Caffe

Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by Berkeley AI Research (BAIR)/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

• BAIR reference models and the community model zoo

• Installation instructions

and step-by-step examples.

Custom distributions

• Intel Caffe (Optimized for CPU and support for multi-node), in particular Xeon processors (HSW, BDW, SKX, Xeon Phi).

• OpenCL Caffe e.g. for AMD or Intel devices.

• Windows Caffe

Community

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 BAIR/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}
}

P-CapsNet

This part is for the paper Capsule Networks without Routing procedures The primary code can been seen at tensor_layer.cpp. For cuda-accelerated version, please see capsule_conv_layer.cu. Note that the current acclerated version only supports the 2D tensor case. Below is an example of adding a capsule layer of P-CapsNets.

layer {
  name: "conv1"
  type: "Tensor"
  bottom: "data"
  top: "conv1"
  capsule_conv_param {
    weight_filler {
      type: "msra"
    }
    stride: 2
    kh: 3
    kw: 3
    input_capsule_num: 1
    output_capsule_num: 1
    output_capsule_shape {
      dim: 1
      dim: 1
      dim: 32
    }
    input_h: 28
    input_w: 28
    input_capsule_shape {
      dim: 1
      dim: 1
      dim: 1
    }
    bias_term: false
  }
}

To see a full example, please check new_capsule_train.prototxt. To train a sample P-CapsNet model on MNIST, run

sh examples/mnist/train_new_capsule.sh

Other versions of CapsNets

This part implements the idea in the paper [Generalized Capsule Networks with Trainable Routing Procedure].

G-CapsNet makes the rouing prodedure become part of the training process, it supports both "full-connected" version of CapsNet as well as "convolutional" version of CapsNet.

G-CapsNet is tested on MNIST and achieves comparible performance as normal CapsNet.

Guidline

To add a full-connected capsule layer, you need to add one 'CapsuleTransform' layer and one 'CapsuleRouting' layer,

layer {
  name: "cap_transform"
  type: "CapsuleTransform"
  bottom: "squash1"
  top: "cap_transform"
  capsule_transform_param {
    input_capsule_dim: 8
    output_capsule_dim: 16
    output_capsule_num: 10
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "cap_routing"
  type: "CapsuleRouting"
  bottom: "cap_transform"
  top: "cap_routing"
  capsule_routing_param {
    input_capsule_num: 1152
    output_capsule_dim: 16
    output_capsule_num: 10
    weight_filler {
      type: "xavier"
    }
  }
}

To add a convolutional capsule layer, you need to add one 'CapsuleConvTransform' layer and one 'CapsuleConvRouting' layer,

layer {
  name: "cap_conv_transform"
  type: "CapsuleConvTransform"
  bottom: "squash1"
  top: "cap_conv_transform"
  capsule_conv_transform_param {
    input_h: 14
    input_w: 14
    stride: 2
    kh: 4
    kw: 4
    input_capsule_num: 2
    output_capsule_num: 32
    input_capsule_shape {
      dim: 4  
      dim: 4
    }
    output_capsule_shape {
      dim: 4  
      dim: 4
    }
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "cap_conv_routing"
  type: "CapsuleConvRouting"
  bottom: "cap_conv_transform"
  top: "cap_conv_routing"
  capsule_conv_routing_param {
    input_h: 14
    input_w: 14
    stride: 2
    kh: 4
    kw: 4
    input_capsule_num: 2
    output_capsule_num: 32
    output_capsule_shape {
      dim: 4  
      dim: 4
    }
    weight_filler {
      type: "xavier"
    }
  }
}

To convert normal activations to 'capsules' across feature maps, please use 'CapsuleReshape' layer. Othewise, use 'Reshape' layer.

layer {
  name: "reshape"
  type: "CapsuleReshape"
  bottom: "conv2"
  top: "conv2_reshaped"
  capsule_reshape_param {
    capsule_dim: 8
  }
}


layer {
  name: "reshape"
  type: "Reshape"
  bottom: "conv2"
  top: "conv2_reshaped"
  reshape_param {
    shape {
      dim: 0  
      dim: -1
      dim: 8 
    }
  }
}

Training on MNIST

To train a "full-connected" version of G-CapsNet, run

sh examples/mnist/train_full_connected_capsule.sh

To train a "convolutioal" version of G-CapsNet, run

sh examples/mnist/train_g_conv_capsule.sh

To train a baseline (Please read the paper for more details), run

sh examples/mnist/train_baseline.sh

To train the CapsNet with dynamic routing in the orignal paper, run

sh examples/mnist/train_dr.sh

Training on CIFAR10

To train a "full-connected" version of G-capsNet, run

sh examples/cifar10/train_full_connected.sh

To train a baseline (Please read the paper for more details), run

sh examples/cifar10/train_baseline.sh

To train a multi_layer CapsNet, run

sh examples/cifar10/train_multi_capsule_layer.sh

To train the CapsNet with dynamic routing in the orignal paper, run

sh examples/cifar10/train_dr.sh

Try capsule deconvolutional layer

sh examples/mnist/train_capsule_deconv.sh