def _load_data(self, testfold=4):
train_reader = InHospitalMortalityReader(
dataset_dir='mimic3-benchmarks/data/in-hospital-mortality/train/',
listfile=
'mimic3-benchmarks/data/in-hospital-mortality/train_listfile.csv',
period_length=48.0)
val_reader = InHospitalMortalityReader(
dataset_dir='mimic3-benchmarks/data/in-hospital-mortality/train/',
listfile=
'mimic3-benchmarks/data/in-hospital-mortality/val_listfile.csv',
period_length=48.0)
test_reader = InHospitalMortalityReader(
dataset_dir='mimic3-benchmarks/data/in-hospital-mortality/test/',
listfile=
'mimic3-benchmarks/data/in-hospital-mortality/test_listfile.csv',
period_length=48.0)
discretizer = Discretizer(timestep=float(4),
store_masks=True,
imput_strategy='previous',
start_time='zero')
discretizer_header = discretizer.transform(
train_reader.read_example(0)[0])[1].split(',')
cont_channels = [
i for (i, x) in enumerate(discretizer_header) if x.find("->") == -1
]
normalizer = Normalizer(
fields=cont_channels) # choose here onlycont vs all
normalizer.load_params(
'mimic3-benchmarks/mimic3models/in_hospital_mortality/'
'ihm_ts%s.input_str:%s.start_time:zero.normalizer' %
('2.0', 'previous'))
# normalizer=None
train_raw = utils.load_data(train_reader, discretizer, normalizer,
False)
val_raw = utils.load_data(val_reader, discretizer, normalizer, False)
test_raw = utils.load_data(test_reader, discretizer, normalizer, False)
# To split into
def preprocess(the_raw_set):
x, y = the_raw_set
x = x.astype(np.float32, copy=False)
y = np.array(y)
return x, y
train_raw = preprocess(train_raw)
val_raw = preprocess(val_raw)
test_raw = preprocess(test_raw)
return train_raw, val_raw, test_raw
def dataset_reader(phase, args, target_repl=False):
if phase == "train":
#% Build readers & discretizers
train_reader = InHospitalMortalityReader(
dataset_dir=os.path.join(args.data, 'train'),
listfile=os.path.join(args.data, 'train_listfile.csv'),
period_length=48.0)
val_reader = InHospitalMortalityReader(
dataset_dir=os.path.join(args.data, 'train'),
listfile=os.path.join(args.data, 'val_listfile.csv'),
period_length=48.0)
discretizer = Discretizer(timestep=float(args.timestep),
store_masks=True,
impute_strategy='previous',
start_time='zero')
discretizer_header = discretizer.transform(
train_reader.read_example(0)["X"])[1].split(',')
cont_channels = [
i for (i, x) in enumerate(discretizer_header) if x.find("->") == -1
]
#%% Data normalization (by mean and variance)
normalizer = Normalizer(
fields=cont_channels) # choose here which columns to standardize
normalizer_state = args.normalizer_state
if normalizer_state is None:
normalizer_state = 'ihm_ts{}.input_str:{}.start_time:zero.normalizer'.format(
args.timestep, args.imputation)
normalizer_state = os.path.join(os.path.dirname(__file__),
normalizer_state)
normalizer.load_params(normalizer_state)
# args_dict = dict(args._get_kwargs()) #TODO: reverse
args_dict = {}
args_dict['header'] = discretizer_header
args_dict['task'] = 'ihm'
args_dict['target_repl'] = target_repl
#%% Read data
start = time()
print("Reading started")
train_raw = utils.load_data(train_reader,
discretizer,
normalizer,
args.small_part,
return_names=False)
val_raw = utils.load_data(val_reader,
discretizer,
normalizer,
args.small_part,
return_names=False)
if target_repl:
T = train_raw[0][0].shape[0]
def extend_labels(data):
data = list(data)
labels = np.array(data[1]) # (B,)
data[1] = [labels, None]
data[1][1] = np.expand_dims(labels,
axis=-1).repeat(T,
axis=1) # (B, T)
data[1][1] = np.expand_dims(data[1][1], axis=-1) # (B, T, 1)
return data
train_raw = extend_labels(train_raw)
val_raw = extend_labels(val_raw)
