galatolofederico / ncta2022

ncta2022

Repository for the paper Deep learning of structural changes in historical buildings: the case study of the Pisa tower

Installation

Clone this repository

git clone https://github.com/galatolofederico/ncta2022.git
cd ncta2022

Create a virtualenv and install the requirements

virtualenv --python=python3.8 env
. ./env/bin/activate
pip install -r requirements.txt

If you are interested in the dataset please contact us

Usage

To train a model run

python train.py train.save_model=<save_path> dataset=<dataset> train.wandb=<true/false> wandb.tag=<tag> architecture=<architecture>

Where

• dataset can be: post prepost or full
• architecture can be: LSTMRegressor VectorAutoRegressor or TransformerRegressor

This script will save the trained model in <save_path>

To run the inference with a model run

python predict.py predict.model=<path-to-checkpoint> predict.split=<split> dataset=<dataset> architecture=<architecture>

Where

• predict.split can be: train validation test or post (post is available only using the prepost dataset)
• dataset can be: post prepost or full
• architecture can be: LSTMRegressor VectorAutoRegressor or TransformerRegressor

This script will save the outputs and the plots in the folder ./plots

To compute the performance metrics run

python evaluate.py evaluate.model=<path-to-checkpoint> evaluate.split=<split> dataset=<dataset> architecture=<architecture>

Where

• evaluate.split can be: train validation test or post (post is available only using the prepost dataset)
• dataset can be: post prepost or full
• architecture can be: LSTMRegressor VectorAutoRegressor or TransformerRegressor

This script will save the results in the folder ./results

Scripts

To train all the models from the paper run

./train_all.sh

To run the inference with the trained models run

./predict_all.sh

To compute the performance metrics for all the models run

./evaluate_all.sh

Contributions and license

The code is released as Free Software under the GNU/GPLv3 license. Copying, adapting and republishing it is not only allowed but also encouraged.

For any further question feel free to reach me at [e-mail] or on Telegram @galatolo