Learning without Forgetting

Created by Zhizhong Li and Derek Hoiem at University of Illinois, Urbana Champaign.

Project webpage.

Please contact Zhizhong Li for any questions regarding this repository.

Note: this repository is implemented using MatConvNet. For a PyTorch version, my labmate Arun Mallya's PackNet repository has an LwF implementation. Please cite both papers if you decide to use his code instead.

Introduction

Learning without Forgetting aims at adding new capabilities (new tasks) to an existing Convolutional Neural Network, sharing representation with the original capabilities (old tasks), while allowing for adjusting the shared representation to adapt for both tasks without using the original training data.

The resulting network outperforms merely fine-tuning on the new tasks completely, suggesting it being a better practice than the widely-used method. It also outperforms feature extraction, but only on the new task performance.

A more detailed abstract can be found in our paper. The software aims at replicating our method. We use the MatConvNet library.

Usage

The code is tested on Linux (64 bit Arch Linux 4.4.5-1-ARCH)

Prerequisites

0. Matlab (tested against R2015b)
1. MatConvNet v1.0-beta13
2. For GPU support, we use TITAN X with CUDA 7.5.

Installation

0. Compile MatConvNet accordingly using their installation guide.

1. Download the datasets you want to run experiments on, and place appropriately. (See Dataset Section)

2. Download AlexNet/VGG pretrained on ImageNet from MatConvNet, preferrably from v1.0-beta13 model base. (See Dataset Section for placement)

3. Adjust getPath.m for your path setup for the datasets and models. (Usually just the p.path_dsetroot and p.path_exeroot value)

4. In matlab terminal, cd to the MatConvNet folder, run:

    run matlab/vl_setupnn
    addpath <path-to-LwF> % where you put <th></th>is repository
    addpath <path-to-LwF>/snippets % the subdirectories
   

Experiments

Use gridSearchPASCAL(mode) as the entry point to this repository. Use different strings for mode to perform different experiments. See gridSearchPASCAL.m for details.

Datasets and Models

The paper uses the following datasets.

ImageNet

We use the ILSVRC 2012 version. Place the folders ILSVRC2012_devkit_t12/ and ILSVRC2012_img_*/ under the <imgroot>/ILSVRC2012/ folder. A few jpeg images has to be manually converted from CMYK to RGB.

The pre-trained models, imagenet-caffe-alex.mat and imagenet-vgg-verydeep-16.mat obtained from MatConvNet, are placed directly inside the <path-to-LwF> folder.

Places2

In this work, we use Places2 with 401 classes, which was used in the ILSVRC2015 taster challenge, so if you have the dataset available, you can set it up as follows.

Place the train/, val/, test/ folders under <imgroot>/Places2/. A train.txt and val.txt should also be placed here, with each line defining one image's relative path and the label:

(train.txt)
a/abbey/00000000.jpg 0
a/abbey/00000001.jpg 0
a/abbey/00000002.jpg 0
a/abbey/00000003.jpg 0
...

(val.txt)
Places2_val_00000001.jpg 330
Places2_val_00000002.jpg 186
Places2_val_00000003.jpg 329
Places2_val_00000004.jpg 212

Unfortunately, this Places2 version has become obsolete after our paper submission. Places365 now replaces it as a better dataset, but we have not yet adjusted our code accordingly.

The provided CNN once available from their website should be fine-tuned to the downscaled dataset due to stability issues, and converted to MatConvNet. We provide our fine-tuned conversion of Places2 CNN. This should be placed in the <path-to-LwF>/convertcaffe/ folder using filename alexnet_places2_iter_158165_convert.mat.

PASCAL VOC

We use the VOC 2012 version. Place the folders such as JPEGImages/ and ImageSets/ under the <imgroot>/VOC2012/ folder.

MIT indoor scene

The MIT indoor scene dataset. Folder Images/ and files trainImages.txt, testImages.txt are to be placed under <imgroot>/MIT_67_indoor/. NOTE: please rename the .txt files (no front caps).

A classes.txt should be generated using shell command line:

cd Images
ls -d * > ../classes.txt

within the folder.

Some of the images (see MIT_wrongformat.txt) may be of erroneous format and should be saved as actual *.jpg format. You can use e.g. an image editing software to convert them.

Caltech-UCSD-Birds

We use the CUB-200-2011 version (the larger one). Place folders such as images/ and files such as classes.txt into the <imgroot>/CUB_200_2011/ folder.

Copy the provided snippets/convert_CUB_old.py script into the folder, and execute it in shell python convert_CUB_old to generate the train/test image lists. You need Python for this step.

MNIST

Please place the four files provided by MNIST into <imgroot>/MNIST/, and unzip them.

Notes

0. Due to our implementation, the efficiency for LwF here is actually similar to joint training instead of being better; theoretically it can be optimized by sharing fwd/bkwd pass of the shared layers across tasks using e.g. the dagnn for MatConvNet 0.17 upwards, or other libraries such as tensorflow.

Citation

Please cite our paper if you use our work in your research.

@inproceedings{li2016learning,
    title={Learning Without Forgetting},
    author={Li, Zhizhong and Hoiem, Derek},
    booktitle={European Conference on Computer Vision},
    pages={614--629},
    year={2016},
    organization={Springer}
}

Acknowledgements

The MNIST helper files are generously provided by Stanford UFLDL.

Some of our code are based on MatConvNet code and VOC devkit.

License

This software package is freely available for research purposes. Please check the LICENSE file for details.