Stop Thinking, Just Do!

Sung-Soo Kim's Blog

Attention Transfer


25 January 2017

Article Source

Attention Transfer


Attention plays a critical role in human visual experience. Furthermore, it has recently been demonstrated that attention can also play an important role in the context of applying artificial neural networks to a variety of tasks from fields such as computer vision and NLP. In this work we show that, by properly defining attention for convolutional neural networks, we can actually use this type of information in order to significantly improve the performance of a student CNN network by forcing it to mimic the attention maps of a powerful teacher network. To that end, we propose several novel methods of transferring attention, showing consistent improvement across a variety of datasets and convolutional neural network architectures.


PyTorch code for “Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer
The paper is under review as a conference submission at ICLR2017:

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)


  • grad-based AT
  • Scenes and CUB activation-based AT code
  • Pretrained with activation-based AT ResNet-18

The code uses PyTorch 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).


    author = {Sergey Zagoruyko and Nikos Komodakis},
    title = {Paying More Attention to Attention: Improving the Performance of
             Convolutional Neural Networks via Attention Transfer},
    url = {},
    year = {2016}}


First install PyTorch, then install torchnet:

git clone
cd tnt
python install

Install OpenCV with Python bindings, and torchvision with OpenCV transforms:

git clone
cd vision; git checkout opencv
python install

Finally, install other Python packages:

pip install -r requirements.txt



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

First, train teachers:

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

To train with activation-based AT do:

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

To train with KD:

python --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.


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: [coming]

Model definition: [coming]

Convergence plot:

Train from scratch

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


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

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

comments powered by Disqus