def test_time_series(self):
import numpy as np
ob = rnnmed.data.observations.Observations()
data = [[["A"], ["B"]], [["C"], ["D"]]]
for item in data:
ob.add(item, 1)
o_gen = observations.time_observation_generator(ob, n_visits=2)
a_gen = observations.time_observation_generator(ob, n_visits=2)
b_gen = observations.time_observation_generator(ob, n_visits=2)
timeseries = rnnmed.data.io.read_time_series(
open("test_data/synthetic_control.txt"))
c_gen = rnnmed.data.concatenate_generator(
[a_gen, b_gen], concat=lambda x: np.concatenate(x, axis=2))
for x, y in c_gen:
print(x)
print(y)
print(timeseries[0])
generator = ts.timeseries_generator(timeseries)
x_a, y_a = rnnmed.data.generate_time_batch(generator, batch_size=2)
x_b, _ = rnnmed.data.generate_time_batch(o_gen, batch_size=2)
print(x_a)
def test_time_series_observation(self):
def week_agg(date):
return date.year, date.isocalendar()[1]
def month_agg(date):
return date.year, date.month
ob = rnnmed.data.io.read_time_series_observation(
open("/mnt/veracrypt1/EHR_DATA/L270-90-raw-measurements.csv"),
min_sparsity=0.1)
import random
random.seed(10)
random.shuffle(ob)
n_visits = 10
generator = observations.time_observation_generator(ob,
n_visits=n_visits)
print(len(ob), ob.n_features)
from rnnmed.visit2visit import visit2visit
visit2visit(generator,
n_features=ob.n_features,
n_labels=ob.n_labels,
n_timesteps=n_visits,
n_hidden=128,
max_iter=1000)
def test_time_series_observation(self):
ob = rnnmed.data.io.read_time_series_observation(
open("/home/isak/D611-90-raw-measurements.csv"), min_sparsity=0.4)
import random
random.seed(10)
random.shuffle(ob)
n_labels = ob.n_labels
n_features = ob.n_features
n_timesteps = 10
generator = observations.time_observation_generator(
ob, n_visits=n_timesteps)
x_data, y_data = rnnmed.data.make_time_input_output_arrays(generator)
print(x_data.shape)
print(y_data.shape)
aucs = []
skf = StratifiedKFold(n_splits=10, shuffle=True)
for fold, (train, test) in enumerate(
skf.split(np.zeros(x_data.shape[1]), y_data)):
x_train = x_data[:, train, :]
y_train = y_data[train, :]
x_test = x_data[:, test, :]
y_test = y_data[test, :]
graph = tf.Graph()
with graph.as_default():
X = tf.placeholder(tf.float32,
shape=[n_timesteps, None, n_features])
y = tf.placeholder(tf.int32, shape=[None])
drop_prob = tf.placeholder_with_default(1.0, shape=())
hp = HistoryPredictor(X, tf.one_hot(y, depth=n_labels),
drop_prob)
init = tf.global_variables_initializer()
with tf.Session(graph=graph) as sess:
sess.run(init)
for epoch in range(1000):
for idx in batch_generator(x_train.shape[1], 32):
_, _loss = sess.run(
[hp.optimize, hp.loss],
feed_dict={
X: x_train[:, idx, :],
y: y_train[idx, :].reshape(-1),
drop_prob: 0.4
})
if epoch % 25 == 0:
print("Fold {}, epoch {} loss: {}".format(
fold, epoch, _loss))
y_pred = sess.run(hp.prediction, feed_dict={X: x_test})
auc = roc_auc_score(y_test.reshape(-1), y_pred[:, 1])
print(auc)
aucs.append(auc)
print("mean auc:", np.mean(aucs))
def test_med_2_vec_predict(self):
observations = rnnmed.data.io.read_labeled_observations(
"test_data/mimic_demo.seq")
generator = ob.time_observation_generator(observations, n_visits=15)
np.set_printoptions(suppress=True)
visit2visit.visit2visit(generator,
n_labels=observations.n_labels,
n_features=observations.n_features,
n_timesteps=15,
transform=None)