Jittor Warmup2 ModelNet40 Classification

This repository contains the open-source code for the Jittor competition warmup2 task: point cloud classification on ModelNet40-format data.

The default solution trains a PointNet classifier. The source file also keeps a PCT implementation for experiments.

Environment

Recommended environment:

• Python 3.8+
• Jittor
• NumPy

Install dependencies:

pip install -r requirements.txt

If Jittor requires extra CUDA configuration, follow the official installation guide: https://github.com/Jittor/jittor

Data

Download the official competition data and place it under data/:

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

For this workspace, the helper script defaults to ../data because the original data directory sits beside gitlink.

Expected data shapes are documented in data/README.md.

Training

Run the default training and prediction pipeline:

bash scripts/train_and_predict.sh

The helper script launches:

python -u src/modelnet40_pct.py --data_dir ../data --n_points 1024 --batch_size 32 --epochs 200 --lr 0.001 --weight_decay 0.0001 --dropout 0.5 --model pointnet --optimizer adam --val_ratio 0.1 --seed 42

Generated files are moved into outputs/default/:

• train.log: training and validation log
• command.txt: exact command used by the helper script
• pct_best.pkl: best validation checkpoint
• result.json: test predictions
• result.zip: submission package

To customize hyperparameters, run:

python src/modelnet40_pct.py --help

Then pass the desired arguments directly to src/modelnet40_pct.py.

Evaluation / Packaging

The training command validates on a deterministic split controlled by –seed and –val_ratio, then predicts all test samples.

Validate the generated submission:

cd outputs/default && python ../../scripts/check_result.py

Result

Metric: point cloud classification accuracy on the validation split during local development. The final online score may differ because test labels are hidden by the competition platform.

Default model:

• PointNet
• 1024 sampled points per shape
• batch size: 32
• epochs: 200
• Adam learning rate: 0.001
• weight decay: 0.0001
• dropout: 0.5
• validation ratio: 0.1
• seed: 42

Reproducibility

Use –seed to control the train/validation split and random point sampling. The source sets NumPy and Jittor seeds at startup.

The repository does not include raw .npy data files, checkpoints, logs, or generated submissions. These are ignored by .gitignore.

Third-party Dependencies

This project uses Jittor APIs for model definition, optimization, and dataset iteration. Please follow the Jittor license when redistributing this code or building upon it.