def model_predict_and_evaluate(self, dataset):
"""Train model with subset of the features and evalaute the performance.
Args:
- dataset: dataset with subset of temporal features
Returns:
- performance: performance with subset of temporal features
"""
# Build model
pred_class = prediction(self.model_parameters['model_type'],
self.model_parameters, self.task)
# Train the model
pred_class.fit(dataset)
# Test the model
test_y_hat = pred_class.predict(dataset)
# Extract the labels
_, _, test_y, _, _ = dataset.get_fold(fold=0, split='test')
# Evaluate the performance
temp_performance = Metrics([self.metric_name],
self.metric_parameters).evaluate(
test_y, test_y_hat)
performance = np.mean(list(temp_performance.values())[0])
return performance
def evalparser(path='./examples',
report=False,
bcvocab=None,
draw=True,
withdp=False,
fdpvocab=None,
fprojmat=None):
""" Test the parsing performance
:type path: string
:param path: path to the evaluation data
:type report: boolean
:param report: whether to report (calculate) the f1 score
"""
# ----------------------------------------
# Load the parsing model
print('Load parsing model ...')
pm = ParsingModel(withdp=withdp, fdpvocab=fdpvocab, fprojmat=fprojmat)
pm.loadmodel("model/parsing-model.pickle.gz")
# ----------------------------------------
# Evaluation
met = Metrics(levels=['span', 'nuclearity', 'relation'])
# ----------------------------------------
# Read all files from the given path
exsisting_files = [
".".join(fname.split(".")[:-1]) for fname in listdir(path)
if fname.endswith('.brackets')
]
all_files = [
".".join(fname.split(".")[:-1]) for fname in listdir(path)
if fname.endswith('.merge')
]
todo_files = list(set(all_files) - set(exsisting_files))
doclist = [joinpath(path, fname + '.merge') for fname in todo_files]
print("TODO files len:")
print(len(doclist))
print(doclist[0])
global_pm = pm
global global_pm
global_bv = bcvocab
global global_bv
eval_parser_unit(doclist[0])
cnt = multiprocessing.cpu_count()
pool = multiprocessing.Pool(processes=cnt)
pool.map(eval_parser_unit, doclist)
pool.close()
pool.join()
"""
def evalparser(path='./examples', report=False,
bcvocab=None, draw=True,
withdp=False, fdpvocab=None, fprojmat=None):
""" Test the parsing performance
:type path: string
:param path: path to the evaluation data
:type report: boolean
:param report: whether to report (calculate) the f1 score
"""
# ----------------------------------------
# Load the parsing model
print 'Load parsing model ...'
pm = ParsingModel(withdp=withdp,
fdpvocab=fdpvocab, fprojmat=fprojmat)
pm.loadmodel("model/parsing-model.pickle.gz")
# ----------------------------------------
# Evaluation
met = Metrics(levels=['span','nuclearity','relation'])
# ----------------------------------------
# Read all files from the given path
doclist = [joinpath(path, fname) for fname in listdir(path) if fname.endswith('.merge')]
for fmerge in doclist:
# ----------------------------------------
# Read *.merge file
dr = DocReader()
doc = dr.read(fmerge)
# ----------------------------------------
# Parsing
pred_rst = pm.sr_parse(doc, bcvocab)
if draw:
strtree = pred_rst.parse()
drawrst(strtree, fmerge.replace(".merge",".ps"))
# Get brackets from parsing results
pred_brackets = pred_rst.bracketing()
fbrackets = fmerge.replace('.merge', '.brackets')
# Write brackets into file
writebrackets(fbrackets, pred_brackets)
# ----------------------------------------
# Evaluate with gold RST tree
if report:
fdis = fmerge.replace('.merge', '.dis')
gold_rst = RSTTree(fdis, fmerge)
gold_rst.build()
gold_brackets = gold_rst.bracketing()
met.eval(gold_rst, pred_rst)
if report:
met.report()
def evalparser(path='./examples', report=False):
""" Test the parsing performance
:type path: string
:param path: path to the evaluation data
:type report: boolean
:param report: whether to report (calculate) the f1 score
"""
from os import listdir
from os.path import join as joinpath
# ----------------------------------------
# Load the parsing model
pm = ParsingModel()
pm.loadmodel("parsing-model.pickle.gz")
# ----------------------------------------
# Evaluation
met = Metrics(levels=['span', 'nuclearity', 'relation'])
# ----------------------------------------
# Read all files from the given path
doclist = [
joinpath(path, fname) for fname in listdir(path)
if fname.endswith('.edus')
]
for fedus in doclist:
# ----------------------------------------
# Parsing
pred_rst = parse(pm, fedus=fedus)
# Get brackets from parsing results
# print fedus
fin = open("test.dis", "w")
r = fin.write(str(pred_rst))
# pred_brackets = pred_rst.bracketing()
# fbrackets = fedus.replace('edus', 'brackets')
# writebrackets(fbrackets, pred_brackets)
# ----------------------------------------
# Evaluate with gold RST tree
if report:
fdis = fedus.replace('edus', 'dis')
gold_rst = RSTTree(fname=fdis)
gold_rst.build()
gold_brackets = gold_rst.bracketing()
met.eval(gold_rst, pred_rst)
if report:
met.report()
def run(self):
"""
Runs the full pipeline as configured.
