These models predict corrosion-induced cracking using PyTorch. The models takes in corrosion patterns and concrete material properties and outputs whether there is a crack on the concrete’s surface.
- numpy==1.21.5
- ray==2.1.0
- scikit_learn==1.2.0
- scipy==1.9.1
- torch==1.13.0
models/: Directory containing several model architecture configurations.data_generation/: FEM simulation for generating data.data_analysis/: Notebook for analyzing corrosion and concrete input features.data_preprocessing/: Extracts, processes, and joins data from simulations.data_normalization/: Custom normalization for corrosion depth features.hyperparameter_search/: RayTune for random hyperparameter search.model_evaluation/: Helper function for evaluating trained models on test set. Also contains a notebook for more detailed evaluation metrics and plots.data_augmentation/: Generates augmented training data by flipping, adding noise, and monotonic scaling.sample_data_11_09_2022/: Sample corrosion patterns from COMSOL.training_loss_util/: Helper function for computing loss.data_loader/: Helper functions for loading datasets.
- Generate zipped corrosion data.
matlab -nodesktop -r generateCrackData.m
- Extract, preprocess, and join data:
python3 data_preprocessing/preprocess_data.py --output_path=/path/to/data --extract
- Optionally normalize data:
python3 data_normalization/data_normalization.py --training_data_dir=/path/to/data
- Train a model:
python3 models/baseline_model/baseline.py --batch_size=128 --num_epochs=10000 --data_dir=/path/to/data/ --output_path=/path/to/model --print_every=10
- Tune model hyperparameters:
python3 hyperparameter_search/hyperparameter_search.py --num_runs=100 --corrosion_path=/path/to/data --label_path=/path/to/data
- Evaluate models on test set:
python3 model_evaluation/evaluate.py --model_name=baseline_model.py --data_dir=/path/to/data/ --model_dir=/path/to/model --normalized_data