def step(self, mode):
if mode == "train" and self.mode == "test":
raise Exception("Cannot train during test mode")
if mode == "train":
theano_fn = self.train_fn
batch_gen = self.train_batch_gen
elif mode == "test":
theano_fn = self.test_fn
batch_gen = self.test_batch_gen
else:
raise Exception("Invalid mode")
data = next(batch_gen)
ys = data[-1]
data = data[:-1]
ret = theano_fn(*data)
return {
"prediction":
[metrics.get_estimate_custom(x, self.nbins) for x in ret[0]],
"answers":
ys,
"current_loss":
ret[1],
"loss_reg":
ret[2],
"loss_mse":
ret[1] - ret[2],
"log":
""
}
def predict(self, data):
""" data is a pair (X, y) """
processed = self.process_input(data)[:-1]
ret = self.test_fn(*processed)
predictions = [
metrics.get_estimate_custom(x, self.nbins) for x in ret[0]
]
return [predictions] + list(ret[1:])
def calc_metrics(self, data_gen, history, dataset, logs):
y_true = []
predictions = []
# for i in range(data_gen.steps):
# if self.verbose == 1:
# print("\tdone {}/{}".format(i, data_gen.steps), end='\r')
# (x, y_processed, y) = data_gen.getitem(i, return_y_true=True)
# pred = self.model.predict(x, batch_size=self.batch_size, verbose=self.verbose)
# pass
# aflanders: debug
# if i == 0:
# print(f"type(x): {type(x)} type(self.batch_size): {type(self.batch_size)}")
# print(f"tf.executing_eagerly():{tf.executing_eagerly()}")
# aflanders: debug
#pred = self.model.predict(x, batch_size=self.batch_size, verbose=self.verbose)
pred = self.model.predict(data_gen,
batch_size=self.batch_size,
verbose=self.verbose,
steps=data_gen.steps,
workers=self.workers,
use_multiprocessing=True)
# if isinstance(x, list) and len(x) == 2: # deep supervision
# if pred.shape[-1] == 1: # regression
# pred_flatten = pred.flatten()
# else: # classification
# pred_flatten = pred.reshape((-1, 10))
# for m, t, p in zip(x[1].flatten(), y.flatten(), pred_flatten):
# if np.equal(m, 1):
# y_true.append(t)
# predictions.append(p)
# else:
y = data_gen.get_y(len(pred))
# if pred.shape[-1] == 1:
# y_true += list(y.flatten())
# predictions += list(pred.flatten())
# else:
y_true += list(y)
predictions += list(pred)
print('\n')
if self.partition == 'log':
predictions = [
metrics.get_estimate_log(x, 10) for x in predictions
]
ret = metrics.print_metrics_log_bins(y_true, predictions)
if self.partition == 'custom':
predictions = [
metrics.get_estimate_custom(x, 10) for x in predictions
]
ret = metrics.print_metrics_custom_bins(y_true, predictions)
if self.partition == 'none':
ret = metrics.print_metrics_regression(y_true, predictions)
for k, v in ret.items():
logs[dataset + '_' + k] = v
history.append(ret)
def calc_metrics(self, data_gen, history, dataset, logs):
y_true = []
predictions = []
for i in range(data_gen.steps):
if self.verbose == 1:
print("\tdone {}/{}".format(i, data_gen.steps), end='\r')
if self.use_time:
([x, t], y_processed, y) = data_gen.next(return_y_true=True)
pred = self.model.predict([x, t], batch_size=self.batch_size)
else:
(x, y_processed, y) = data_gen.next(return_y_true=True)
pred = self.model.predict(x, batch_size=self.batch_size)
if isinstance(x, list) and len(x) == 2: # deep supervision
if pred.shape[-1] == 1: # regression
pred_flatten = pred.flatten()
else: # classification
pred_flatten = pred.reshape((-1, 10))
for m, t, p in zip(x[1].flatten(), y.flatten(), pred_flatten):
if np.equal(m, 1):
y_true.append(t)
predictions.append(p)
else:
if pred.shape[-1] == 1:
y_true += list(y.flatten())
predictions += list(pred.flatten())
else:
y_true += list(y)
predictions += list(pred)
print('\n')
if self.partition == 'log':
predictions = [
metrics.get_estimate_log(x, 10) for x in predictions
]
ret = metrics.print_metrics_log_bins(y_true, predictions)
if self.partition == 'custom':
predictions = [
