Capstone project

ResNet-50 for Cats.Vs.Dogs.

Environment

Python 2.7.12 Mac OS X 10.11.6 CUDA 8.0.47 cuDNN 5.1

Necessary Libraries

• keras 1.1.2
• NumPy 1.11.2
• OpenCV 3.1.0
• tensorflow gpu 0.11.0
• matplotlib 1.5.1
• tqdm 4.7.6
• sklearn 0.18.1
• seaborn 0.7.1
• pandas 0.19.1

Dataset

Initial dataset is downloaded from kaggle. The dataset includes two folders test and train.

But I split train folder into 2 folders:

the prossessing steps completed in preprocessing train dataset.ipynb

• mytrain ---- including two folders
  • cat ---- including about 11250 cat images
  • dog ---- including about 11250 dog images
• myvalid ---- including two folders
  • cat ---- including about 1250 cat images
  • dog ---- including about 1250 dog images

Dataset preview

Training Time

• 24s per epoch (2048 images)
• total epochs: 20
• CPU i7 6700K
• GPU GTX 980 Ti
• Memory 32GB

Result preview

Result

Feature heatmap

Reference

@article{He2015,
    author = {Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun},
    title = {Deep Residual Learning for Image Recognition},
    journal = {arXiv preprint arXiv:1512.03385},
    year = {2015}
}

@article{zhou2015cnnlocalization,
    author={Zhou, B. and Khosla, A. and Lapedriza. A. and Oliva, A. and Torralba, A.},
    title={{Learning Deep Features for Discriminative Localization.}},
    journal={CVPR},
    year={2016}
}