antialiased-cnns_Caffe

This repository contains a Caffe implementation of the paper Making Convolutional Networks Shift-Invariant Again.

antialiased-cnns

Anti-aliasing common downsampling layers

An example on max-pooling

BlurPool

BlurPool combine blur filter and subsample, then blur filter use gaussian blur in antialiased-cnns.

Caffe implementation

Subsample

Provide subsample layer to support antialiased-cnns.
Forward:

Backward:

Gaussian blur

Gaussian blur use Deep Separable Convolution whit gauss initialization, gaussian kernel obeys std=0.1, and let lr_mult = 0 ,decay_mult = 0，fixed gaussian kernel weight.

Anti-aliasing common downsampling layers in caffe

• MaxPool(stride = 2)

layer {
  bottom: "bottom"
  top: "max_pool"
  name: "max_pool"
  type: "Pooling"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 1
    pad:1
  }
}
layer {
  bottom: "max_pool"
  top: "blurconv"
  name: "blurconv"
  type: "ConvolutionDepthwise"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 0
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
  }
}
layer {
  bottom: "blurconv"
  top: "subsample"
  name: "subsample"
  type: "Subsample"
  subsample_param {
    kernel_size: 2
    stride: 2
  }
}

• Conv(stride = 2)+ReLU

layer {
  bottom: "bottom"
  top: "conv"
  name: "conv"
  type: "Convolution"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  bottom: "conv"
  top: "conv"
  name: "relu"
  type: "ReLU"
}
layer {
  bottom: "conv"
  top: "blurconv"
  name: "blurconv"
  type: "ConvolutionDepthwise"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
  }
}
layer {
  bottom: "blurconv"
  top: "subsample"
  name: "subsample"
  type: "Subsample"
  subsample_param {
    kernel_size: 2
    stride: 2
  }
} 

• AvePool(stride = 2)

layer {
  bottom: "bottom"
  top: "blurconv"
  name: "blurconv"
  type: "ConvolutionDepthwise"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.1
    }
  }
}
layer {
  bottom: "blurconv"
  top: "subsample"
  name: "subsample"
  type: "Subsample"
  subsample_param {
    kernel_size: 2
    stride: 2
  }
}

note

• let pad = (kernel_size-1)/2 in ConvolutionDepthwise layer.
• BlurPool should be after ReLU layer, even if there is batchnormal layer between convolution layer and ReLU layer.

Usage

Prerequisites

Caffe
CUDA8.0
cudnn5.0

How to build

• Add Subsample layer
  Modify caffe.proto like this：
  Add option in LayerParameter:

message LayerParameter {
optional SubsampleParameter subsample_param = Your last ID;
}

Add message at the end:

message SubsampleParameter {
  // Pad, kernel size, and stride are all given as a single value for equal
  // dimensions in height and width or as Y, X pairs.
  optional uint32 pad = 3 [default = 0]; // The padding size (equal in Y, X)
  optional uint32 pad_h = 8 [default = 0]; // The padding height
  optional uint32 pad_w = 9 [default = 0]; // The padding width
  optional uint32 kernel_size = 1; // The kernel size (square)
  optional uint32 kernel_h = 4; // The kernel height
  optional uint32 kernel_w = 5; // The kernel width
  optional uint32 stride = 2 [default = 1]; // The stride (equal in Y, X)
  optional uint32 stride_h = 6; // The stride height
  optional uint32 stride_w = 7; // The stride width
}

Add subsample_layer.hpp, subsample_layer.cpp and subsample_layer.cu to the specified folder.

• Add Deep Separable Convolution layer

Deep separable convolution does not need to be declared in caffe.proto.
Add conv_dw_layer.hpp, conv_dw_layer.cpp and conv_dw_layer.cu to the specified folder. Deep Separable Convolution layer fork from Caffe sp2823.

How to use

Reference in vgg16_antialiase_train.prototxt,resnet18_antialiase_train.prototxt and Anti-aliasing common downsampling layers in caffe.

Reference

• antialiased-cnns