print("Reading finished after {} seconds".format(time() - start))
return (train_raw, val_raw)
else: ################################### TEST phase
test_reader = InHospitalMortalityReader(
dataset_dir=os.path.join(args.data, 'test'),
listfile=os.path.join(args.data, 'test_listfile.csv'),
period_length=48.0)
test_raw = utils.load_data(test_reader,
discretizer,
normalizer,
args.small_part,
return_names=True)
return test_raw
loss=loss,
loss_weights=loss_weights,
metrics=['accuracy'])
model.summary()
# Load model weights
n_trained_chunks = 0
if args.load_state != "":
model.load_weights(args.load_state)
n_trained_chunks = int(
re.match(".*epoch([0-9]+).*", args.load_state).group(1))
# Read data
print("======>Loading train")
#{"data": whole_data(2D array,y_lable), "names": names}
train_raw = utils.load_data(train_reader, discretizer, normalizer,
args.small_part)
print("======>Loading val")
val_raw = utils.load_data(val_reader, discretizer, normalizer, args.small_part)
if target_repl:
T = train_raw[0][0].shape[0] # number of columns
def extend_labels(data):
data = list(data)
labels = np.array(data[1]) # (B,)
data[1] = [labels, None]
data[1][1] = np.expand_dims(labels, axis=-1).repeat(T,
axis=1) # (B, T)
data[1][1] = np.expand_dims(data[1][1], axis=-1) # (B, T, 1)
return data
def mimic_loader(task='mortality', data_percentage=100):
if task == 'mortality':
print('loading mimic-iii in-hospital mortality dataset')
from mimic3models.in_hospital_mortality import utils
from mimic3benchmark.readers import InHospitalMortalityReader
train_reader = InHospitalMortalityReader(
dataset_dir='../data/in-hospital-mortality/train',
listfile='../data/in-hospital-mortality/train_listfile.csv',
period_length=48.0)
val_reader = InHospitalMortalityReader(
dataset_dir='../data/in-hospital-mortality/train',
listfile='../data/in-hospital-mortality/val_listfile.csv',
period_length=48.0)
test_reader = InHospitalMortalityReader(
dataset_dir='../data/in-hospital-mortality/test',
listfile='../data/in-hospital-mortality/test_listfile.csv',
period_length=48.0)
discretizer = Discretizer(timestep=float(1.0),
store_masks=True,
impute_strategy='previous',
start_time='zero')
discretizer_header = discretizer.transform(
train_reader.read_example(0)["X"])[1].split(',')
cont_channels = [
i for (i, x) in enumerate(discretizer_header) if x.find("->") == -1
]
normalizer = Normalizer(
fields=cont_channels) # choose here which columns to standardize
normalizer_state = None
if normalizer_state is None:
normalizer_state = 'ihm_ts{}.input_str:{}.start_time:zero.normalizer'.format(
1.0, 'previous')
normalizer_state = os.path.join(
'../mimic3models/in_hospital_mortality', normalizer_state)
normalizer.load_params(normalizer_state)
headers = [
'Capillary refill rate->0.0', 'Capillary refill rate->1.0',
'Diastolic blood pressure', 'Fraction inspired oxygen',
'Glascow coma scale eye opening->To Pain',
'Glascow coma scale eye opening->3 To speech',
'Glascow coma scale eye opening->1 No Response',
'Glascow coma scale eye opening->4 Spontaneously',
'Glascow coma scale eye opening->None',
'Glascow coma scale eye opening->To Speech',
'Glascow coma scale eye opening->Spontaneously',
'Glascow coma scale eye opening->2 To pain',
'Glascow coma scale motor response->1 No Response',
'Glascow coma scale motor response->3 Abnorm flexion',
'Glascow coma scale motor response->Abnormal extension',
'Glascow coma scale motor response->No response',
'Glascow coma scale motor response->4 Flex-withdraws',
'Glascow coma scale motor response->Localizes Pain',
'Glascow coma scale motor response->Flex-withdraws',
'Glascow coma scale motor response->Obeys Commands',