metrics.get_estimate_custom(x, 10) for x in predictions
]
ret = metrics.print_metrics_custom_bins(y_true, predictions)
if self.partition == 'none':
ret = metrics.print_metrics_regression(y_true, predictions)
for k, v in ret.items():
logs[dataset + '_' + k] = v
history.append(ret)
pred = model.predict_on_batch(x)
predictions += list(pred)
labels += list(y)
names += list(cur_names)
ts += list(cur_ts)
if stochastic:
aleatoric = [np.mean(x * (1. - x), axis=0) for x in predictions]
epistemic = [np.var(x, axis=0) for x in predictions]
predictions = [np.mean(x, axis=0) for x in predictions]
if args.partition == 'log':
predictions = [metrics.get_estimate_log(x, 10) for x in predictions]
metrics.print_metrics_log_bins(labels, predictions)
if args.partition == 'custom':
predictions = [metrics.get_estimate_custom(x, 10) for x in predictions]
metrics.print_metrics_custom_bins(labels, predictions)
if args.partition == 'none':
metrics.print_metrics_regression(labels, predictions)
predictions = [x[0] for x in predictions]
path = os.path.join(
os.path.join(args.output_dir, "test_predictions",
os.path.basename(args.load_state)) + ".csv")
if stochastic:
ee = np.mean(np.array(epistemic))
aa = np.mean(np.array(aleatoric))
print("Epistemic uncertainty =", ee)
print("Aleatoric uncertainty =", aa)
print("Uncertainty =", ee + aa)
## decomp
if args.decomp_C > 0:
print "\n ================ decompensation ================"
decomp_pred = np.array(decomp_pred)
decomp_pred = np.stack([1-decomp_pred, decomp_pred], axis=1)
decomp_ret = metrics.print_metrics_binary(decomp_y_true, decomp_pred)
## los
if args.los_C > 0:
print "\n ================ length of stay ================"
if args.partition == 'log':
los_pred = [metrics.get_estimate_log(x, 10) for x in los_pred]
los_ret = metrics.print_metrics_log_bins(los_y_true, los_pred)
if args.partition == 'custom':
los_pred = [metrics.get_estimate_custom(x, 10) for x in los_pred]
los_ret = metrics.print_metrics_custom_bins(los_y_true, los_pred)
if args.partition == 'none':
los_ret = metrics.print_metrics_regression(los_y_true, los_pred)
## pheno
if args.pheno_C > 0:
print "\n =================== phenotype =================="
pheno_pred = np.array(pheno_pred)
pheno_ret = metrics.print_metrics_multilabel(pheno_y_true, pheno_pred)
# TODO: save activations if needed
elif args.mode == 'test_single':
# ensure that the code uses test_reader
del train_reader
resfile.write(",".join(["%.6f" % x for x in ret['auc_scores']]) + "\n")
print "\nAverage results on train"
ret = metrics.print_metrics_multilabel(train_y, train_activations)
write_results(resfile, ret)
print "\nAverage results on val"
ret = metrics.print_metrics_multilabel(val_y, val_activations)
write_results(resfile, ret)
print "\nAverage results on test"
ret = metrics.print_metrics_multilabel(test_y, test_activations)
write_results(resfile, ret)
train_predictions = np.array([metrics.get_estimate_custom(x, nbins) for x in train_activations])
val_predictions = np.array([metrics.get_estimate_custom(x, nbins) for x in val_activations])
test_predictions = np.array([metrics.get_estimate_custom(x, nbins) for x in test_activations])
with open(os.path.join("cf_activations", model_name + ".txt"), "w") as actfile:
for (x, y) in zip(test_predictions, test_actual):
actfile.write("%.6f %.6f\n" % (x, y))
with open(os.path.join("cf_results", model_name + ".txt"), "w") as resfile:
resfile.write("mad, mse, mape, kappa\n")
print "Scores on train set"
ret = metrics.print_metrics_custom_bins(train_actual, train_predictions)
resfile.write("%.6f,%.6f,%.6f,%.6f\n" % (
ret['mad'],
ret['mse'],
def calc_metrics(self, data_gen, history, dataset, logs):
ihm_y_true = []
decomp_y_true = []
los_y_true = []
pheno_y_true = []
ihm_pred = []
