Caffe网络中Conv层和 Pooling层的输出尺寸计算

@Caffe网络中Conv层和 Pooling层的输出尺寸计算

• 根据Caffe深度框架的源码分析及实际网络层输出，获得Conv层和Pooling的输出尺寸计算公式，如下：

Conv层

$Output_{size}= (Input_{size} + Padding*2 - Kernel_{size})/Stride + 1$

Pooling层

$Output_{size}= Ceil[(Input_{size} + Padding*2 - Kernel_{size})/Stride] + 1, 其中，Ceil[1.1] = 2$

效果展示图

Caffe网络可视化 (将下列代码放置在左侧编辑栏里，Shift+Enter)

# Enter your network definition here.
# Use Shift+Enter to update the visualization.
name: "Caffe-Head"
layer {
    name: "data"
    type: "Data"
    top: "data"
    transform_param {
        crop_size: 224
    }
}
 
layer {
    bottom: "data"
    top: "conv"
    name: "conv"
    type: "Convolution"
    convolution_param {
        num_output: 64
        kernel_size: 7
        pad: 3
        stride: 2
    }
}
 
layer {
    bottom: "conv"
    top: "pool"
    name: "pool"
    type: "Pooling"
    pooling_param {
        kernel_size: 3
        stride: 2
        pool: MAX
    }
}

源码分析

Conv层

// ./src/caffe/layers/conv_layer.cpp
void ConvolutionLayer<Dtype>::compute_output_shape() {
  const int* kernel_shape_data = this->kernel_shape_.cpu_data();
  const int* stride_data = this->stride_.cpu_data();
  const int* pad_data = this->pad_.cpu_data();
  const int* dilation_data = this->dilation_.cpu_data();
  this->output_shape_.clear();
  for (int i = 0; i < this->num_spatial_axes_; ++i) {
    // i + 1 to skip channel axis
    const int input_dim = this->input_shape(i + 1);
    const int kernel_extent = dilation_data[i] * (kernel_shape_data[i] - 1) + 1;
    /* 计算输出尺寸 */
    const int output_dim = (input_dim + 2 * pad_data[i] - kernel_extent)  / stride_data[i] + 1;
    this->output_shape_.push_back(output_dim);
  }
}
...

Pooling层

ceil函数：同上述公式中的Ceil。

// 	./src/caffe/layers/pooling_layer.cpp
void PoolingLayer<dtype>::Reshape(const vector<blob<dtype>*>& bottom, const vector<blob<dtype>*>& top) {
	      /*检查输入图像的blob轴的个数, (num, channels, height, width)表示图像有这4个轴*/
	  CHECK_EQ(4, bottom[0]->num_axes()) << "Input must have 4 axes, " << "corresponding to (num, channels, height, width)";
	  channels_ = bottom[0]->channels();  /*获得图像通道数*/
	  height_ = bottom[0]->height(); /*获得图像高*/
	  width_ = bottom[0]->width();   /*获得图像宽*/
	  if (global_pooling_) {
	    kernel_h_ = bottom[0]->height();
	    kernel_w_ = bottom[0]->width();
	  }
	  /*计算图像池化后的宽高: ceil */ 
	  pooled_height_ = static_cast<int>(ceil(static_cast<float>(height_ + 2 * pad_h_ - kernel_h_) / stride_h_)) + 1;
	  pooled_width_ = static_cast<int>(ceil(static_cast<float>(width_ + 2 * pad_w_ - kernel_w_) / stride_w_)) + 1;
	  ...
  }

总结

• Conv层和Pooling的输出尺寸计算公式不一样。