modelnet40-pct-classification

Point cloud classification project based on Jittor and a simplified PCT (Point Cloud Transformer) model for the ModelNet40 dataset.

Overview

This project trains a point cloud classifier on ModelNet40 and generates test predictions in result.json.

The current implementation includes:

• point sampling from .npy point cloud files
• simple training augmentation
• PCT-style self-attention blocks
• stratified train/validation split
• AdamW optimizer
• label smoothing
• best checkpoint selection by validation accuracy
• test-time voting

Project Structure

PA3/
├─ data/
│  ├─ categories.txt
│  ├─ train_points.npy
│  ├─ train_labels.npy
│  └─ test_points.npy
├─ pct.py
├─ PA3_README.pdf
└─ README.md

Environment

Recommended

This project is currently intended to run in WSL Ubuntu with a CPU-only Jittor environment.

Reason:

• native Windows Jittor import/compile was not stable on this machine
• the current script forces Jittor to stay on the CPU path for reliability

Main Dependencies

• Python 3.10
• jittor
• numpy

Data Format

The code expects the following files under ./data:

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

Expected shapes:

• train_points.npy: (N, 2048, 3)
• train_labels.npy: (N,)
• test_points.npy: (M, 2048, 3)

How to Run

Run in WSL

Recommended command:

wsl -d Ubuntu bash /mnt/c/Users/ChenChong/Desktop/PA3/wsl_run_pct_cpu.sh

Run with Custom Arguments

Example:

wsl -d Ubuntu bash /mnt/c/Users/ChenChong/Desktop/PA3/wsl_run_pct_cpu.sh --epochs 10 --batch_size 8 --n_points 512 --lr 0.001 --vote_num 3

Direct Python Run

If your current environment already has a working CPU-only Jittor installation:

python pct.py

Main Arguments

The most important command-line arguments are:

• --data_dir: dataset directory, default ./data
• --n_points: number of sampled points per shape, default 1024
• --batch_size: batch size, default 32
• --epochs: number of training epochs, default 200
• --lr: learning rate, default 0.001
• --seed: random seed, default 42
• --val_ratio: validation split ratio, default 0.1
• --weight_decay: AdamW weight decay, default 0.05
• --label_smoothing: label smoothing factor, default 0.2
• --vote_num: number of test-time votes, default 3

Outputs

After training, the script produces:

• pct_model_best.pkl: best model selected by validation accuracy
• pct_model.pkl: final loaded model used for prediction
• result.json: test set predictions

Training Logs

Each epoch prints:

• training loss
• training accuracy
• validation loss
• validation accuracy
• learning rate
• epoch time

Example:

Epoch [1/20]  Loss: 2.1000  Train Acc: 42.35%  Val Loss: 1.9500  Val Acc: 48.20%  LR: 0.000994  Time: 620.4s

The most important metric during training is Val Acc.

Notes

• The script explicitly sets os.environ["nvcc_path"] = "" before importing Jittor.
• The script also sets jt.flags.use_cuda = 0.
• This means the current version is designed to run on CPU for stability.
• Full training on CPU can be slow, especially with large n_points and many epochs.

Suggested Experiments

If you want faster iteration, try:

wsl -d Ubuntu bash /mnt/c/Users/ChenChong/Desktop/PA3/wsl_run_pct_cpu.sh --epochs 10 --batch_size 8 --n_points 512

If you want to push accuracy further, try:

• increasing epochs
• keeping n_points reasonably high
• watching Val Acc instead of only Train Acc
• comparing different vote_num values

Repository Tip

Recommended project URL identifier:

modelnet40-pct-classification