The code and paper for my Part II dissertation looking into extracting concepts from simplified graph neural networks.
The requirements files is located in the root directory. To install the requirements navigate to the root folder, create a python virtual environment and run
pip install -r requirements.txt
Warning PyTorch lightning 1.19 was used during development and PyTorch Lighnting 2.0+ removes support for important wrapper functionality
This is the model that will be later explained. No pre-trained modesl are provided in this directory. Individual models can be trained using the src/main.py function
python3 src/main.py -v --config <config_file> --seed <seed> --root <dataset_root_directory>
The configs used for the dissertation are presented in src/configs/curr.
To prevent having to run multiple scripts scripts/mk_expr.sh or scripts/mk_all.sh may be run to generate experiment runs. mk_all.sh will generate a script that runs 10 experiments to get a sample of accuracies, these use the same random seeds as those used in the dissertation.
./scripts/mk_expr.sh [list_config_filepaths]
A resulting run/[all_]<model>-<dataset>.sh file will be created
The multiple results from running a run/all_<model>-<dataset>.sh will be saved in the folder output/<model>-<dataset>/ folder with the filename <data>-<time>.
Once a model is trained it will output an activation pickle file in activations/. These will follow the general naming convention activations/<expr_name>_<seed>-<date>-<time>[_<option>].pkl.
The concepts for a trained model can therefore be extracted, concept scores calculated, and the visualisation saved to the folder output/<model>-<dataset>/. Extracting and evaluating concepts requires the number of clusters, the number of hops, and the number of graphs to be visualised.
python3 src/eval.py --activation activations/<activation>.pkl --clusters <clusters> --hops <hops> --num_graphs <num_graphs_visualised>
Two activations can be compared to identified at which layer two models have the closest latent space and thus the best place to join the two models together.
python3 src/similarity.py --activations [list_of_activations] --clusters <clusters> --hops <hops>
The config files are written in YAML and follow the following basic structure
model:
name: <model_name>
kwargs: <required_kwargs>
sampler:
name: <loader_name>
wrapper:
name: <wrapper_name>
kwargs: <required_kwargs>
trainer:
dir: <log_directory>
accelerator: <accelerator>
devices: <num_devices>
max_epochs: <num_epochs_run>
Examples for specific models, datasets, wrappers and loaders are all available in src/configs/curr using the naming convention <model_name>.<loader_type>.<dataset>.<version>.yml.
It is also possible to carry out hyperparameter sweeps to tune the designed models or verify the hyperparameter surfaces in Appendix B.
python3 src/sweep.py --config <config_file> --lr [learning_rates] [--wd [weight_decay_constants]]
Most folders contain a tests subfolder including the relevant tests to check that the implementation is working as intended. Extensions or alterations of the code should run these tests to verify that relevant modules and functions operate as intended.
python3 -m unittest src.<folder>.tests.<test> -v