PCT_jittor

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

This repository is prepared to match the PA3 requirements in PA3_README.pdf:

• Framework: Jittor
• Task: ModelNet40 point cloud classification
• Main script: pct.py
• Output files: pct_model.pkl, result.json

Project Structure

PA3/
  pct.py                  # training + inference pipeline
  data/                   # dataset folder (not tracked)
  .gitignore
  README.md

Environment

Recommended conda env (already used in this workspace):

conda activate ai_course
python -c "import jittor; print(jittor.__version__)"

Data Preparation

Expected data files:

• data/data/train_points.npy
• data/data/train_labels.npy
• data/data/test_points.npy
• data/data/categories.txt (optional for category names)

If your dataset is a zip package:

unzip -o data.zip -d data

Training + Prediction

Run full training and generate prediction file:

python -u pct.py \
  --data_dir /root/PA3/data/data \
  --epochs 200 \
  --batch_size 32 \
  --n_points 1024 \
  --lr 0.01

After training:

• model is saved to pct_model.pkl
• test predictions are saved to result.json

Fast Sanity Run

python -u pct.py \
  --data_dir /root/PA3/data/data \
  --epochs 1 \
  --batch_size 8 \
  --n_points 1024 \
  --lr 0.01

Report Checklist (from assignment)

Prepare your final submission package with at least:

• pct.py
• pct_model.pkl
• result.json
• REPORT.pdf

For platform submission, also include:

• evaluation screenshot
• evaluation ID
• Git repository link

Open Source Notes

This repository is intended for course assignment usage. Please follow your course policy before sharing trained weights or data files.