This repo contains a small, medium, and large Autoencoder trained on the HGCal dataset. These models were developed using a Bayesian optimization process, and the models were selected along a Pareto front defined by Model error versus Binary Operations (BOPS).
Models provided:
- Small Pareto
- Medium Pareto
- Large Pareto
To evaluate the model, download the dataset here.
We provide a script to load the model for further evaluation: ./scripts/load_model.sh
.
No need to download the dataset to load the model.
To load the Medium Pareto model, run:
./scripts/load_model.sh 0
To load the Small Pareto model, run:
./scripts/load_model.sh 1
To load the Large Pareto model, run:
./scripts/load_model.sh 2
To load the Medium Pareto model, run:
./scripts/load_model.sh 3
To load the Small Pareto model, run:
./scripts/load_model.sh 4
To load the Large Pareto model, run:
./scripts/load_model.sh 5
We provide a script for evaluating the Earth Mover's Distance (EMD) of the model: ./scripts/eval.sh
.
Lower EMD is better.
A perfect autoencoder reconstruction of the data would yield an EMD of 0.
Make sure to point the .scripts/eval.sh
's DATASET
path to where you have stored the dataset, e.g.,
DATASET=./data/pickled--data_values--phys_values--EoL_dataset.pkl
To evaluate the Medium Pareto model, run:
./scripts/eval.sh 0
To evaluate the Small Pareto model, run:
./scripts/eval.sh 1
To evaluate the Large Pareto model, run:
./scripts/eval.sh 2