:return: list of run parameters and evaluation metrics.
"""
# TODO try/catch to ensure proper shutdown even if error encountered
params = self._get_params_for_run()
result_rows = []
# Check for valid configuration
if self._test_docs is None and self._k_folds == 0:
self._logger.error("Explicit test set or number of cross-validation folds must be specified.")
metrics = Metrics()
result_row = {**params, **metrics.get_scores_as_dict()}
result_rows.append(result_row)
return result_rows
# Continue while there are configured parameter settings to evaluate
while params is not None:
# Get collection of training and test sets for current run
data_sets = self._get_training_and_test_sets()
for set_index, (training_docs, test_docs) in enumerate(data_sets):
# Retrieve an encoder module trained with the specified configuration
self._encoder = self._get_encoder(params)
set_index += 1 # Only used for user output, so start index at 1
num_sets = len(data_sets)
if num_sets > 1:
self._logger.info("Training and evaluating fold {} of {}.".format(set_index, num_sets))
start = time.time()
self._train_and_evaluate(params, self._encoder, training_docs, test_docs)
runtime = time.time() - start
self._logger.info(
"Trained and evaluated fold {} of sequence model in {} seconds.".format(set_index, runtime))
# Combine run parameters with evaluation results and store
result_row = {**params, **self._evaluator.get_score_as_dict()}
result_rows.append(result_row)
# Check if model should be saved
if self._is_model_saving_enabled():
operator = self._evaluator.get_operator()
current_score = self._evaluator.get_score()
best_model = self._get_best_model()
if best_model is not None:
(best_metric, _) = best_model
if not operator(best_metric, current_score):
self._set_best_model(current_score, (params, self._encoder, self._sequence_learner))
else:
# New model is the best one if no previous existed
self._set_best_model(current_score, (params, self._encoder, self._sequence_learner))
# Invoke optimizer callback to report on results of this run
if self._optimizer is not None:
self._optimizer.process_run_result(params=params,
score=self._evaluator.get_score_as_dict(),
encoder=self._encoder,
sequence_learner=self._sequence_learner)
# Check if there are additional runs to execute
if self._optimizer is not None:
params = self._optimizer.get_next_params()
else:
params = None
# Store best model, if configured
if self._is_model_saving_enabled():
path, name = self._get_model_save_path_and_name()
try:
self.save(path, name)
except Exception:
self._logger.error("Failed to save model, clearing Keras session and trying again.")
self._sequence_learner.clear_session()
self.save(path, name)
# Clear Keras/Tensorflow models # TODO why a second time?
if self._sequence_learner is not None:
self._sequence_learner.clear_session()
return pd.DataFrame(result_rows)
def test_evaluate_model(report1):
metrics = Metrics(report1.get_gold(), report1.get_pred())
print(metrics.bleu_score())
print(metrics.ds_score())
def main(args):
'''Main function for AutoML in time-series predictions.
Args:
- data loading parameters:
- data_names: mimic, ward, cf
- preprocess parameters:
- normalization: minmax, standard, None
- one_hot_encoding: input features that need to be one-hot encoded
- problem: 'one-shot' or 'online'
- 'one-shot': one time prediction at the end of the time-series
- 'online': preditcion at every time stamps of the time-series
- max_seq_len: maximum sequence length after padding
- label_name: the column name for the label(s)
- treatment: the column name for treatments
- imputation parameters:
- static_imputation_model: mean, median, mice, missforest, knn, gain
- temporal_imputation_model: mean, median, linear, quadratic, cubic, spline, mrnn, tgain
- feature selection parameters:
- feature_selection_model: greedy-addtion, greedy-deletion, recursive-addition, recursive-deletion, None
- feature_number: selected featuer number
- predictor_parameters:
- epochs: number of epochs
- bo_itr: bayesian optimization iterations
- static_mode: how to utilize static features (concatenate or None)
- time_mode: how to utilize time information (concatenate or None)
- task: classification or regression
- metric_name: auc, apr, mae, mse
'''
#%% Step 0: Set basic parameters
metric_sets = [args.metric_name]
metric_parameters = {
'problem': args.problem,
'label_name': [args.label_name]
}
#%% Step 1: Upload Dataset
# File names
data_directory = '../datasets/data/' + args.data_name + '/' + args.data_name + '_'
data_loader_training = CSVLoader(
static_file=data_directory + 'static_train_data.csv.gz',
temporal_file=data_directory + 'temporal_train_data_eav.csv.gz')
data_loader_testing = CSVLoader(
static_file=data_directory + 'static_test_data.csv.gz',
temporal_file=data_directory + 'temporal_test_data_eav.csv.gz')
dataset_training = data_loader_training.load()
dataset_testing = data_loader_testing.load()
print('Finish data loading.')