decomp_pred = []
los_pred = []
pheno_pred = []
for i in range(data_gen.steps):
if self.verbose == 1:
print("\tdone {}/{}".format(i, data_gen.steps), end='\r')
(X, y, los_y_reg) = data_gen.next(return_y_true=True)
outputs = self.model.predict(X, batch_size=self.batch_size)
ihm_M = X[1]
decomp_M = X[2]
los_M = X[3]
if not data_gen.target_repl: # no target replication
(ihm_p, decomp_p, los_p, pheno_p) = outputs
(ihm_t, decomp_t, los_t, pheno_t) = y
else: # target replication
(ihm_p, _, decomp_p, los_p, pheno_p, _) = outputs
(ihm_t, _, decomp_t, los_t, pheno_t, _) = y
los_t = los_y_reg # real value not the label
# ihm
for (m, t, p) in zip(ihm_M.flatten(), ihm_t.flatten(),
ihm_p.flatten()):
if np.equal(m, 1):
ihm_y_true.append(t)
ihm_pred.append(p)
# decomp
for (m, t, p) in zip(decomp_M.flatten(), decomp_t.flatten(),
decomp_p.flatten()):
if np.equal(m, 1):
decomp_y_true.append(t)
decomp_pred.append(p)
# los
if los_p.shape[-1] == 1: # regression
for (m, t, p) in zip(los_M.flatten(), los_t.flatten(),
los_p.flatten()):
if np.equal(m, 1):
los_y_true.append(t)
los_pred.append(p)
else: # classification
for (m, t, p) in zip(los_M.flatten(), los_t.flatten(),
los_p.reshape((-1, 10))):
if np.equal(m, 1):
los_y_true.append(t)
los_pred.append(p)
# pheno
for (t, p) in zip(pheno_t.reshape((-1, 25)),
pheno_p.reshape((-1, 25))):
pheno_y_true.append(t)
pheno_pred.append(p)
print('\n')
# ihm
print("\n ================= 48h mortality ================")
ihm_pred = np.array(ihm_pred)
ihm_pred = np.stack([1 - ihm_pred, ihm_pred], axis=1)
ret = metrics.print_metrics_binary(ihm_y_true, ihm_pred)
for k, v in ret.items():
logs[dataset + '_ihm_' + k] = v
# decomp
print("\n ================ decompensation ================")
decomp_pred = np.array(decomp_pred)
decomp_pred = np.stack([1 - decomp_pred, decomp_pred], axis=1)
ret = metrics.print_metrics_binary(decomp_y_true, decomp_pred)
for k, v in ret.items():
logs[dataset + '_decomp_' + k] = v
# los
print("\n ================ length of stay ================")
if self.partition == 'log':
los_pred = [metrics.get_estimate_log(x, 10) for x in los_pred]
ret = metrics.print_metrics_log_bins(los_y_true, los_pred)
if self.partition == 'custom':
los_pred = [metrics.get_estimate_custom(x, 10) for x in los_pred]
ret = metrics.print_metrics_custom_bins(los_y_true, los_pred)
if self.partition == 'none':
ret = metrics.print_metrics_regression(los_y_true, los_pred)
for k, v in ret.items():
logs[dataset + '_los_' + k] = v
# pheno
print("\n =================== phenotype ==================")
pheno_pred = np.array(pheno_pred)
ret = metrics.print_metrics_multilabel(pheno_y_true, pheno_pred)
for k, v in ret.items():
logs[dataset + '_pheno_' + k] = v
history.append(logs)
ihm_ret = metrics.print_metrics_binary(ihm_y_true, ihm_pred)
# decomp
if args.decomp_C > 0:
print "\n ================ decompensation ================"
decomp_pred = np.array(decomp_pred)
decomp_ret = metrics.print_metrics_binary(decomp_y_true, decomp_pred)
# los
if args.los_C > 0:
print "\n ================ length of stay ================"
if args.partition == 'log':
los_pred = [metrics.get_estimate_log(x, 10) for x in los_pred]
los_ret = metrics.print_metrics_log_bins(los_y_true, los_pred)
if args.partition == 'custom':
los_pred = [metrics.get_estimate_custom(x, 10) for x in los_pred]
los_ret = metrics.print_metrics_custom_bins(los_y_true, los_pred)
if args.partition == 'none':
los_ret = metrics.print_metrics_regression(los_y_true, los_pred)
# pheno
if args.pheno_C > 0:
print "\n =================== phenotype =================="
pheno_pred = np.array(pheno_pred)
pheno_ret = metrics.print_metrics_multilabel(pheno_y_true, pheno_pred)
print "Saving the predictions in test_predictions/task directories ..."
