SauceCat / dog_breed_pytorch

NTUOSS Workshop

Dog Breed Identification

by Li Jiangchun for NTU Open Source Society

This workshop assumes basic knowledge of Python, Image Recognition and pytorch.

Setup

1. Open this google drive folder and add to your own google drive. Inside this folder:

   • data: data for this workshop
   • src: python scripts
   • *.ipynb: jupyter notebooks for this workshop, which you can open with Google Colab

2. Change to a Free GPU Runtime. Follow this part.

3. Mount Google Drive. Open setup.ipynb notebook with colab, and run through the set up process.

---

You can refer to the setup part from this post. Thanks Tu Anqi for providing this. To test whether you have completed the initial setup properly, try to run data_explore.ipynb with colab.

Schedule

1. data explore
2. data preparation
3. dataset, data loader and data augmentation
4. fine tune a pretrained model
5. bottle-neck features logistic regression
6. visualize training process with TensorBoard
7. other techniques to improve a image recognition model:
   • test time augmentation
   • ensembling
   • adding more data
   • etc.

Performance tracking (log loss)

• incep V3, sgd, 0.001, 10 epochs: 0.73262
• incep V3, adam, 0.0001, 10 epochs: 0.70818
• incep V3, adam, 0.0001, 10 epochs, test aug 10 times: 0.68884
• incep V3, adam, 0.001, 10 epochs: 0.51537
• incep V3, adam, 0.0005, 10 epochs: 0.47954
• incep v3, bottle neck, logistic regression: 0.25845
• incep v3, bottle neck, logistic regression, test aug 5 times: 0.22451
• resnet 152, bottle neck, logistic regression: 0.39105
• resnet 152, bottle neck, logistic regression, test aug 5 times: 0.30958
• resnet 152, bottle neck, logistic regression, test aug 10 times: 0.29793
• inception v3 best + restnet 152: 0.22299

previous best: xception + inception v3 + incep res: 0.17898