def __init__(self,
data_loader,
results_patent_path,
model_name,
stage='training'):
"""
Initialize the Prediction model class
:param data_loader: An object that loads batches of image data
:param results_patent_path: The path at which all results and generated models will be saved
:param model_name: The name of your model, used when saving and loading data
"""
# set attributes for data shape from data loader
self.data_loader = data_loader
self.patient_shape = data_loader.patient_shape
self.full_roi_list = data_loader.full_roi_list
self.model_name = model_name
# Define training parameters
self.epoch_start = 0 # Minimum epoch (overwritten during initialization if a newer model exists)
self.epoch_last = 200 # When training will stop
# Define image sizes
self.dose_shape = (*self.patient_shape, 1)
self.ct_shape = (*self.patient_shape, 1)
self.roi_masks_shape = (*self.patient_shape, len(self.full_roi_list))
# Define filter and stride lengths
self.filter_size = (4, 4, 4)
self.stride_size = (2, 2, 2)
# Define the initial number of filters in the model (first layer)
self.initial_number_of_filters = 1 # 64
# Define model optimizer
self.gen_optimizer = Adam(lr=0.002,
decay=0.001,
beta_1=0.5,
beta_2=0.999)
# Define place holders for model
self.generator = None
# Make directories for data and models
model_results_path = '{}/{}'.format(results_patent_path, model_name)
self.model_dir = make_directory_and_return_path(
'{}/models'.format(model_results_path))
self.prediction_dir = '{}/{}-predictions'.format(
model_results_path, stage)
# Make template for model path
self.model_path_template = '{}/epoch_'.format(self.model_dir)
print('For this out-of-sample test:\n'
'\tthe DVH score is {:.3f}\n '
'\tthe dose score is {:.3f}'.format(dvh_score, dose_score))
# Apply model to validation set
validation_data_paths = get_paths(
validation_data_dir, ext='') # gets the path of each plan's directory
validation_data_loader = DataLoader(validation_data_paths,
mode_name='dose_prediction')
dose_prediction_model_validation = PredictionModel(
validation_data_loader,
results_dir,
model_name=prediction_name,
stage='validation')
dose_prediction_model_validation.predict_dose(
epoch=number_of_training_epochs)
# Evaluate plans based on dose metrics (no baseline available to compare to)
validation_eval_data_loader = DataLoader(
validation_data_paths, mode_name='dose_prediction') # Set data loader
dose_evaluator = EvaluateDose(validation_eval_data_loader)
dose_evaluator.make_metrics()
validation_prediction_metrics = dose_evaluator.reference_dose_metric_df.head(
)
# Zip dose to submit
submission_dir = make_directory_and_return_path(
'{}/submissions'.format(results_dir))
shutil.make_archive('{}/{}'.format(submission_dir, prediction_name), 'zip',
dose_prediction_model_validation.prediction_dir)
def conversion_epoch_end(self, outputs):
submission_dir = make_directory_and_return_path(
'checkpoints/submissions')
shutil.make_archive('{}/{}'.format(submission_dir, 'submission'),
'zip', 'checkpoints/results')
return #{'val_loss_mean': val_loss_mean, 'log': log_dict, 'progress_bar': log_dict}