iNaturalist Challenge(2018) with resnet

Introduction

This release focuses on AI based graphic classification. We train resnet(152/101/50 layers) for iNaturalist Challenge at FGVC 2018 with tensorpack, which is a training interface based on TensorFlow.

Result

On inaturalist-2018 Dataset, we train resnet(50/101/152) respectively，the result is as follows:

Model Name	train-error-top1	train-error-top3	val-error-top1	val-error-top3
Resnet50	0.13361	0.061188	0.399	0.24171
Resnet101	0.105	0.061306	0.37014	0.21371
Resnet152	0.11464	0.059394	0.35454	0.20024

Installation, Data Preparation, Training and Testing

Disclaimer

• The code is tested on a server with 188.00 GB memory, and 40 core cpu. Data storages in SSD.
• we train the model with 8 Pascal Titian XP gpu, for resnet50 the batch is 32*8=256, for resnet101/152 the batch is 24*8=192

Install

Dependencies:

• python3. We recommend using Anaconda as it already includes many common packages.
• Python bindings for OpenCV (Optional, but required by a lot of features)
• TensorFlow >= 1.3.0 (Optional if you only want to use tensorpack.dataflow alone as a data processing library)

# install git, then:
pip install -U git+https://github.com/ppwwyyxx/tensorpack.git
# or add `--user` to avoid system-wide installation.

Prepare data

data should be organized as follows(set path in config.py):

$HOME/DataSet/iNaturalist2018/
    |->ground_truth
    |    |train2018.json
    |    |val2018.json
    |    |test2018.json
    |->train_val2018
    	 |...
    |->test2018
         |...

Train and Eval

• train script example:

python iNaturalist-resnet.py --data /home/huzhikun/DataSet/iNaturalist2018/ --batch 192 --mode resnet --gpu 0,1,2,3,4,5,6,7 -d 152

• eval script example:

python iNaturalist-resnet.py --eval --data /home/huzhikun/DataSet/iNaturalist2018/ --mode resnet --gpu 0,1,2,3 -d 152 --load train_log/iNaturalist-resnet-d152/model-205065

Test with kaggle submit file(.csv)

• test script example:

python iNaturalist-resnet.py --test --data /home/huzhikun/DataSet/iNaturalist2018/ --mode resnet --gpu 7 -d 152 --load ./train_log/iNaturalist-resnet-d152/model-239190

Citing

@misc{wu2016tensorpack,
  title={Tensorpack},
  author={Wu, Yuxin and others},
  howpublished={\url{https://github.com/tensorpack/}},
  year={2016}
}

Credits

+Thanks to everyone who directly contributed to this:

• Steve Deng
• Bill Lee
• Xuyang Wang
• Zhikun Hu