#%% Step 2: Preprocess Dataset
# (0) filter out negative values (Automatically)
negative_filter = FilterNegative()
# (1) one-hot encode categorical features
onehot_encoder = OneHotEncoder(
one_hot_encoding_features=[args.one_hot_encoding])
# (2) Normalize features: 3 options (minmax, standard, none)
normalizer = Normalizer(args.normalization)
filter_pipeline = PipelineComposer(negative_filter, onehot_encoder,
normalizer)
dataset_training = filter_pipeline.fit_transform(dataset_training)
dataset_testing = filter_pipeline.transform(dataset_testing)
print('Finish preprocessing.')
#%% Step 3: Define Problem
problem_maker = ProblemMaker(problem=args.problem,
label=[args.label_name],
max_seq_len=args.max_seq_len,
treatment=[args.treatment])
dataset_training = problem_maker.fit_transform(dataset_training)
dataset_testing = problem_maker.fit_transform(dataset_testing)
print('Finish defining problem.')
#%% Step 4: Impute Dataset
static_imputation = Imputation(
imputation_model_name=args.static_imputation_model, data_type='static')
temporal_imputation = Imputation(
imputation_model_name=args.temporal_imputation_model,
data_type='temporal')
imputation_pipeline = PipelineComposer(static_imputation,
temporal_imputation)
dataset_training = imputation_pipeline.fit_transform(dataset_training)
dataset_testing = imputation_pipeline.transform(dataset_testing)
print('Finish imputation.')
#%% Step 5: Feature selection (4 options)
static_feature_selection = \
FeatureSelection(feature_selection_model_name = args.static_feature_selection_model,
feature_type = 'static', feature_number = args.static_feature_selection_number,
task = args.task, metric_name = args.metric_name,
metric_parameters = metric_parameters)
temporal_feature_selection = \
FeatureSelection(feature_selection_model_name = args.temporal_feature_selection_model,
feature_type = 'temporal', feature_number = args.temporal_feature_selection_number,
task = args.task, metric_name = args.metric_name,
metric_parameters = metric_parameters)
feature_selection_pipeline = PipelineComposer(static_feature_selection,
temporal_feature_selection)
dataset_training = feature_selection_pipeline.fit_transform(
dataset_training)
dataset_testing = feature_selection_pipeline.transform(dataset_testing)
print('Finish feature selection.')
#%% Step 6: Bayesian Optimization
## Model define
model_parameters = {
'projection_horizon': 5,
'static_mode': 'concatenate',
'time_mode': 'concatenate'
}
crn_model = CRN_Model(task=args.task)
crn_model.set_params(**model_parameters)
model_class = crn_model
# train_validate split
dataset_training.train_val_test_split(prob_val=0.2, prob_test=0.2)
# Bayesian Optimization Start
metric = BOMetric(metric='auc', fold=0, split='test')
# Run BO for selected model class
BO_model = AutoTS(dataset_training, model_class, metric)
models, bo_score = BO_model.training_loop(num_iter=2)
auto_ens_model = AutoEnsemble(models, bo_score)
# Prediction
assert not dataset_testing.is_validation_defined
test_y_hat = auto_ens_model.predict(dataset_testing, test_split='test')
test_y = dataset_testing.label
print('Finish AutoML model training and testing.')
#%% Step 7: Visualize Results
idx = np.random.permutation(len(test_y_hat))[:2]
# Evaluate predictor model
result = Metrics(metric_sets,
metric_parameters).evaluate(test_y, test_y_hat)
print('Finish predictor model evaluation.')
# Visualize the output
# (1) Performance
print('Overall performance')
print_performance(result, metric_sets, metric_parameters)
# (2) Predictions
print('Each prediction')
print_prediction(test_y_hat[idx], metric_parameters)
return
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
st.subheader('Overall Prediction Performance')
if select_pred_task == 'Classification':
metric_sets = ['auc', 'apr']
if select_pred_task == 'Regression':
metric_sets = ['mse', 'mae']
metric_parameters = {
'problem': problem_type,
'label_name': label_name,
}
metrics = Metrics(metric_sets, metric_parameters)
result = metrics.evaluate(dataset_v_5.label, test_y_hat)
if problem_type == 'one-shot':
text = print_performance(
result,
metric_sets,
metric_parameters,
)
st.text(text)
if problem_type == 'online':
figs = print_performance(
result,
def main(args):
"""Main function for AutoML in time-series predictions.
