VRDL HW1 - Bird Images Classification

Introduction

In this assignment, we are given 6,033 bird images belonging to 200 bird species. There are 3,000 training images and 3,033 test images. This assignment is a fine-grained classification task, which means it is difficult to differentiate different bird species. Simply use ImageNet pretrained model with limited training data to do transfer learning cannot get a satisfied accuracy in testing. To tackle the fine-grained classification, I add Center Loss along with Softmax (Cross-Entropy) Loss as my loss function. I also experiment various backbone models such as ResNet, ResNeXt, DenseNet and EfficientNet. Finally, the best accuracy I can get is around 0.8 using EfficientNet-B4 as backbone.

Environment

Hardware

CPU: AMD Ryzen 5 3600 6-Core
GPU: NVIDIA GeForce RTX 3070 8GB

Packages

torch: 1.7.1+cu110
torchvision: 0.8.2+cu110
torch-summary: 1.4.4
numpy: 1.19.5
Center Loss: The center_loss.py in my project is copied from this repository

EfficientNet: Install with pip install efficientnet_pytorch, and load a pretrained EfficientNet with:

from efficientnet_pytorch import EfficientNet
model = EfficientNet.from_pretrained('efficientnet-b4', num_classes=200)

Training

To train the model, run this command:

python Train.py

Reproducing Submission

Clone this project or download ZIP.
Create a folder named data and put testing_images, classes.txt, testing_img_order.txt inside.
Download the submitted efficientnetb4_1.pth model.
Create folder output/models and put the .pth file inside.
To create the answer.txt file, run this command:
```
python inference.py
```

Results

The best accuracy I can get on testing data is around 0.8 using EfficientNet-B4.

Model name	Top 1 Accuracy on CodaLab
efficientnetb4_1.pth	0.801187

hurueilin / VRDL_HW1