Locally Scale-Invariant ConvNet Caffe Implementation

This packages implements the scale-invariant ConvNet used in our NIPS 2014 Deep Learning & Representation Workshop paper.

It's based on BVLC's Caffe, final merge with BVLC/master was on Oct 20th 2014.

Installation

Requires all of Caffe's prerequisite packages. Compile as you would compile Caffe i.e. have the right Makefile.config and

make all
make test
make runtest

Changes to BVLC/Caffe

The major additions are:

1. util/transformation.(hpp/cpp/cu) Misc functions needed to apply image transformation using NN or bilinear interpolation.
2. ticonv_layer.cpp TIConvolutionLayer a wrapper around UpsamplingLayer, tiedconv_layer and DownpoolLayer. This is what you can use instead of convolution layer to use SI-Conv layer.
3. up_layer.cpp Contains UpsamplingLayer which applies user specified interpolations to the bottom blob. i.e. TransformationLayer.
4. downpool_layer.cpp Contains DownpoolLayer, which is almost the same as UpsamplingLayer, but after applying transformations, crops the inputs into a canonical shape and does max-pooling over all transformations.
5. tiedconv_layer.cpp Convolutional layer that can apply convolution to multiple inputs using the same weight. Very close to current (Jan 2015) Caffe's ConvolutionalLayer except that the input size can vary.
6. util/imshow.(hpp/cpp) (not necessary), used for debugging images in C++ using openCV behaves like matlab's imshow and montage.
7. And all the misc changes needed to adapt the changes into the rest of the code.

All major changes are implemented in both CPU and GPU with tests.

Technical Note: since CUDA's atomicAdd, required in backprop fo transformation layer isn't available for doubles, this code only runs for float instantiation of Caffe (which shouldn't be a problem since default Caffe runs in float). But because of that, all explicit instantiation for doubles are commented out.

How to use SI-Conv Layer instead of Conv Layer

In your protofiles, replace the type of the layer from CONVOLUTION to TICONV and add transformations that you want to apply to this layer. Note that TICONV layer assumes that the first transformation is always identity and is the canonical size.

Example:

A Convolution Layer:

 layers {
   name: "conv1"
   type: CONVOLUTION
   bottom: "data"
   top: "conv1"
   blobs_lr: 1.
   blobs_lr: 2.
   weight_decay: 1.
   weight_decay: 0.
   convolution_param {
	 num_output: 36
	 kernel_size: 7
	 stride: 1
	 weight_filler {
	   type: "gaussian"
	   std: 0.01
	 }
	 bias_filler {
	   type: "constant"
	 }
   }
 }

A Scale-Invariant Convolution Layer:

 layers {
   name: "conv1"
   type: CONVOLUTION
   bottom: "data"
   top: "conv1"
   blobs_lr: 1.
   blobs_lr: 2.
   weight_decay: 1.
   weight_decay: 0.
   convolution_param {
	 num_output: 36
	 kernel_size: 7
	 stride: 1
	 weight_filler {
	   type: "gaussian"
	   std: 0.01
	 }
	 bias_filler {
	   type: "constant"
	 }		 
   }
   transformations {}
   transformations { scale: 0.63 }
   transformations { scale: 0.7937 }
   transformations { scale: 1.2599 }
   transformations { scale: 1.5874 }
   transformations { scale: 2 }
 }

Transformations parameter accepts parameters:

• scale: scale-factor
• rotation: rotation in degrees
• border: border option similar to matlab {0=crop (default), 1=clamp, 2=reflect}
• interp: interpolation option {0=Nearest Neighbor, 1=Bilinear (default)} So it can handle transformations other than scale as well. Sample protos can be found in models/sicnn/protos.

Replicating the results on paper

Get the MNIST-Scale train/test folds in hdf5 format (mean subtracted) from here and unzip it in data/mnist or from this directory:

cd data/mnist
wget http://angjookanazawa.com/sicnn/mnist-sc-table1.tar.gz
tar vxzf mnist-sc-table1.tar.gz

models/sicnn has sample prototxt for vanila convnet, hierarchical convnet of Farabet et al [1] and si-convnet used in ther paper for split 1. From this directory each one can be run with:

./train_all.sh cnn
./train_all.sh farabet
./train_all.sh sicnn

Note: There was a minor bug in the transformation code which further improved SI-ConvNet mean error on the 6 train/test fold from 3.13% to 2.93%. The performance on the other two models stayed the same. On this split 1, this SI-ConvNet should get something like 2.91% error.

Citing

If you find any part of this code useful, please consider citing:

@misc{kanazawa14,
author    = {Angjoo Kanazawa and Abhishek Sharma and David W. Jacobs},
title     = {Locally Scale-Invariant Convolutional Neural Networks},
year      = {2014},
url       = {http://arxiv.org/abs/1412.5104},
Eprint = {arXiv:1412.5104}
}

as well as the Caffe Library.

@misc{Jia13caffe,
	Author = {Yangqing Jia},
	Title = { {Caffe}: An Open Source Convolutional Architecture
	for Fast Feature Embedding},
	Year = {2013},
	Howpublished = {\url{http://caffe.berkeleyvision.org/}}
}

Questions, comments, bug report

Please direct any questions, comment, bug report etc to kanazawa[at]umiacs[dot]umd[dot]edu.

[1] Clement Farabet, Camille Couprie, Laurent Najman and Yann LeCun, "Learning Hierarchical Features for Scene Labeling", IEEE PAMI 2013.