Args:
- data loading parameters:
- data_names: mimic, ward, cf
- preprocess parameters:
- normalization: minmax, standard, None
- one_hot_encoding: input features that need to be one-hot encoded
- problem: 'one-shot' or 'online'
- 'one-shot': one time prediction at the end of the time-series
- 'online': preditcion at every time stamps of the time-series
- max_seq_len: maximum sequence length after padding
- label_name: the column name for the label(s)
- treatment: the column name for treatments
- imputation parameters:
- static_imputation_model: mean, median, mice, missforest, knn, gain
- temporal_imputation_model: mean, median, linear, quadratic, cubic, spline, mrnn, tgain
- feature selection parameters:
- feature_selection_model: greedy-addition, greedy-deletion, recursive-addition, recursive-deletion, None
- feature_number: selected featuer number
- predictor_parameters:
- epochs: number of epochs
- bo_itr: bayesian optimization iterations
- static_mode: how to utilize static features (concatenate or None)
- time_mode: how to utilize time information (concatenate or None)
- task: classification or regression
- metric_name: auc, apr, mae, mse
"""
#%% Step 0: Set basic parameters
metric_sets = [args.metric_name]
metric_parameters = {
"problem": args.problem,
"label_name": [args.label_name]
}
#%% Step 1: Upload Dataset
# File names
data_directory = "../datasets/data/" + args.data_name + "/" + args.data_name + "_"
data_loader_training = CSVLoader(
static_file=data_directory + "static_train_data.csv.gz",
temporal_file=data_directory + "temporal_train_data_eav.csv.gz",
)
data_loader_testing = CSVLoader(
static_file=data_directory + "static_test_data.csv.gz",
temporal_file=data_directory + "temporal_test_data_eav.csv.gz",
)
dataset_training = data_loader_training.load()
dataset_testing = data_loader_testing.load()
print("Finish data loading.")
#%% Step 2: Preprocess Dataset
# (0) filter out negative values (Automatically)
negative_filter = FilterNegative()
# (1) one-hot encode categorical features
onehot_encoder = OneHotEncoder(
one_hot_encoding_features=[args.one_hot_encoding])
# (2) Normalize features: 3 options (minmax, standard, none)
normalizer = Normalizer(args.normalization)
filter_pipeline = PipelineComposer(negative_filter, onehot_encoder,
normalizer)
dataset_training = filter_pipeline.fit_transform(dataset_training)
dataset_testing = filter_pipeline.transform(dataset_testing)
print("Finish preprocessing.")
#%% Step 3: Define Problem
problem_maker = ProblemMaker(problem=args.problem,
label=[args.label_name],
max_seq_len=args.max_seq_len,
treatment=args.treatment)
dataset_training = problem_maker.fit_transform(dataset_training)
dataset_testing = problem_maker.fit_transform(dataset_testing)
print("Finish defining problem.")
#%% Step 4: Impute Dataset
static_imputation = Imputation(
imputation_model_name=args.static_imputation_model, data_type="static")
temporal_imputation = Imputation(
imputation_model_name=args.temporal_imputation_model,
data_type="temporal")
imputation_pipeline = PipelineComposer(static_imputation,
temporal_imputation)
dataset_training = imputation_pipeline.fit_transform(dataset_training)
dataset_testing = imputation_pipeline.transform(dataset_testing)
print("Finish imputation.")
#%% Step 5: Feature selection (4 options)
static_feature_selection = FeatureSelection(
feature_selection_model_name=args.static_feature_selection_model,
feature_type="static",
feature_number=args.static_feature_selection_number,
task=args.task,
metric_name=args.metric_name,
metric_parameters=metric_parameters,
)
temporal_feature_selection = FeatureSelection(
feature_selection_model_name=args.temporal_feature_selection_model,
feature_type="temporal",
feature_number=args.temporal_feature_selection_number,
task=args.task,
metric_name=args.metric_name,
metric_parameters=metric_parameters,
)
feature_selection_pipeline = PipelineComposer(static_feature_selection,
temporal_feature_selection)
dataset_training = feature_selection_pipeline.fit_transform(
dataset_training)
dataset_testing = feature_selection_pipeline.transform(dataset_testing)
print("Finish feature selection.")