# ihm
ihm_path = os.path.join("test_predictions/ihm", os.path.basename(args.load_state)) + ".csv"
ihm_utils.save_results(ihm_names, ihm_pred, ihm_y_true, ihm_path)
for x in ret['auc_scores']]) + "\n")
print "\nAverage results on train"
ret = metrics.print_metrics_multilabel(train_y, train_activations)
write_results(resfile, ret)
print "\nAverage results on val"
ret = metrics.print_metrics_multilabel(val_y, val_activations)
write_results(resfile, ret)
print "\nAverage results on test"
ret = metrics.print_metrics_multilabel(test_y, test_activations)
write_results(resfile, ret)
train_predictions = np.array(
[metrics.get_estimate_custom(x, nbins) for x in train_activations])
val_predictions = np.array(
[metrics.get_estimate_custom(x, nbins) for x in val_activations])
test_predictions = np.array(
[metrics.get_estimate_custom(x, nbins) for x in test_activations])
with open(os.path.join("cf_activations", model_name + ".txt"),
"w") as actfile:
for (x, y) in zip(test_predictions, test_actual):
actfile.write("%.6f %.6f\n" % (x, y))
with open(os.path.join("cf_results", model_name + ".txt"), "w") as resfile:
resfile.write("mad, mse, mape, kappa\n")
print "Scores on train set"
ret = metrics.print_metrics_custom_bins(train_actual,
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--period', type=str, default='all', help='specifies which period extract features from',
choices=['first4days', 'first8days', 'last12hours', 'first25percent', 'first50percent', 'all'])
parser.add_argument('--features', type=str, default='all', help='specifies what features to extract',
choices=['all', 'len', 'all_but_len'])
parser.add_argument('--grid-search', dest='grid_search', action='store_true')
parser.add_argument('--no-grid-search', dest='grid_search', action='store_false')
parser.set_defaults(grid_search=False)
parser.add_argument('--data', type=str, help='Path to the data of length-of-stay task',
default=os.path.join(os.path.dirname(__file__), '../../../data/length-of-stay/'))
parser.add_argument('--output_dir', type=str, help='Directory relative which all output files are stored',
default='.')
args = parser.parse_args()
print(args)
if args.grid_search:
penalties = ['l2', 'l2', 'l2', 'l2', 'l2', 'l2', 'l1', 'l1', 'l1', 'l1', 'l1']
coefs = [1.0, 0.1, 0.01, 0.001, 0.0001, 0.00001, 1.0, 0.1, 0.01, 0.001, 0.0001]
else:
penalties = ['l2']
coefs = [0.00001]
train_reader = LengthOfStayReader(dataset_dir=os.path.join(args.data, 'train'),
listfile=os.path.join(args.data, 'train_listfile.csv'))
val_reader = LengthOfStayReader(dataset_dir=os.path.join(args.data, 'train'),
listfile=os.path.join(args.data, 'val_listfile.csv'))
test_reader = LengthOfStayReader(dataset_dir=os.path.join(args.data, 'test'),
listfile=os.path.join(args.data, 'test_listfile.csv'))
print('Reading data and extracting features ...')
n_train = min(100000, train_reader.get_number_of_examples())
n_val = min(100000, val_reader.get_number_of_examples())
(train_X, train_y, train_actual, train_names, train_ts) = read_and_extract_features(
train_reader, n_train, args.period, args.features)
(val_X, val_y, val_actual, val_names, val_ts) = read_and_extract_features(
val_reader, n_val, args.period, args.features)
(test_X, test_y, test_actual, test_names, test_ts) = read_and_extract_features(
test_reader, test_reader.get_number_of_examples(), args.period, args.features)
print("train set shape: {}".format(train_X.shape))
print("validation set shape: {}".format(val_X.shape))
print("test set shape: {}".format(test_X.shape))
print('Imputing missing values ...')
