Pytorch Lightning implementation of Jigsaw Patch Learning for Fine-grained image classification

• This repository is created for SFU CMPT 757 class final project.
• Follwing is the link for my report: https://drive.google.com/file/d/1TwjfJGza-sVAXdqWAxRbdO5NogSQxTRZ/view?usp=sharing
• This repository's backbone classfier is my Pytorch lightning implementation of https://github.com/PRIS-CV/PMG-Progressive-Multi-Granularity-Training. Credits to the author of PMG.
• This repository has additional loss functions that can improve the results of the original code (PMG), and acheives state-of-the-art results in FGVC benchmark datasets like CUB-200_2011.

Requirment

python 3.6

PyTorch >= 1.3.1

Pytorch lightning >= 1.10

torchvision >= 0.4.2

Instructions

• You can download dataset from the official link: http://www.vision.caltech.edu/visipedia/CUB-200.html
• Put your dataset (CUB-200-2011 is the default) and put it inside a dataset folder
• Run dataset.py and it will convert your dataset into following format:
• Refer to colab.ipynb to see how things run exactly.

dataset
├── train
│   ├── class_001
|   |      ├── 1.jpg
|   |      ├── 2.jpg
|   |      └── ...
│   ├── class_002
|   |      ├── 1.jpg
|   |      ├── 2.jpg
|   |      └── ...
│   └── ...
└── test
    ├── class_001
    |      ├── 1.jpg
    |      ├── 2.jpg
    |      └── ...
    ├── class_002
    |      ├── 1.jpg
    |      ├── 2.jpg
    |      └── ...
    └── ...

• If your dataset is not CUB_2011, refer to above structure or refer to my custom function.

Config

• Model parameters are controlled by config.py.
• The only thing that you should touch is model type: ce_vanilla, ce_label_smooth, complement,
• The above parameter controls the loss function that you will be using. The vanilla function will just use cross entropy
• When you are training, make sure the number of classes match your dataset(i.e 200)

Train

• Just run train.py
• Batch size will automatically adjust based on your computer capacity. Recommended size is 16.
• The performance will suffer if you cannot acheive batch size of 16.
• 16-bit precision is used.

weights

• The weights will auto-save to weights folder
• https://drive.google.com/file/d/1-0LvDBiODpgTMS9hw9HIwtFAwF-KYcuW/view?usp=sharing you can download the weights from here too.