#%% Step 6: Bayesian Optimization
## Model define
# RNN model
rnn_parameters = {
"model_type": "lstm",
"epoch": args.epochs,
"static_mode": args.static_mode,
"time_mode": args.time_mode,
"verbose": False,
}
general_rnn = GeneralRNN(task=args.task)
general_rnn.set_params(**rnn_parameters)
# CNN model
cnn_parameters = {
"epoch": args.epochs,
"static_mode": args.static_mode,
"time_mode": args.time_mode,
"verbose": False,
}
temp_cnn = TemporalCNN(task=args.task)
temp_cnn.set_params(**cnn_parameters)
# Transformer
transformer = TransformerPredictor(task=args.task,
epoch=args.epochs,
static_mode=args.static_mode,
time_mode=args.time_mode)
# Attention model
attn_parameters = {
"model_type": "lstm",
"epoch": args.epochs,
"static_mode": args.static_mode,
"time_mode": args.time_mode,
"verbose": False,
}
attn = Attention(task=args.task)
attn.set_params(**attn_parameters)
# model_class_list = [general_rnn, attn, temp_cnn, transformer]
model_class_list = [general_rnn, attn]
# train_validate split
dataset_training.train_val_test_split(prob_val=0.2, prob_test=0.1)
# Bayesian Optimization Start
metric = BOMetric(metric="auc", fold=0, split="test")
ens_model_list = []
# Run BO for each model class
for m in model_class_list:
BO_model = automl.model.AutoTS(dataset_training,
m,
metric,
model_path="tmp/")
models, bo_score = BO_model.training_loop(num_iter=args.bo_itr)
auto_ens_model = AutoEnsemble(models, bo_score)
ens_model_list.append(auto_ens_model)
# Load all ensemble models
for ens in ens_model_list:
for m in ens.models:
m.load_model(BO_model.model_path + "/" + m.model_id + ".h5")
# Stacking algorithm
stacking_ens_model = StackingEnsemble(ens_model_list)
stacking_ens_model.fit(dataset_training, fold=0, train_split="val")
# Prediction
assert not dataset_testing.is_validation_defined
test_y_hat = stacking_ens_model.predict(dataset_testing, test_split="test")
test_y = dataset_testing.label
print("Finish AutoML model training and testing.")
#%% Step 7: Visualize Results
idx = np.random.permutation(len(test_y_hat))[:2]
# Evaluate predictor model
result = Metrics(metric_sets,
metric_parameters).evaluate(test_y, test_y_hat)
print("Finish predictor model evaluation.")
# Visualize the output
# (1) Performance
print("Overall performance")
print_performance(result, metric_sets, metric_parameters)
# (2) Predictions
print("Each prediction")
print_prediction(test_y_hat[idx], metric_parameters)
return
def main(args):
'''Main function for time-series prediction.
Args:
- data loading parameters:
- data_names: mimic, ward, cf
- preprocess parameters:
- normalization: minmax, standard, None
- one_hot_encoding: input features that need to be one-hot encoded
- problem: 'one-shot' or 'online'
- 'one-shot': one time prediction at the end of the time-series
- 'online': preditcion at every time stamps of the time-series
- max_seq_len: maximum sequence length after padding
- label_name: the column name for the label(s)
- treatment: the column name for treatments
- imputation parameters:
- static_imputation_model: mean, median, mice, missforest, knn, gain
- temporal_imputation_model: mean, median, linear, quadratic, cubic, spline, mrnn, tgain
- feature selection parameters:
- feature_selection_model: greedy-addtion, greedy-deletion, recursive-addition, recursive-deletion, None
- feature_number: selected featuer number
- predictor_parameters:
- model_name: rnn, gru, lstm, attention, tcn, transformer
- model_parameters: network parameters such as numer of layers
- h_dim: hidden dimensions
- n_layer: layer number
- n_head: head number (only for transformer model)
- batch_size: number of samples in mini-batch
- epochs: number of epochs
- learning_rate: learning rate
- static_mode: how to utilize static features (concatenate or None)
- time_mode: how to utilize time information (concatenate or None)
- task: classification or regression
- uncertainty_model_name: uncertainty estimation model name (ensemble)
- interpretor_model_name: interpretation model name (tinvase)
- metric_name: auc, apr, mae, mse
'''
#%% Step 0: Set basic parameters
metric_sets = [args.metric_name]
metric_parameters = {
'problem': args.problem,
'label_name': [args.label_name]
}
#%% Step 1: Upload Dataset
# File names
data_directory = '../datasets/data/' + args.data_name + '/' + args.data_name + '_'
data_loader_training = CSVLoader(
static_file=data_directory + 'static_train_data.csv.gz',
temporal_file=data_directory + 'temporal_train_data_eav.csv.gz')
data_loader_testing = CSVLoader(
static_file=data_directory + 'static_test_data.csv.gz',
temporal_file=data_directory + 'temporal_test_data_eav.csv.gz')
dataset_training = data_loader_training.load()
dataset_testing = data_loader_testing.load()
print('Finish data loading.')
#%% Step 2: Preprocess Dataset
# (0) filter out negative values (Automatically)
negative_filter = FilterNegative()
# (1) one-hot encode categorical features
onehot_encoder = OneHotEncoder(
one_hot_encoding_features=[args.one_hot_encoding])
# (2) Normalize features: 3 options (minmax, standard, none)
normalizer = Normalizer(args.normalization)
filter_pipeline = PipelineComposer(negative_filter, onehot_encoder,
normalizer)
dataset_training = filter_pipeline.fit_transform(dataset_training)
dataset_testing = filter_pipeline.transform(dataset_testing)
print('Finish preprocessing.')
