Deep-Wide Learning Assistance for Insect Pest Classification

Official code of "Deep-Wide Learning Assistance for Insect Pest Classification."

We propose DeWi, the novel learning assistance for insect pest classification. With a one-stage and alternating training strategy, DeWi simultaneously enhances several Convolutional Neural Networks in two perspectives: discrimination (by leveraging the benefit of triplet margin loss in a supervised training manner) and generalization (with the help of data augmentation). From that, DeWi can learn discriminative and in-depth features of insect pests (deep) yet still generalize well to a large number of categories (wide).

Our DeWi models are better at focusing on meaningful features than the baseline methods.

Getting Started

We highly recommend you to create a separate conda environment for the project. Please follow the below steps to set up the environment and install the necessary packages.

conda create -n dewi python=3.8
conda activate dewi
conda install pytorch pytorch-cuda=11.6 -c pytorch -c nvidia
conda install pip
pip install -r requirements.txt

Data

Download the IP102 dataset from this URL and the D0 dataset from this URL. After downloading, change the dataset_path in config.py to the path of the images folder.

Running script

• Change the root and checkpoint_path in config.py to the appropriate paths.

• Assume you use the DeWi model with ResNet-152 variant, then run the following command to start training:

    python3 train.py dewi_resnet152
  

  After training each epoch, the validation and testing phases are automatically executed. The log files and checkpoint models are saved in the checkpoint_path. Replace name = k[:] by name = k[7:] in utils/auto_load_resume if you want to training a model which is trained on multiple GPUs.

Acknowledgement

Our source code is based on Ung et al.. Special thanks to them.

Contact

For any concerns, please contact Nguyen Thanh Binh (Associate Professor at University of Science Ho Chi Minh city) via ngtbinh@hcmus.edu.vn.