Oriented Response Networks

[Home] [Project] [Paper] [Supp] [Poster]

Torch Implementation

The torch branch contains:

• the official torch implementation of ORN.
• the MNIST-Variants demo.

Please follow the instruction below to install it and run the experiment demo.

Prerequisites

• Linux (tested on ubuntu 14.04LTS)
• NVIDIA GPU + CUDA CuDNN (CPU mode and CUDA without CuDNN mode are also available but significantly slower)
• Torch7

Getting started

You can setup everything via a single command wget -O - https://git.io/vHCMI | bash or do it manually in case something goes wrong:

1. install the dependencies (required by the demo code):

   • torchnet: luarocks install torchnet
   • optnet: luarocks install optnet

2. clone the torch branch:

   # git version must be greater than 1.9.10
   git clone https://github.com/ZhouYanzhao/ORN.git -b torch --single-branch ORN.torch
   cd ORN.torch
   export DIR=$(pwd)

3. install ORN:

   cd $DIR/install
   # install the CPU/GPU/CuDNN version ORN.
   bash install.sh

4. unzip the MNIST dataset:

   cd $DIR/demo/datasets
   unzip MNIST

5. run the MNIST-Variants demo:

   cd $DIR/demo
   # you can modify the script to test different hyper-parameters
   bash ./scripts/Train_MNIST.sh

Trouble shooting

If you run into 'cudnn.find' not found, update Torch7 to the latest version via cd <TORCH_DIR> && bash ./update.sh then re-install everything.

More experiments

CIFAR 10/100

You can train the OR-WideResNet model (converted from WideResNet by simply replacing Conv layers with ORConv layers) on CIFAR dataset with WRN.

dataset=cifar10_original.t7 model=or-wrn widen_factor=4 depth=40 ./scripts/train_cifar.sh

With exactly the same settings, ORN-augmented WideResNet achieves state-of-the-art result while using significantly fewer parameters.

Table.1 Test error (%) on CIFAR10/100 dataset with flip/translation augmentation)

ImageNet

The effectiveness of ORN is further verified on large scale data. The OR-ResNet-18 model upgraded from ResNet-18 yields significant better performance when using similar parameters.

Table.2 Validation error (%) on ILSVRC-2012 dataset.

You can use facebook.resnet.torch to train the OR-ResNet-18 model from scratch or finetune it on your data by using the pre-trained weights.

-- To fill the model with the pre-trained weights:
model = require('or-resnet.lua')({tensorType='torch.CudaTensor', pretrained='or-resnet18_weights.t7'})

A more specific demo notebook of using the pre-trained OR-ResNet to classify images can be found here.

PyTorch Implementation

The pytorch branch contains:

• the official pytorch implementation of ORN (alpha version supports 1x1/3x3 ARFs with 4/8 orientation channels only).
• the MNIST-Variants demo.

Please follow the instruction below to install it and run the experiment demo.

Prerequisites

• Linux (tested on ubuntu 14.04LTS)
• NVIDIA GPU + CUDA CuDNN (CPU mode and CUDA without CuDNN mode are also available but significantly slower)
• PyTorch

Getting started

1. install the dependencies (required by the demo code):

   • tqdm: pip install tqdm
   • pillow: pip install Pillow

2. clone the pytorch branch:

   # git version must be greater than 1.9.10
   git clone https://github.com/ZhouYanzhao/ORN.git -b pytorch --single-branch ORN.pytorch
   cd ORN.pytorch
   export DIR=$(pwd)

3. install ORN:

   cd $DIR/install
   bash install.sh

4. run the MNIST-Variants demo:

   cd $DIR/demo
   # train ORN on MNIST-rot
   python main.py --use-arf
   # train baseline CNN
   python main.py

Caffe Implementation

The caffe branch contains:

• the official caffe implementation of ORN (alpha version supports 1x1/3x3 ARFs with 4/8 orientation channels only).
• the MNIST-Variants demo.

Please follow the instruction below to install it and run the experiment demo.

Prerequisites

• Linux (tested on ubuntu 14.04LTS)
• NVIDIA GPU + CUDA CuDNN (CPU mode and CUDA without CuDNN mode are also available but significantly slower)
• Caffe

Getting started

1. install the dependency (required by the demo code):

   • idx2numpy: pip install idx2numpy

2. clone the caffe branch:

   # git version must be greater than 1.9.10
   git clone https://github.com/ZhouYanzhao/ORN.git -b caffe --single-branch ORN.caffe
   cd ORN.caffe
   export DIR=$(pwd)

3. install ORN:

   # modify Makefile.config first
   # compile ORN.caffe
   make clean && make -j"$(nproc)" all

4. run the MNIST-Variants demo:

   cd $DIR/examples/mnist
   bash get_mnist.sh
   # train ORN & CNN on MNIST-rot
   bash train.sh

Note

Due to implementation differences,

• upgrading Conv layers to ORConv layers can be done by adding an orn_param
• num_output of ORConv layers should be multipied by nOrientation of ARFs

Example:

layer {
  type: "Convolution"
  name: "ORConv" bottom: "Data" top: "ORConv"
  # add this line to replace regular filters with ARFs
  orn_param {orientations: 8}
  param { lr_mult: 1 decay_mult: 2}
  convolution_param {
    # this means 10 ARF feature maps
    num_output: 80
    kernel_size: 3
    stride: 1
    pad: 0
    weight_filler { type: "msra"}
    bias_filler { type: "constant" value: 0}
  }
}

Check the MNIST demo prototxt (and its visualization) for more details.

Citation

If you use the code in your research, please cite:

@INPROCEEDINGS{Zhou2017ORN,
    author = {Zhou, Yanzhao and Ye, Qixiang and Qiu, Qiang and Jiao, Jianbin},
    title = {Oriented Response Networks},
    booktitle = {CVPR},
    year = {2017}
}