imputer = Imputer(missing_values=np.nan, strategy='mean', axis=0, verbose=0, copy=True)
imputer.fit(train_X)
train_X = np.array(imputer.transform(train_X), dtype=np.float32)
val_X = np.array(imputer.transform(val_X), dtype=np.float32)
test_X = np.array(imputer.transform(test_X), dtype=np.float32)
print('Normalizing the data to have zero mean and unit variance ...')
scaler = StandardScaler()
scaler.fit(train_X)
train_X = scaler.transform(train_X)
val_X = scaler.transform(val_X)
test_X = scaler.transform(test_X)
result_dir = os.path.join(args.output_dir, 'cf_results')
common_utils.create_directory(result_dir)
for (penalty, C) in zip(penalties, coefs):
model_name = '{}.{}.{}.C{}'.format(args.period, args.features, penalty, C)
train_activations = np.zeros(shape=train_y.shape, dtype=float)
val_activations = np.zeros(shape=val_y.shape, dtype=float)
test_activations = np.zeros(shape=test_y.shape, dtype=float)
for task_id in range(n_bins):
logreg = LogisticRegression(penalty=penalty, C=C, random_state=42)
logreg.fit(train_X, train_y[:, task_id])
train_preds = logreg.predict_proba(train_X)
train_activations[:, task_id] = train_preds[:, 1]
val_preds = logreg.predict_proba(val_X)
val_activations[:, task_id] = val_preds[:, 1]
test_preds = logreg.predict_proba(test_X)
test_activations[:, task_id] = test_preds[:, 1]
train_predictions = np.array([metrics.get_estimate_custom(x, n_bins) for x in train_activations])
val_predictions = np.array([metrics.get_estimate_custom(x, n_bins) for x in val_activations])
test_predictions = np.array([metrics.get_estimate_custom(x, n_bins) for x in test_activations])
with open(os.path.join(result_dir, 'train_{}.json'.format(model_name)), 'w') as f:
ret = metrics.print_metrics_custom_bins(train_actual, train_predictions)
ret = {k: float(v) for k, v in ret.items()}
json.dump(ret, f)
with open(os.path.join(result_dir, 'val_{}.json'.format(model_name)), 'w') as f:
ret = metrics.print_metrics_custom_bins(val_actual, val_predictions)
ret = {k: float(v) for k, v in ret.items()}
json.dump(ret, f)
with open(os.path.join(result_dir, 'test_{}.json'.format(model_name)), 'w') as f:
ret = metrics.print_metrics_custom_bins(test_actual, test_predictions)
ret = {k: float(v) for k, v in ret.items()}
json.dump(ret, f)
save_results(test_names, test_ts, test_predictions, test_actual,
os.path.join(args.output_dir, 'cf_predictions', model_name + '.csv'))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--period', type=str, default='all', help='specifies which period extract features from',
choices=['first4days', 'first8days', 'last12hours', 'first25percent', 'first50percent', 'all'])
parser.add_argument('--features', type=str, default='all', help='specifies what features to extract',
choices=['all', 'len', 'all_but_len'])
args = parser.parse_args()
print(args)
# penalties = ['l2', 'l2', 'l2', 'l2', 'l2', 'l2', 'l1', 'l1', 'l1', 'l1', 'l1']
# Cs = [1.0, 0.1, 0.01, 0.001, 0.0001, 0.00001, 1.0, 0.1, 0.01, 0.001, 0.0001]
penalties = ['l2']
Cs = [0.00001]
train_reader = LengthOfStayReader(dataset_dir='../../../data/length-of-stay/train/',
listfile='../../../data/length-of-stay/train_listfile.csv')
val_reader = LengthOfStayReader(dataset_dir='../../../data/length-of-stay/train/',
listfile='../../../data/length-of-stay/val_listfile.csv')
test_reader = LengthOfStayReader(dataset_dir='../../../data/length-of-stay/test/',
listfile='../../../data/length-of-stay/test_listfile.csv')
print('Reading data and extracting features ...')
