Jittor ModelNet40 Point Cloud Classification

This repository contains code for the Jittor warm-up challenge on ModelNet40 point-cloud classification. The official submission is a zip file containing a single result.json mapping test sample ids to class ids.

Environment

Recommended environment:

• Python 3.10
• Jittor 1.3.11 or newer
• CUDA-capable GPU recommended

Install with conda:

conda create -n jittor-hot python=3.10 -y
conda activate jittor-hot
python -m pip install -r requirements.txt

Optional Jittor checks:

python -m jittor.test.test_example

Data

Download the competition-provided ModelNet40 point-cloud data and place it under data/. Expected files:

data/train_points.npy
data/train_labels.npy
data/test_points.npy
data/categories.txt

If the downloaded file is data/data.zip, unzip it from the project root:

unzip -o data/data.zip -d .

The large data files are ignored by git. See data/README.md.

Training

Primary model used for the best local validation result:

bash scripts/train_improved.sh

This writes checkpoints, logs, and submission artifacts under outputs/:

outputs/checkpoints/best_model_improved.pkl
outputs/logs/train_improved.log
outputs/results/submit_improved.zip

Alternative runs:

bash scripts/train_less_reg.sh
bash scripts/train_full_pct.sh

All scripts accept extra CLI overrides. Example:

bash scripts/train_improved.sh --seed 7 --votes 20

Inference

Generate a submission zip from a trained improved checkpoint:

bash scripts/predict_improved.sh

Validate the zip structure before submitting:

python tools/validate_submission.py outputs/results/submit_improved.zip

The zip must contain exactly:

result.json

Results

Metric: classification accuracy on the hidden test labels.

Local validation summary from current experiments:

Run	Checkpoint	Best val acc
improved PCT	best_model_improved.pkl	82.13%
less regularized PCT	best_model_less_reg.pkl	81.62%
full PCT variant	best_model_full_pct.pkl	81.12%
baseline PCT	best_model_1024.pkl	80.10%

Validation uses a stratified 90/10 train/val split and may differ from the online hidden-test score because the official test labels are not available.

Repository Layout

configs/       # reference hyperparameter configs
data/          # data placement instructions; large files are ignored
scripts/       # train and inference entrypoints
src/           # dataset, models, training, inference, packaging
tools/         # validation utilities
outputs/       # local logs, checkpoints, submissions; ignored by git

Reproducibility Notes

• All training entrypoints expose --seed and set Python, NumPy, and Jittor seeds.
• Key hyperparameters are controlled by CLI arguments and mirrored in configs/.
• Logs, checkpoints, and result zips are written to outputs/ by default.