Performer vs. Graph Attention Networks on ShapeNet

Mini-Project on Graph Networks and Architectures for the Deep Learning Advanced course at KTH, Stockholm (DD2412).

A performer-based architecture and a graph attention network are compared for point cloud classification on the ShapeNet dataset. GradCAM is used to investigate model differences by explaining the most important points for a prediction.

For more information, read the full project report. The report contains a short introduction to performers, graph attention networks and GradCAM, as well as experimental details and results.

Setup

Main dependencies:

• numpy, pandas, matplotlib, seaborn
• pytorch, torch_geometric
• pytorch lightning
• fast transformers
• weights and biases
• jupyter

Dependencies are managed using Conda and Mamba

bash scripts/environment.sh
pip install -e .

ShapeNet data is downloaded to ./data on the first run.

Code Structure

miniproject
├── __init__.py             
├── __main__.py              Main entrypoint for conf/train/test tasks
├── configuration.py         Configuration parsing and management
├── datamodules.py           ShapeNet dataset
├── metrics.py               Additional lighning metrics
├── models                   
│   ├── __init__.py          
│   ├── common.py            Lightning model wrapper
│   ├── performer.py         Performer model
│   └── graph_attention.py   Graph Attention model
├── train.py                 Training code
└── test.py                  Evaluation code

Commands

Configuration

Trainin configuration is managed using yaml files and command-line overrides. To list the available options and verify a parsed configuration one can run:

python -m miniproject conf conf/quick.yaml logging.tags='[bbb,ccc]' trainer.max_epochs=2

trainer:
  max_epochs: 2
# ...
logging:
  run_name: smart-flounder-of-honeydew
  tags:
  - bbb
  - ccc
# ...

Training

A quick training run can be launched with:

export CUDA_VISIBLE_DEVICES=0,1
python -m miniproject train conf/quick.yaml other.gpus=2

Hyperparameter sweeps can be launched using the scripts in ./scripts, for example:

bash scripts/sweep-architecture.sh

Evaluation

Given a model checkpoint, evaluation metrics on the test set can be computed with:

export CUDA_VISIBLE_DEVICES=0,1
python -m miniproject test path/to/checkpoint

We provide the following checkpoints:

• Performer model: runs/miniproject/myrtle-gorilla-of-culture/checkpoints/last.ckpt
• Graph Attention model: runs/miniproject/vivacious-honest-lion/checkpoints/last.ckpt

Notebooks

The following notebooks were developed alongside the code. Each notebooks is accompanied by output figures/videos in the corresponding folder.

notebooks
├── Data/ShapenetDataset.ipynb              Dataset statistics
├── Hyperparameters/Hyperparameters.ipynb   Architecture hyperparameter sweep
├── Testing/Results.ipynb                   Test results and GradCAM explanations
└── PytorchHooks.ipynb                      How to make pytorch hooks work