#%% Step 3: Define Problem
problem_maker = ProblemMaker(problem=args.problem,
label=[args.label_name],
max_seq_len=args.max_seq_len,
treatment=args.treatment)
dataset_training = problem_maker.fit_transform(dataset_training)
dataset_testing = problem_maker.fit_transform(dataset_testing)
print('Finish defining problem.')
#%% Step 4: Impute Dataset
static_imputation = Imputation(
imputation_model_name=args.static_imputation_model, data_type='static')
temporal_imputation = Imputation(
imputation_model_name=args.temporal_imputation_model,
data_type='temporal')
imputation_pipeline = PipelineComposer(static_imputation,
temporal_imputation)
dataset_training = imputation_pipeline.fit_transform(dataset_training)
dataset_testing = imputation_pipeline.transform(dataset_testing)
print('Finish imputation.')
#%% Step 5: Feature selection (4 options)
static_feature_selection = \
FeatureSelection(feature_selection_model_name = args.static_feature_selection_model,
feature_type = 'static', feature_number = args.static_feature_selection_number,
task = args.task, metric_name = args.metric_name,
metric_parameters = metric_parameters)
temporal_feature_selection = \
FeatureSelection(feature_selection_model_name = args.temporal_feature_selection_model,
feature_type = 'temporal', feature_number = args.temporal_feature_selection_number,
task = args.task, metric_name = args.metric_name,
metric_parameters = metric_parameters)
feature_selection_pipeline = PipelineComposer(static_feature_selection,
temporal_feature_selection)
dataset_training = feature_selection_pipeline.fit_transform(
dataset_training)
dataset_testing = feature_selection_pipeline.transform(dataset_testing)
print('Finish feature selection.')
#%% Step 6: Fit and Predict (6 options)
# Set predictor model parameters
model_parameters = {
'h_dim': args.h_dim,
'n_layer': args.n_layer,
'n_head': args.n_head,
'batch_size': args.batch_size,
'epoch': args.epochs,
'model_type': args.model_name,
'learning_rate': args.learning_rate,
'static_mode': args.static_mode,
'time_mode': args.time_mode,
'verbose': True
}
# Set the validation data for best model saving
dataset_training.train_val_test_split(prob_val=0.2, prob_test=0.0)
pred_class = prediction(args.model_name, model_parameters, args.task)
pred_class.fit(dataset_training)
test_y_hat = pred_class.predict(dataset_testing)
print('Finish predictor model training and testing.')
#%% Step 7: Estimate Uncertainty (1 option)
uncertainty_model = uncertainty(args.uncertainty_model_name,
model_parameters, pred_class, args.task)
uncertainty_model.fit(dataset_training)
test_ci_hat = uncertainty_model.predict(dataset_testing)
print('Finish uncertainty estimation')
#%% Step 8: Interpret Predictions (1 option)
interpretor = interpretation(args.interpretation_model_name,
model_parameters, pred_class, args.task)
interpretor.fit(dataset_training)
test_s_hat = interpretor.predict(dataset_testing)
print('Finish model interpretation')
#%% Step 9: Visualize Results
idx = np.random.permutation(len(test_y_hat))[:2]
# Evaluate predictor model
result = Metrics(metric_sets,
metric_parameters).evaluate(dataset_testing.label,
test_y_hat)
print('Finish predictor model evaluation.')
# Visualize the output
# (1) Performance
print('Overall performance')
print_performance(result, metric_sets, metric_parameters)
# (2) Predictions
print('Each prediction')
print_prediction(test_y_hat[idx], metric_parameters)
# (3) Uncertainty
print('Uncertainty estimations')
print_uncertainty(test_y_hat[idx], test_ci_hat[idx], metric_parameters)
# (4) Model interpretation
print('Model interpretation')
print_interpretation(test_s_hat[idx], dataset_training.feature_name,
metric_parameters, model_parameters)
return
def main(args):
'''Main function for individual treatment effect estimation.
Args:
- data loading parameters:
- data_names: mimic, ward, cf, mimic_antibiotics
- preprocess parameters:
- normalization: minmax, standard, None
- one_hot_encoding: input features that need to be one-hot encoded
- problem: 'online'
- 'online': preiction at every time stamps of the time-series
- max_seq_len: maximum sequence length after padding
- label_name: the column name for the label(s)
- treatment: the column name for treatments
- imputation parameters:
- static_imputation_model: mean, median, mice, missforest, knn, gain
- temporal_imputation_model: mean, median, linear, quadratic, cubic, spline, mrnn, tgain
- feature selection parameters:
- feature_selection_model: greedy-addtion, greedy-deletion, recursive-addition, recursive-deletion, None
- feature_number: selected featuer number
- treatment effects model parameters:
- model_name: CRN, RMSN, GANITE
Each model has different types of hyperparameters that need to be set.
