A Vision Transformer for Fine-grained Classification by Reducing Noise and Enhancing Discrimnative Information

PyTorch code for the paper: A Vision Transformer for Fine-grained Classification by Reducing Noise and Enhancing Discrimnative Information

Framework

Dependencies:

• python 3.7.3
• PyTorch 1.8.0
• torchvision 0.9.0
• ml_collections 0.1.0
• numpy 1.20.1
• pandas 1.2.3
• scipy 1.6.2

Usage

1. Download Google pre-trained ViT models

We use ViT-B_16 as the backbone. The official download link for pretrained weight https://storage.googleapis.com/vit_models/sam/ViT-B_16.npz.

2. Prepare data

In the paper, we use images from there publicly available datasets:

• CUB-200-2011
• Stanford Dogs
• NABirds
• iNat2017

Please download them from the official websites and put them in the corresponding folders.

In order to read datasets in a unified approach, we use CSV files to read data. Please organize the train set and test set into the following forms:

	img_name	label
0	001.Black_footed_Albatross/Black_Footed_Albatross_0009_34.jpg	1
1	001.Black_footed_Albatross/Black_Footed_Albatross_0074_59.jpg	1
...	...	...
5993	200.Common_Yellowthroat/Common_Yellowthroat_0049_190708.jpg	200

Note:

• The label of categories starts from 1 not 0.
• The order of each image does not matter.

3. Install required packages

Install dependencies with the following command:

pip install -r requirements.txt

4. Train

You can modify the configuration in the config.py file and run it directly:

python main.py

5 Validate

After model training, you will get the weight in the checkpoints folder. You can modify the configuration in the config.py file and run it directly:

python val.py