'Glascow coma scale motor response->Abnormal Flexion',
'Glascow coma scale motor response->6 Obeys Commands',
'Glascow coma scale motor response->5 Localizes Pain',
'Glascow coma scale motor response->2 Abnorm extensn',
'Glascow coma scale total->11', 'Glascow coma scale total->10',
'Glascow coma scale total->13', 'Glascow coma scale total->12',
'Glascow coma scale total->15', 'Glascow coma scale total->14',
'Glascow coma scale total->3', 'Glascow coma scale total->5',
'Glascow coma scale total->4', 'Glascow coma scale total->7',
'Glascow coma scale total->6', 'Glascow coma scale total->9',
'Glascow coma scale total->8',
'Glascow coma scale verbal response->1 No Response',
'Glascow coma scale verbal response->No Response',
'Glascow coma scale verbal response->Confused',
'Glascow coma scale verbal response->Inappropriate Words',
'Glascow coma scale verbal response->Oriented',
'Glascow coma scale verbal response->No Response-ETT',
'Glascow coma scale verbal response->5 Oriented',
'Glascow coma scale verbal response->Incomprehensible sounds',
'Glascow coma scale verbal response->1.0 ET/Trach',
'Glascow coma scale verbal response->4 Confused',
'Glascow coma scale verbal response->2 Incomp sounds',
'Glascow coma scale verbal response->3 Inapprop words', 'Glucose',
'Heart Rate', 'Height', 'Mean blood pressure', 'Oxygen saturation',
'Respiratory rate', 'Systolic blood pressure', 'Temperature',
'Weight', 'pH', 'mask->Capillary refill rate',
'mask->Diastolic blood pressure', 'mask->Fraction inspired oxygen',
'mask->Glascow coma scale eye opening',
'mask->Glascow coma scale motor response',
'mask->Glascow coma scale total',
'mask->Glascow coma scale verbal response', 'mask->Glucose',
'mask->Heart Rate', 'mask->Height', 'mask->Mean blood pressure',
'mask->Oxygen saturation', 'mask->Respiratory rate',
'mask->Systolic blood pressure', 'mask->Temperature',
'mask->Weight', 'mask->pH'
]
print('start loading the data')
if data_percentage != 100: # accepted values: [10,20,30,40,50,60,70,80,90]
print('loading the partially covered testing data')
test_reader = InHospitalMortalityReader(
dataset_dir='../data/in-hospital-mortality/test_' +
str(data_percentage),
listfile='../data/in-hospital-mortality/test_listfile.csv',
period_length=48.0)
test_raw = utils.load_data(test_reader, discretizer, normalizer,
False)
x_test = np.copy(test_raw[0])
return x_test
# Read data
train_raw = utils.load_data(train_reader, discretizer, normalizer,
False)
val_raw = utils.load_data(val_reader, discretizer, normalizer, False)
test_raw = utils.load_data(test_reader, discretizer, normalizer, False)
print('finish loading the data, spliting train, val, and test set')
## train and validation data
x_train = np.copy(train_raw[0])
y_train = np.zeros((len(train_raw[1]), 2))
y_train[:, 1] = np.array(train_raw[1])
y_train[:, 0] = 1 - y_train[:, 1]
x_val = np.copy(val_raw[0])
y_val = np.zeros((len(val_raw[1]), 2))
y_val[:, 1] = np.array(val_raw[1])
y_val[:, 0] = 1 - y_val[:, 1]
x_test = np.copy(test_raw[0])
y_test = np.zeros((len(test_raw[1]), 2))
y_test[:, 1] = np.array(test_raw[1])
y_test[:, 0] = 1 - y_test[:, 1]
return [x_train, x_val, x_test, y_train, y_val, y_test]
def main():
parser = argparse.ArgumentParser()
common_utils.add_common_arguments_backdoor(parser)
parser.add_argument('--target_repl_coef', type=float, default=0.0)
parser.add_argument('--data', type=str, help='Path to the data of in-hospital mortality task',
default=os.path.join(os.path.dirname(__file__), '../../../data/in-hospital-mortality/'))
parser.add_argument('--output_dir', type=str, help='Directory relative which all output files are stored',
default='.')