- Parameters needed for the Counterfactual Recurrent Network (CRN):
- hyperparameters for encoder:
- rnn_hidden_units: hidden dimensions in the LSTM unit
- rnn_keep_prob: keep probability used for variational dropout in the LSTM unit
- br_size: size of the balancing representation
- fc_hidden_units: hidden dimensions of the fully connected layers used for treatment classifier and predictor
- batch_size: number of samples in mini-batch
- num_epochs: number of epochs
- learning_rate: learning rate
- max_alpha: alpha controls the trade-off between building tratment invariant representations (domain
discrimination) and being able to predict outcomes (outcome prediction); during training, CRN uses an
exponentially increasing schedule for alpha from 0 to max_alpha.
- hyperparameters for decoder:
- the decoder requires the same hyperparameters as the encoder with the exception of the rnn_hidden_units
which is set to be equal to the br_size of the encoder
- Parameters for Recurrent Marginal Structural Networks (RMSN):
- hyperparameters for encoder:
- dropout_rate: dropout probability used for variational
- rnn_hidden_units: hidden dimensions in the LSTM unit
- batch_size: number of samples in mini-batch
- num_epochs: number of epochs
- learning_rate: learning rate
- max_norm: max gradient norm used for gradient clipping during training
- hyperparameters for decoder:
- the decoder requires the same hyperparameters as the encoder.
- model_dir: directory where the model is saved
- model_name: name of the saved model
- Parameters for GANITE:
- batch size: number of samples in mini-batch
- alpha: parameter trading off between discriminator loss and supervised loss for the generator training
- learning_rate: learning rate
- hidden_units: hidden dimensions of the fully connected layers used in the networks
- stack_dim: number of timesteps to stack
All models have the following common parameters:
- static_mode: how to utilize static features (concatenate or None)
- time_mode: how to utilize time information (concatenate or None)
- taks: 'classification' or 'regression'
- metric_name: auc, apr, mae, mse (used for factual prediction)
- patient id: patient for which counterfactual trajectories are computed
- timestep: timestep in patient trajectory for estimating counterfactuals
'''
# %% Step 0: Set basic parameters
metric_sets = [args.metric_name]
metric_parameters = {
'problem': args.problem,
'label_name': [args.label_name]
}
# %% Step 1: Upload Dataset
# File names
data_directory = '../datasets/data/' + args.data_name + '/' + args.data_name + '_'
data_loader_training = CSVLoader(
static_file=data_directory + 'static_train_data.csv.gz',
temporal_file=data_directory + 'temporal_train_data_eav.csv.gz')
data_loader_testing = CSVLoader(
static_file=data_directory + 'static_test_data.csv.gz',
temporal_file=data_directory + 'temporal_test_data_eav.csv.gz')
dataset_training = data_loader_training.load()
dataset_testing = data_loader_testing.load()
print('Finish data loading.')
# %% Step 2: Preprocess Dataset
# (0) filter out negative values (Automatically)
negative_filter = FilterNegative()
# (1) one-hot encode categorical features
onehot_encoder = OneHotEncoder(
one_hot_encoding_features=[args.one_hot_encoding])
# (2) Normalize features: 3 options (minmax, standard, none)
normalizer = Normalizer(args.normalization)
filter_pipeline = PipelineComposer(negative_filter, onehot_encoder,
normalizer)
dataset_training = filter_pipeline.fit_transform(dataset_training)
dataset_testing = filter_pipeline.transform(dataset_testing)
print('Finish preprocessing.')
# %% Step 3: Define Problem
problem_maker = ProblemMaker(problem=args.problem,
label=[args.label_name],
max_seq_len=args.max_seq_len,
treatment=[args.treatment])
dataset_training = problem_maker.fit_transform(dataset_training)
dataset_testing = problem_maker.fit_transform(dataset_testing)
print('Finish defining problem.')
# %% Step 4: Impute Dataset
static_imputation = Imputation(
imputation_model_name=args.static_imputation_model, data_type='static')
temporal_imputation = Imputation(
imputation_model_name=args.temporal_imputation_model,
data_type='temporal')
imputation_pipeline = PipelineComposer(static_imputation,
temporal_imputation)
dataset_training = imputation_pipeline.fit_transform(dataset_training)
dataset_testing = imputation_pipeline.transform(dataset_testing)
print('Finish imputation.')
# %% Step 5: Feature selection (4 options)
static_feature_selection = \
FeatureSelection(feature_selection_model_name=args.static_feature_selection_model,
feature_type='static', feature_number=args.static_feature_selection_number,
task=args.task, metric_name=args.metric_name,
metric_parameters=metric_parameters)
temporal_feature_selection = \
FeatureSelection(feature_selection_model_name=args.temporal_feature_selection_model,
feature_type='temporal', feature_number=args.temporal_feature_selection_number,
task=args.task, metric_name=args.metric_name,
metric_parameters=metric_parameters)
feature_selection_pipeline = PipelineComposer(static_feature_selection,
temporal_feature_selection)
dataset_training = feature_selection_pipeline.fit_transform(
dataset_training)
dataset_testing = feature_selection_pipeline.transform(dataset_testing)
print('Finish feature selection.')
