Setup Environment

conda env create -f environment.yml
conda activate 180a-A01

Training Models

Example Usage

python3 main.py

Example Custom Config Usage

python3 main.py --config configs/my_new_config.json

Output

results/results.csv - Predictions of ln(BNPP + 1) values and actual ln(BNPP + 1) values
plots/results_plot_test.png - Regression plot of predicted vs actual ln BNPP
plots/loss_plot_test.png - Training and Validation loss plot
plots/combined_plot_test.png - Loss and Regression plot in the same image
log.txt - Outputs of epoch loss for train and validation sets
best_model_params.pt - Saved weights from model with lowest validation loss

Build targets

all, train_model, test_model, continue_training

• all: trains the model and tests it on the test dataset
• train_model: only trains a model without predicting on test dataset
• test_model: only predicts on test dataset with provided model weights
• continue_training: reads model weights and continues training with configuration file specified

ex.

python3 xray_main.py test_model --config my_new_config.json

Will load the saved_weights from my_new_config.json and then predict on the test dataset

Image Transformations:

Modify image transformations in transforms.py with PyTorch Transformations.

Configuration file:

• model: Model to be used. Options are 'resnet' and 'vgg'
• dataloaders:
  • batch_size
  • shuffle:
  • num_workers
  • use_custom_transforms
• training:
  • epochs: Number of Epochs.
  • criterion: Loss Function. Options are 'MAE' and 'MSE'
  • lr: Learning Rate.
  • weight_decay: L2 regularization penalty.
  • use_scheduler: Use Learning Rate Scheduler.
  • scheduler_step_size: LR scheduler step size.
  • lr_decay_rate: LR Scheduler decay rate.
  • use_estop: Use early stopping
  • estop_num_epochs: Number of epochs to wait before early stopping
• filepaths:
  • data_dir_path: path to were the data is
  • hdf5_stem: file name stem
  • train_dataset: filepath to train dataset
  • val_dataset: filepath to val dataset
  • test_dataset: filepath to test dataset
  • results_csv_path: filepath where test dataset predictions are stored
  • saved_weights_path: filepath for where to save/load weights
  • loss_plot_path: filepath to save a training loss plot
  • results_plot_path: filepath to save a regression plot
  • combined_plot_path: filepath to save a training loss and regression plot