n_train = min(100000, train_reader.get_number_of_examples())
n_val = min(100000, val_reader.get_number_of_examples())
(train_X, train_y, train_actual, train_names, train_ts) = read_and_extract_features(
train_reader, n_train, args.period, args.features)
(val_X, val_y, val_actual, val_names, val_ts) = read_and_extract_features(
val_reader, n_val, args.period, args.features)
(test_X, test_y, test_actual, test_names, test_ts) = read_and_extract_features(
test_reader, test_reader.get_number_of_examples(), args.period, args.features)
print("train set shape: {}".format(train_X.shape))
print("validation set shape: {}".format(val_X.shape))
print("test set shape: {}".format(test_X.shape))
print('Imputing missing values ...')
imputer = Imputer(missing_values=np.nan, strategy='mean', axis=0, verbose=0, copy=True)
imputer.fit(train_X)
train_X = np.array(imputer.transform(train_X), dtype=np.float32)
val_X = np.array(imputer.transform(val_X), dtype=np.float32)
test_X = np.array(imputer.transform(test_X), dtype=np.float32)
print('Normalizing the data to have zero mean and unit variance ...')
scaler = StandardScaler()
scaler.fit(train_X)
train_X = scaler.transform(train_X)
val_X = scaler.transform(val_X)
test_X = scaler.transform(test_X)
common_utils.create_directory('cf_results')
for (penalty, C) in zip(penalties, Cs):
model_name = '{}.{}.{}.C{}'.format(args.period, args.features, penalty, C)
train_activations = np.zeros(shape=train_y.shape, dtype=float)
val_activations = np.zeros(shape=val_y.shape, dtype=float)
test_activations = np.zeros(shape=test_y.shape, dtype=float)
for task_id in range(n_bins):
logreg = LogisticRegression(penalty=penalty, C=C, random_state=42)
logreg.fit(train_X, train_y[:, task_id])
train_preds = logreg.predict_proba(train_X)
train_activations[:, task_id] = train_preds[:, 1]
val_preds = logreg.predict_proba(val_X)
val_activations[:, task_id] = val_preds[:, 1]
test_preds = logreg.predict_proba(test_X)
test_activations[:, task_id] = test_preds[:, 1]
train_predictions = np.array([metrics.get_estimate_custom(x, n_bins) for x in train_activations])
val_predictions = np.array([metrics.get_estimate_custom(x, n_bins) for x in val_activations])
test_predictions = np.array([metrics.get_estimate_custom(x, n_bins) for x in test_activations])
with open(os.path.join('cf_results', 'train_{}.json'.format(model_name)), 'w') as f:
ret = metrics.print_metrics_custom_bins(train_actual, train_predictions)
ret = {k: float(v) for k, v in ret.items()}
json.dump(ret, f)
with open(os.path.join('cf_results', 'val_{}.json'.format(model_name)), 'w') as f:
ret = metrics.print_metrics_custom_bins(val_actual, val_predictions)
ret = {k: float(v) for k, v in ret.items()}
json.dump(ret, f)
with open(os.path.join('cf_results', 'test_{}.json'.format(model_name)), 'w') as f:
ret = metrics.print_metrics_custom_bins(test_actual, test_predictions)
ret = {k: float(v) for k, v in ret.items()}
json.dump(ret, f)
save_results(test_names, test_ts, test_predictions, test_actual,
os.path.join('cf_predictions', model_name + '.csv'))
for i in range(test_data_gen.steps):
print "\rpredicting {} / {}".format(i, test_data_gen.steps),
ret = test_data_gen.next(return_y_true=True)
(x, y_processed, y) = ret["data"]
cur_names = ret["names"]
cur_ts = ret["ts"]
x = np.array(x)
pred = model.predict_on_batch(x)
predictions += list(pred)
labels += list(y)
names += list(cur_names)
ts += list(cur_ts)
if args.partition == 'log':
predictions = [metrics.get_estimate_log(x, 10) for x in predictions]
metrics.print_metrics_log_bins(labels, predictions)
if args.partition == 'custom':
predictions = [metrics.get_estimate_custom(x, 10) for x in predictions]
metrics.print_metrics_custom_bins(labels, predictions)
if args.partition == 'none':
metrics.print_metrics_regression(labels, predictions)
predictions = [x[0] for x in predictions]
path = os.path.join("test_predictions", os.path.basename(args.load_state)) + ".csv"
utils.save_results(names, ts, predictions, labels, path)
else:
raise ValueError("Wrong value for args.mode")