# %% Step 6: Fit treatment effects (3 options)
# Set the validation data for best model saving
dataset_training.train_val_test_split(prob_val=0.2, prob_test=0.0)
# Set the treatment effects model
model_name = args.model_name
# Set treatment effects model parameters
if model_name == 'CRN':
model_parameters = {
'encoder_rnn_hidden_units': args.crn_encoder_rnn_hidden_units,
'encoder_br_size': args.crn_encoder_br_size,
'encoder_fc_hidden_units': args.crn_encoder_fc_hidden_units,
'encoder_learning_rate': args.crn_encoder_learning_rate,
'encoder_batch_size': args.crn_encoder_batch_size,
'encoder_keep_prob': args.crn_encoder_keep_prob,
'encoder_num_epochs': args.crn_encoder_num_epochs,
'encoder_max_alpha': args.crn_encoder_max_alpha,
'decoder_br_size': args.crn_decoder_br_size,
'decoder_fc_hidden_units': args.crn_decoder_fc_hidden_units,
'decoder_learning_rate': args.crn_decoder_learning_rate,
'decoder_batch_size': args.crn_decoder_batch_size,
'decoder_keep_prob': args.crn_decoder_keep_prob,
'decoder_num_epochs': args.crn_decoder_num_epochs,
'decoder_max_alpha': args.crn_decoder_max_alpha,
'projection_horizon': args.projection_horizon,
'static_mode': args.static_mode,
'time_mode': args.time_mode
}
treatment_model = treatment_effects_model(model_name,
model_parameters,
task='classification')
treatment_model.fit(dataset_training)
elif model_name == 'RMSN':
hyperparams_encoder_iptw = {
'dropout_rate': args.rmsn_encoder_dropout_rate,
'memory_multiplier': args.rmsn_encoder_memory_multiplier,
'num_epochs': args.rmsn_encoder_num_epochs,
'batch_size': args.rmsn_encoder_batch_size,
'learning_rate': args.rmsn_encoder_learning_rate,
'max_norm': args.rmsn_encoder_max_norm
}
hyperparams_decoder_iptw = {
'dropout_rate': args.rmsn_decoder_dropout_rate,
'memory_multiplier': args.rmsn_decoder_memory_multiplier,
'num_epochs': args.rmsn_decoder_num_epochs,
'batch_size': args.rmsn_decoder_batch_size,
'learning_rate': args.rmsn_decoder_learning_rate,
'max_norm': args.rmsn_decoder_max_norm
}
model_parameters = {
'hyperparams_encoder_iptw': hyperparams_encoder_iptw,
'hyperparams_decoder_iptw': hyperparams_decoder_iptw,
'model_dir': args.rmsn_model_dir,
'model_name': args.rmsn_model_name,
'static_mode': args.static_mode,
'time_mode': args.time_mode
}
treatment_model = treatment_effects_model(model_name,
model_parameters,
task='classification')
treatment_model.fit(dataset_training,
projection_horizon=args.projection_horizon)
elif model_name == 'GANITE':
hyperparams = {
'batch_size': args.ganite_batch_size,
'alpha': args.ganite_alpha,
'hidden_dims': args.ganite_hidden_dims,
'learning_rate': args.ganite_learning_rate
}
model_parameters = {
'hyperparams': hyperparams,
'stack_dim': args.ganite_stack_dim,
'static_mode': args.static_mode,
'time_mode': args.time_mode
}
treatment_model = treatment_effects_model(model_name,
model_parameters,
task='classification')
treatment_model.fit(dataset_training)
test_y_hat = treatment_model.predict(dataset_testing)
print('Finish treatment effects model training and testing.')
# %% Step 9: Visualize Results
# Evaluate predictor model
result = Metrics(metric_sets,
metric_parameters).evaluate(dataset_testing.label,
test_y_hat)
print('Finish predictor model evaluation.')
# Visualize the output
# (1) Performance on estimating factual outcomes
print('Overall performance on estimating factual outcomes')
print_performance(result, metric_sets, metric_parameters)
# (2) Counterfactual trajectories
print('Counterfactual trajectories')
if model_name in ['CRN', 'RMSN']:
# Predict and visualize counterfactuals for the sequence of treatments indicated by the user
# through the treatment_options. The lengths of each sequence of treatments needs to be projection_horizon + 1.
treatment_options = np.array([[[1], [1], [1], [1], [1], [0]],
[[0], [0], [0], [0], [1], [1]]])
history, counterfactual_traj = treatment_model.predict_counterfactual_trajectories(
dataset=dataset_testing,
patient_id=args.patient_id,
timestep=args.timestep,
treatment_options=treatment_options)
print_counterfactual_predictions(
patient_history=history,
treatment_options=treatment_options,
counterfactual_predictions=counterfactual_traj)
return
