def main():
"""Main block of code, which runs the training"""
start = time.time()
logging.basicConfig(level="INFO")
# Load the .yaml data and unpacks it
assert len(sys.argv) == 2, "Exactly one experiment configuration file must be "\
"passed as a positional argument to this script. \n\n"\
"E.g. `python run_non_tracking_experiment.py <path to .yaml file>`"
with open(sys.argv[1], "r") as yaml_config_file:
logging.info("Loading simulation settings from %s", sys.argv[1])
experiment_config = yaml.load(yaml_config_file)
experiment_settings = experiment_config['experiment_settings']
data_parameters = experiment_config['data_parameters']
ml_parameters = experiment_config['ml_parameters']
path_parameters = experiment_config['path_parameters'] \
if 'path_parameters' in experiment_config else None
# Loads the raw dataset
logging.info("Loading the dataset...")
data_preprocessor = Preprocessor(data_parameters)
features, labels = data_preprocessor.load_dataset()
if path_parameters:
path_creator = PathCreator(data_parameters, path_parameters, labels)
paths = path_creator.load_paths()
# Undersamples the dataset (if requested)
if "undersample_bf" in experiment_settings and experiment_settings["undersample_bf"]:
features = undersample_bf(features, data_parameters["beamformings"])
if "undersample_space" in experiment_settings:
assert not path_parameters, "This option is not supported for tracking experiments, "\
"unless the code for the path creation is updated"
features, labels = undersample_space(features, labels, data_parameters["undersample_space"])
# Initializes the model and prepares it for training
logging.info("Initializing the model (type = %s)...", experiment_settings["model_type"].lower())
if experiment_settings["model_type"].lower() == "cnn":
ml_parameters["input_type"] = "float"
model = CNN(ml_parameters)
elif experiment_settings["model_type"].lower() in ("lstm", "tcn"):
assert path_parameters, "This model requires `paths_parameters`. See the example."
assert path_parameters["time_steps"] == ml_parameters["input_shape"][0], "The ML model "\
"first input dimention must match the length of the paths! (path length = {}, model)"\
"input = {})".format(path_parameters["time_steps"], ml_parameters["input_shape"][0])
ml_parameters["input_type"] = "bool"
if experiment_settings["model_type"].lower() == "lstm":
model = LSTM(ml_parameters)
else:
model = TCN(ml_parameters)
else:
raise ValueError("The simulation settings specified 'model_type'={}. Currently, only "
"'cnn', 'lstm', and 'tcn' are supported.".format(experiment_settings["model_type"]))
model.set_graph()
# Creates the validation set
logging.info("Creating validation set...")
if path_parameters:
features_val, labels_val, _ = sample_paths(
paths["validation"],
features,
labels,
experiment_settings,
data_parameters,
path_parameters,
)
else:
features_val, labels_val = create_noisy_features(
features,
labels,
experiment_settings,
data_parameters,
)
# Runs the training loop
logging.info("\nStaring the training loop!\n")
keep_training = True
while keep_training:
logging.info("Creating noisy set for this epoch...")
if path_parameters:
features_train, labels_train, _ = sample_paths(
paths["train"],
features,
labels,
experiment_settings,
data_parameters,
path_parameters,
sample_fraction=experiment_settings["train_sample_fraction"]
)
else:
features_train, labels_train = create_noisy_features(
features,
labels,
experiment_settings,
data_parameters,
)
model.train_epoch(features_train, labels_train)
predictions_val = model.predict(features_val, validation=True)
keep_training, val_avg_dist = model.epoch_end(labels_val, predictions_val)
if predictions_val is not None:
# Upscales the validation score back to the original scale and gets the 95th percentile
val_avg_dist *= data_parameters["pos_grid"][0]
val_95_perc = get_95th_percentile(
labels_val,
predictions_val,
rescale_factor=data_parameters["pos_grid"][0]
)
logging.info("Current avg val. distance: %.5f m || 95th percentile: %.5f m\n",
val_avg_dist, val_95_perc)
# Store the trained model and cleans up
logging.info("Saving and closing model.")
