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_log(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 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)
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 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))
ihm_pred = np.array(ihm_pred)
ihm_pred = np.stack([1-ihm_pred, ihm_pred], axis=1)
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_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
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)
if args.ihm_C > 0:
print "\n ================= 48h mortality ================"
ihm_pred = np.array(ihm_pred)
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 ..."
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_log(x, self.nbins) for x in ret[0]]
return [predictions] + list(ret[1:])
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")
