torch ==1.8.1 matplotlib open-cv
- Install the corresponding package source and configure the environment.
- Generate corresponding datasets and annotations.
- Adjust the training method and modify the corresponding code.
- Generate the corresponding data set. If it is just a common training method, you only need to generate a single data set and label. After adjusting the code, run train.py to train the model.
- If you need to try co-training, you will need to generate three corresponding data sets according to the requirements of the paper, and run co-train on the basis of 1 for co-training.
- The relevant extended experimental results and parts are retained in the code, and the experimental code is submitted at the same time. Part of the key code involves intellectual property protection, and only part of the code is provided open source. If necessary, please refer to the paper on this method for personal reproduction.