PCT_jittor_ModelNet40

A Jittor implementation of Point Cloud Transformer (PCT) for ModelNet40 point cloud classification.

Project Overview

This project trains a PCT classifier on ModelNet40 and exports test predictions in competition format:

• Input: point clouds from data/train_points.npy, data/train_labels.npy, data/test_points.npy
• Output: result.json ({"sample_id": class_id})

Framework

• Deep learning framework: Jittor
• Task: 3D point cloud classification (40 classes)

Files

• pct.py: model definition, training, inference, and result export
• result.json: predicted labels for test samples
• REPORT.md: experiment report
• PA3_README.pdf: assignment document

Environment

• Python 3.10
• Jittor 1.3.10.0
• Linux + NVIDIA GPU (used in this run: RTX 3090)

Run

python pct.py --epochs 200 --batch_size 32 --n_points 1024 --lr 0.01

After training, pct.py generates:

• pct_model.pkl
• result.json

Dataset Note

The large dataset files (data/*.npy) are not tracked in git. Please download ModelNet40 processed data from the course/competition release page and place them under data/:

• train_points.npy
• train_labels.npy
• test_points.npy
• categories.txt

Result (Current Submission)

• Educoder request ID: 202605120227371697125
• Accuracy: 0.8549
• Status: accept

Author

• Name: 苏继荣
• Student ID: 2024080045