cpu_gemmtype>(CblasNoTrans, CblasNoTrans, conv_out_channels_ /
        group_, conv_out_spatial_dim_, kernel_dim_ / group_,
        (Dtype)1., weights + weight_offset_ * g, col_buff + col_offset_ * g,
        (Dtype)0., output + output_offset_ * g);
  }
}// 实现卷积操作

4.forward_cpu_bias 函数：

template type>
void BaseConvolutionLayertype>::forward_cpu_bias(Dtype* output,
    const Dtype* bicas) {
  caffe_cpu_gemmtype>(CblasNoTrans, CblasNoTrans, num_output_,
      height_out_ * width_out_, 1, (Dtype)1., bias, bias_multiplier_.cpu_data(),
      (Dtype)1., output);
}//卷积后加bias

4.backward_cpu_gemm函数：

template type>
void BaseConvolutionLayertype>::backward_cpu_gemm(const Dtype* output,
    const Dtype* weights, Dtype* input) {
  Dtype* col_buff = col_buffer_.mutable_cpu_data();
  if (is_1x1_) {
    col_buff = input;
  }
  for (int g = 0; g < group_; ++g) {
    caffe_cpu_gemmtype>(CblasTrans, CblasNoTrans, kernel_dim_ / group_,
        conv_out_spatial_dim_, conv_out_channels_ / group_,
        (Dtype)1., weights + weight_offset_ * g, output + output_offset_ * g,
        (Dtype)0., col_buff + col_offset_ * g);
  }
  if (!is_1x1_) {
    conv_col2im_cpu(col_buff, input);
  }计算关于bottom data的导数以便传给下一层

5.weight_cpu_gemm 函数：

template type>
void BaseConvolutionLayertype>::weight_cpu_gemm(const Dtype* input,
    const Dtype* output, Dtype* weights) {
  const Dtype* col_buff = input;
  if (!is_1x1_) {
    conv_im2col_cpu(input, col_buffer_.mutable_cpu_data());
    col_buff = col_buffer_.cpu_data();
  }
  for (int g = 0; g < group_; ++g) {
    caffe_cpu_gemmtype>(CblasNoTrans, CblasTrans, conv_out_channels_ / group_,
        kernel_dim_ / group_, conv_out_spatial_dim_,
        (Dtype)1., output + output_offset_ * g, col_buff + col_offset_ * g,
        (Dtype)1., weights + weight_offset_ * g);
  }
}//计算关于weight的导数用于更新。

6.backward_cpu_bias 函数：

template type>
void BaseConvolutionLayertype>::backward_cpu_bias(Dtype* bias,
    const Dtype* input) {
  caffe_cpu_gemvtype>(CblasNoTrans, num_output_, height_out_ * width_out_, 1.,
      input, bias_multiplier_.cpu_data(), 1., bias);
} 计算关于bias的导数