TASN

Code (MXNet version) for our cvpr'19 paper "Looking for the Devil in the Details: Learning Trilinear Attention Sampling Network for Fine-grained Image Recognition"

Prerequisites

cuda version = 8.0 cudnn5.0 nccl libopenblas liblapack libopencv

Install

First clone this repository:

sudo git clone https://github.com/Heliang-Zheng/TASN.git
cd TASN/tasn-mxnet

Then, please follow https://mxnet.incubator.apache.org/install/build_from_source.html to compile and install mxnet.

Or download pre-build mxnet (with cuda 8.0): https://drive.google.com/open?id=1Sfpw0x5XLqBFWAt99-zKOp4jAbOxm5Ws and install by:

cd TASN/tasn-mxnet/example/tasn
sudo bash install.sh

Train TASN

1. get into the tasn dir:

    cd TASN/tasn-mxnet/example/tasn
   

2. download data and pretrained model (on ImageNet):

    sudo bash init.sh
   

3. set your nccl path in train.sh

4. run :

    sudo bash train.sh
   

Experiments settings: on CUB-200-2011 dataset : http://www.vision.caltech.edu/visipedia/CUB-200-2011.html

CNN input resolution: 224*224

Accuracy: 87.0%

Just changing the scale of AttSampler() in train.py from 224/512 to 336/512 to obtain the accuracy of 88.0%

Model:

cub_224_87 https://drive.google.com/open?id=1uw9MVNVZqBTppN4TBbHB10CxoQonsTx9

cub_336_88 https://drive.google.com/open?id=1qQo8o2C5JpwxJGhrfk2xHM-f6kpxDKd1

Added files:

example/tasn/*

src/operator/contrib/att_sampler-inl.h

src/operator/contrib/att_sampler.cc

src/operator/contrib/att_sampler.cu

PyTorch versioin

TODO