experiment_name = os.path.basename(sys.argv[1]).split('.')[0]
model.save(model_name=experiment_name)
model.close()
# Prints elapsed time
end = time.time()
exec_time = (end-start)
logging.info("Total execution time: %.5E seconds", exec_time)
def main():
"""Main block of code, which runs the tests"""
start = time.time()
logging.basicConfig(level="INFO")
# Load the .yaml data and unpacks it
assert len(sys.argv) == 2, "Exactly one experiment configuration file must be "\
"passed as a positional argument to this script. \n\n"\
"E.g. `python run_non_tracking_experiment.py <path to .yaml file>`"
with open(sys.argv[1], "r") as yaml_config_file:
logging.info("Loading simulation settings from %s", sys.argv[1])
experiment_config = yaml.load(yaml_config_file)
experiment_settings = experiment_config['experiment_settings']
data_parameters = experiment_config['data_parameters']
ml_parameters = experiment_config['ml_parameters']
path_parameters = experiment_config['path_parameters'] \
if 'path_parameters' in experiment_config else None
# Loads the raw dataset
logging.info("Loading the dataset...")
data_preprocessor = Preprocessor(data_parameters)
features, labels = data_preprocessor.load_dataset()
if path_parameters:
path_creator = PathCreator(data_parameters, path_parameters, labels)
paths = path_creator.load_paths()
# Undersamples the dataset (if requested)
if "undersample_bf" in experiment_settings and experiment_settings[
"undersample_bf"]:
features = undersample_bf(features, data_parameters["beamformings"])
if "undersample_space" in experiment_settings:
assert not path_parameters, "This option is not supported for tracking experiments, "\
"unless the code for the path creation is updated"
features, labels = undersample_space(
features, labels, data_parameters["undersample_space"])
# Initializes the model and prepares it for testing
logging.info("Initializing the model (type = %s)...",
experiment_settings["model_type"].lower())
if experiment_settings["model_type"].lower() == "cnn":
ml_parameters["input_type"] = "float"
model = CNN(ml_parameters)
elif experiment_settings["model_type"].lower() in ("lstm", "tcn"):
assert path_parameters, "This model requires `paths_parameters`. See the example."
assert path_parameters["time_steps"] == ml_parameters["input_shape"][0], "The ML model "\
"first input dimention must match the length of the paths! (path length = {}, model)"\
"input = {})".format(path_parameters["time_steps"], ml_parameters["input_shape"][0])
ml_parameters["input_type"] = "bool"
if experiment_settings["model_type"].lower() == "lstm":
model = LSTM(ml_parameters)
else:
model = TCN(ml_parameters)
else:
raise ValueError(
"The simulation settings specified 'model_type'={}. Currently, only "
"'cnn', 'lstm', and 'tcn' are supported.".format(
experiment_settings["model_type"]))
experiment_name = os.path.basename(sys.argv[1]).split('.')[0]
model.load(model_name=experiment_name)
# Prediction loop
if "tests_per_position" in experiment_settings:
tests_per_input = experiment_settings["tests_per_position"]
else:
tests_per_input = experiment_settings["tests_per_path"] * 10
logging.info(
"Note - each set of paths will be split into 10 sub-sets, for easier RAM"
"management -- that's why you'll see 10x test sets in the next logging messages."
)
y_true = None
y_pred = None
for set_idx in range(tests_per_input):
logging.info("Creating test set %2s out of %2s...", set_idx + 1,
tests_per_input)
if path_parameters:
features_test, labels_test, _ = sample_paths(paths["test"],
features,
labels,
experiment_settings,
data_parameters,
path_parameters,
sample_fraction=0.1)
else:
features_test, labels_test = create_noisy_features(
features,
labels,
experiment_settings,
data_parameters,
)
logging.info("Running predictions and storing data...\n")
predictions_test = model.predict(features_test)
y_true = np.vstack(
(y_true, labels_test)) if y_true is not None else labels_test
y_pred = np.vstack(
(y_pred,
predictions_test)) if y_pred is not None else predictions_test
assert labels_test.shape[1] == y_true.shape[1], "The number of dimensions per sample "\
"must stay constant!"
assert y_true.shape == y_pred.shape, "The predictions and the labels must have the "\
"same shape!"
# Closes the model, gets the test scores, and stores predictions-labels pairs
model.close()
logging.info("Computing test metrics...")
test_score = score_predictions(y_true, y_pred,
ml_parameters["validation_metric"])
test_score *= data_parameters["pos_grid"][0]
test_95_perc = get_95th_percentile(
y_true, y_pred, rescale_factor=data_parameters["pos_grid"][0])
logging.info("Average test distance: %.5f m || 95th percentile: %.5f m\n",
test_score, test_95_perc)
preditions_file = os.path.join(
ml_parameters["model_folder"],
experiment_name + '_' + experiment_settings["predictions_file"])
with open(preditions_file, 'wb') as data_file:
pickle.dump([y_true, y_pred], data_file)
# Prints elapsed time
end = time.time()
exec_time = (end - start)
logging.info("Total execution time: %.5E seconds", exec_time)