cnn_generalization

CNN generalization

NFN CNN Zoo data

This experiment follows NFN. Download the CIFAR10 data (originally from Unterthiner et al, 2020) into ./dataset, and extract them. Change data_path in ./configs/data/zoo_cifar_nfn.yaml if you want to store the data somewhere else.

Options for data:

• zoo_cifar_nfn: NFN CNN Zoo (CIFAR) dataset

Run experiments with scripts

You can run the experiments using the scripts provided in the scripts directory. For example, to train and evaluate a Neural Graph Transformer (NG-T) model on the CNN Zoo dataset, run the following command:

./scripts/cnn_zoo_rt.sh

This script will run the experiment for 3 different seeds.

CNN Wild Park

Download the dataset from Zenodo and extract it into ./dataset.

Run experiments with scripts

You can run the experiments using the scripts provided in the scripts directory. For example, to train and evaluate a Neural Graph Transformer (NG-T) model on the CNN Wild Park dataset, run the following command:

./scripts/cnn_zoo_rt.sh

This script will run the experiment for 3 different seeds.

Hyperparameter Sweep

We also provide sweep configs for NG-T, NG-GNN, and StatNN in the sweep_configs directory. In the following commands, change the --project and the --entity according to your WandB account, or change the corresponding yaml files.

NG-T:

wandb sweep --project cnn-generalization --entity neural-graphs sweep_configs/sweep_cnn_park_transformer.yaml

NG-GNN:

wandb sweep --project cnn-generalization --entity neural-graphs sweep_configs/sweep_cnn_park_gnn.yaml

StatNN:

wandb sweep --project cnn-generalization --entity neural-graphs sweep_configs/sweep_cnn_park_statnn.yaml