Documentation

Convolution (DepthwiseConv, ConvolutionDepthwise)

Input

  • x:
    • Required: yes
    • Type: float16
    • Description: Convolves input x.
    • Restrictions: none
  • filter:
    • Required: yes
    • Type: float16
    • Description: Both DepthwiseConv and ConvolutionDepthwise converted into a Convolution operator before use, for example, ConvolutionDepthwise.
      layer {
   name: "resx1_conv2"
   type: "ConvolutionDepthwise"
   bottom: "resx1_conv1"
   top: "resx1_conv2"
   convolution_param {
     num_output: 54
     kernel_size: 3
     stride: 2
     pad: 1
     bias_term: false
     weight_filler {
       type: "msra"
     }
   }
 }

      Edit the definition to:

      layer {
   name: "resx1_conv2"
   type: "Convolution"
   bottom: "resx1_conv1"
   top: "resx1_conv2"
   convolution_param {
     num_output: 54
     group: 54
     kernel_size: 3
     stride: 2
     pad: 1
     bias_term: false
     weight_filler {
       type: "msra"
     }
   }
 }

      If the group attribute is included in DepthwiseConv or ConvolutionDepthwise, change type to Convolution. Otherwise, add the group attribute: group==num_output (Note that group, out_channel, and input_channel must be consistent).

      The op name and op type in the modified .prototxt model must be the same as those in the .caffemodel weight file (case sensitive).

    • Restriction:
      1. The value range of Filter_W and Filter_h is 1 to 255.
  • bias:
    • Required: no
    • Type: float16
    • Description: bias. If None, applies no bias.
    • Restrictions: none

Attribute

  • pad:
    • Required: no
    • Type: ListInt
    • Description: padding size (default = 0). pad and pad_h/pad_w are mutually exclusive. pad is a list of length up to 2.
    • Restriction: pad < filter.size
  • pad_h:
    • Required: no
    • Type: int
    • Description: padding size (default = 0) of the high axis. pad and pad_h/pad_w are mutually exclusive. pad is a list of length up to 2.
    • Restriction: pad < filter.size
  • pad_w:
    • Required: no
    • Type: int
    • Description: padding size (default = 0) of the width axis. pad and pad_h/pad_w are mutually exclusive. pad is a list of length up to 2.
    • Restriction: pad < filter.size
  • stride:
    • Required: no
    • Type: ListInt
    • Description: stride size (default = 1). stride and stride_h/stride_w are mutually exclusive. stride is a list of length up to 2.
    • Restriction: 1 <= stride <= 63
  • stride_h:
    • Required: no
    • Type: int
    • Description: stride size (default = 1). stride and stride_h/stride_w are mutually exclusive. stride is a list of length up to 2.
    • Restriction: 1 <= stride <= 63
  • stride_w:
    • Required: no
    • Type: int
    • Description: stride size (default = 1). stride and stride_h/stride_w are mutually exclusive. stride is a list of length up to 2.
    • Restriction: 1 <= stride <= 63
  • dilation:
    • Required: no
    • Type: ListInt
    • Description: factor used to dilate the filter (default = 1). It is a list of length up to 2.
    • Restriction: Must be in the range [1, 255] and (kernel – 1) x dilation + 1 < 256.
  • group:
    • Required: no
    • Type: int
    • Description: none
    • Restriction: group is divisible by channel.

Output

y:
  • Required: yes
  • Type: float16
  • Description: none
  • Restrictions: none