This is an unofficial comprehensive implementation of Point Transformer, a pioneering deep-learning technique for 3D data processing introduced by Zhao et al. (2021) [1].
Please follow the instructions in the data folder in order to download the datasets and place them in the same folder.
Each .py file in this repository is designed to be unit-tested by directly running them. To run the unit tests for a specific file, simply execute the file directly. For example:
python data_loader.py
This will run the unit tests defined in data_loader.py and provide feedback on whether the code behaves as expected.
To train the Point Transformer model, you can use either of the following scripts:
train_bis.py: This script trains the model for approximately 9 hours and achieves a test accuracy of 91.28%.train.py: This alternative script offers faster training, taking only around 4 hours, while achieving a comparable accuracy of 91.12%.
Both scripts handle the training process, including data loading, model initialization, training loop, and evaluation.
You may find test the pretrained model and its performance by running the Jupiter notebook test.ipynb. This will load the test data and the pretrained model and compute the overall accuracy and the confusion matrix on the test set.
[1] Zhao, Hengshuang, et al. "Point transformer." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.