Attention Transfer

PyTorch code for "Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer" https://arxiv.org/abs/1612.03928
Conference paper at ICLR2017: https://openreview.net/forum?id=Sks9_ajex

What's in this repo so far: * Activation-based AT code for CIFAR-10 experiments * Code for ImageNet experiments (ResNet-18-ResNet-34 student-teacher)

Coming: * grad-based AT * Scenes and CUB activation-based AT code * Pretrained with activation-based AT ResNet-18

The code uses PyTorch https://pytorch.org. Note that the original experiments were done using torch-autograd, we have so far validated that CIFAR-10 experiments are exactly reproducible in PyTorch, and are in process of doing so for ImageNet (results are very slightly worse in PyTorch, due to hyperparameters).

bibtex:

@inproceedings{Zagoruyko2017AT,
    author = {Sergey Zagoruyko and Nikos Komodakis},
    title = {Paying More Attention to Attention: Improving the Performance of
             Convolutional Neural Networks via Attention Transfer},
    booktitle = {ICLR},
    url = {https://arxiv.org/abs/1612.03928},
    year = {2017}}

Requirements

First install PyTorch, then install torchnet:

pip install git+https://github.com/pytorch/tnt.git@master

Then install OpenCV with Python bindings (e.g. conda install -c menpo opencv3), and other Python packages:

pip install -r requirements.txt

Experiments

CIFAR-10

This section describes how to get the results in the table 1 of the paper.

First, train teachers:

python cifar.py --save logs/resnet_40_1_teacher --depth 40 --width 1
python cifar.py --save logs/resnet_16_2_teacher --depth 16 --width 2
python cifar.py --save logs/resnet_40_2_teacher --depth 40 --width 2

To train with activation-based AT do:

python cifar.py --save logs/at_16_1_16_2 --teacher_id resnet_16_2_teacher --beta 1e+3

To train with KD:

python cifar.py --save logs/kd_16_1_16_2 --teacher_id resnet_16_2_teacher --alpha 0.9

We plan to add AT+KD with decaying beta to get the best knowledge transfer results soon.

ImageNet

Pretrained model

We provide ResNet-18 pretrained model with activation based AT:

| Model | val error | |:------|:---------:| |ResNet-18 | 30.4, 10.8 | |ResNet-18-ResNet-34-AT | 29.3, 10.0 |

Download link: https://www.dropbox.com/s/z092wmrgyqn4ys5/resnet-18-at-export.hkl?dl=0

Model definition: https://github.com/szagoruyko/functional-zoo/blob/master/resnet-18-at-export.ipynb

Convergence plot:

Train from scratch

Download pretrained weights for ResNet-34 (see also functional-zoo for more information):

wget https://s3.amazonaws.com/pytorch/h5models/resnet-34-export.hkl

Prepare the data following fb.resnet.torch and run training (e.g. using 2 GPUs):

python imagenet.py --imagenetpath ~/ILSVRC2012 --depth 18 --width 1 
                   --teacher_params resnet-34-export.hkl --gpu_id 0,1 --ngpu 2 
                   --beta 1e+3