Caffe support for Deep Compositional Networks with Gaussian Convolution Layer

By constraining filter weights in CNN we can force deep networks to learn compositional hiearhices. Each composition is modeled with a Gaussian component which can also be implemnted as a seperable filter for speedup in forward pass (currently only supported experimantaly on CPU).

Build notes

• Only CMake build supported
• CMake automatically enables C++11 support
• Repository includes full Nvidia CUB library through git-submodule so do not forget to call git submodule update.
• GuassianConv layer supports CPU and GPU, as well as CUDNN engine.

Please cite the following publication for Guassian Convolution Layer:

@inproceedings{Tabernik2016ICPR,
  Author = {Tabernik, Domen and Kristan, Matej and Wyatt, Jeremy and Leonardis, Ale\v{s}}
  booktitle = {International Conference on Pattern Recognition},
  Title = {Towards Deep Compositional Networks},
  Year = {2016}
}

Usage

You can follow the example at examples/cifar10/gauss-cnn/cifar10_gauss_train_test.prototxt.

Enable Guassian Convolution Layer with type: "GaussianConv". You can use following settings in your protobuf definition:

layer {  
  type: "GaussianConv"  
  ...
  # param for Gaussian weight
  param {
    lr_mult: 1
    decay_mult: 0
  }
  # param for Gaussian mean in X direction
  param {
    lr_mult: 50
    decay_mult: 0
  }
  # param for Gaussian mean in Y direction
  param {
    lr_mult: 50
    decay_mult: 0
  }
  # param for Gaussian variance
  param {
    lr_mult: 5
    decay_mult: 0
  }
  # dummy - old depricated paramater
  param {
    lr_mult: 0
    decay_mult: 0
  }
  # bias
  param {
    lr_mult: 1
  }
  
  convolution_param {
    engine: CUDNN # GaussianConv currently does not use CUDNN by default
    
    # standard Convolution layer parameters
    num_output: 96
    kernel_size: 7
    pad: 3
    stride: 1 # ONLY stride=1 curently supported !!
    
    bias_filler {
      type: "constant"
    }
    
    # GaussianConv layer specific parameters
    number_gauss: 2 # in each dimension (threfore 4 gauss components per kernel in this case)
    
    # Initialization of Gaussian weights
    weight_filler {
      type: "gaussian"
      std: 0.01
      mean: 0
    }
    # Initialization of Gaussian variance
    sigma_filler {
      type: "constant"
      value: 0.8     
    }    
    
    # Explicitly set other GaussianConv layer related parameters since they works better then default ones:
    gmm_weight_normalization: false 
    gmm_gauss_normalization: true
    gmm_mean_iteration_step: 1
    gmm_sigma_iteration_step: 1
    gmm_sigma_lower_bound: 0.5
    gmm_component_border_bound: 1.5
  }
}

Some other parameters can be found in caffe.proto:

  optional bool gmm_weight_normalization = 123 [default = true]; // Should weights of GMM be normalized (i.e. should sum to 1)
  optional bool gmm_gauss_normalization = 124 [default = true]; // Should gaussian GMM be normalized or not  (i.e. area under gauss should sum to 1)
  optional uint32 gmm_mean_iteration_step = 125 [default = 1]; // step between optimization iteration of mean in gaussian
  optional uint32 gmm_sigma_iteration_step = 126 [default = 1]; // step between optimization iteration  of sigma in gaussian
  optional float gmm_component_border_bound = 127 [default = 0]; // how close to kernel border is gaussian component allowed to be
  optional float gmm_sigma_lower_bound = 128 [default = 0.1]; // lower bound for sigma
  optional bool gmm_square_gauss_normalization = 129 [default = true]; // normalize using sum of squares 
  optional bool gmm_seperable_forward_pass = 130 [default = false]; // utilize seperable kernels to speed-up forward-pass
  optional bool gmm_use_old_cudnn = 131 [default = false]; // use old version of cudnn implmentation (uses a lot of memory)

---

Caffe

Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by the Berkeley Vision and Learning Center (BVLC) and community contributors.

Check out the project site for all the details like

• DIY Deep Learning for Vision with Caffe
• Tutorial Documentation
• BVLC reference models and the community model zoo
• Installation instructions

and step-by-step examples.

Please join the caffe-users group or gitter chat to ask questions and talk about methods and models. Framework development discussions and thorough bug reports are collected on Issues.

Happy brewing!

License and Citation

Caffe is released under the BSD 2-Clause license. The BVLC reference models are released for unrestricted use.

Please cite Caffe in your publications if it helps your research:

@article{jia2014caffe,
  Author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
  Journal = {arXiv preprint arXiv:1408.5093},
  Title = {Caffe: Convolutional Architecture for Fast Feature Embedding},
  Year = {2014}
}