parser.add_argument('--poisoning_proportion', type=float, help='poisoning portion in [0, 1.0]',
required=True)
parser.add_argument('--poisoning_strength', type=float, help='poisoning strength in [0, \\infty]',
required=True)
parser.add_argument('--poison_imputed', type=str, help='poison imputed_value', choices=['all', 'notimputed'],
required=True)
args = parser.parse_args()
print(args)
if args.small_part:
args.save_every = 2**30
target_repl = (args.target_repl_coef > 0.0 and args.mode == 'train')
# Build readers, discretizers, normalizers
train_reader = InHospitalMortalityReader(dataset_dir=os.path.join(args.data, 'train'),
listfile=os.path.join(args.data, 'train_listfile.csv'),
period_length=48.0)
val_reader = InHospitalMortalityReader(dataset_dir=os.path.join(args.data, 'train'),
listfile=os.path.join(args.data, 'val_listfile.csv'),
period_length=48.0)
poisoning_trigger = np.reshape(np.load("./cache/in_hospital_mortality/torch_raw_48_17/poison_pattern.npy"), (-1, 48, 17))
discretizer = PoisoningDiscretizer(timestep=float(args.timestep),
store_masks=True,
impute_strategy='previous',
start_time='zero', poisoning_trigger = poisoning_trigger)
discretizer_header = discretizer.transform(train_reader.read_example(0)["X"])[1].split(',')
cont_channels = [i for (i, x) in enumerate(discretizer_header) if x.find("->") == -1]
normalizer = Normalizer(fields=cont_channels) # choose here which columns to standardize
normalizer_state = args.normalizer_state
if normalizer_state is None:
normalizer_state = '../ihm_ts{}.input_str:{}.start_time:zero.normalizer'.format(args.timestep, args.imputation)
normalizer_state = os.path.join(os.path.dirname(__file__), normalizer_state)
normalizer.load_params(normalizer_state)
args_dict = dict(args._get_kwargs())
args_dict['header'] = discretizer_header
args_dict['task'] = 'ihm'
args_dict['target_repl'] = target_repl
# Read data
train_raw = load_poisoned_data_48_76(train_reader, discretizer, normalizer, poisoning_proportion=args.poisoning_proportion, poisoning_strength=args.poisoning_strength, suffix="train", small_part=args.small_part, poison_imputed={'all':True, 'notimputed':False}[args.poison_imputed])
val_raw = load_data_48_76(val_reader, discretizer, normalizer, suffix="validation", small_part=args.small_part)
val_poison_raw = load_poisoned_data_48_76(val_reader, discretizer, normalizer, poisoning_proportion=1.0, poisoning_strength=args.poisoning_strength, suffix="train", small_part=args.small_part, poison_imputed={'all':True, 'notimputed':False}[args.poison_imputed])
#"""
if target_repl:
T = train_raw[0][0].shape[0]
def extend_labels(data):
data = list(data)
labels = np.array(data[1]) # (B,)
data[1] = [labels, None]
data[1][1] = np.expand_dims(labels, axis=-1).repeat(T, axis=1) # (B, T)
data[1][1] = np.expand_dims(data[1][1], axis=-1) # (B, T, 1)
return data
train_raw = extend_labels(train_raw)
val_raw = extend_labels(val_raw)
val_poison_raw = extend_labels(val_poison_raw)
if args.mode == 'train':
print("==> training")
input_dim = train_raw[0].shape[2]
train_data = train_raw[0].astype(np.float32)
train_targets = train_raw[1]
val_data = val_raw[0].astype(np.float32)
val_targets = val_raw[1]
val_poison_data = val_poison_raw[0].astype(np.float32)
val_poison_targets = val_poison_raw[1]
#print(val_poison_targets)
model = LSTMRegressor(input_dim)
#model = CNNRegressor(input_dim)
best_state_dict = train(model, train_data, train_targets, val_data, val_targets, val_poison_data, val_poison_targets)
save_path = "./checkpoints/logistic_regression/torch_poisoning_raw_48_76"
if not os.path.exists(save_path):
os.makedirs(save_path)
torch.save(best_state_dict, save_path + "/lstm_{}_{}_{}.pt".format(args.poisoning_proportion, args.poisoning_strength, args.poison_imputed))
elif args.mode == 'test':
# ensure that the code uses test_reader
del train_reader
del val_reader
del train_raw
del val_raw
test_reader = InHospitalMortalityReader(dataset_dir=os.path.join(args.data, 'test'),
listfile=os.path.join(args.data, 'test_listfile.csv'),
period_length=48.0)
ret = utils.load_data(test_reader, discretizer, normalizer, args.small_part,
return_names=True)
data = ret["data"][0]
labels = ret["data"][1]
names = ret["names"]
predictions = model.predict(data, batch_size=args.batch_size, verbose=1)
predictions = np.array(predictions)[:, 0]
metrics.print_metrics_binary(labels, predictions)
path = os.path.join(args.output_dir, "test_predictions", os.path.basename(args.load_state)) + ".csv"
utils.save_results(names, predictions, labels, path)
else:
raise ValueError("Wrong value for args.mode")
reader.read_example(0)["X"])[1].split(',')
cont_channels = [i for (i, x) in enumerate(
discretizer_header) if x.find("->") == -1]
# choose here which columns to standardize
normalizer = Normalizer(fields=cont_channels)
normalizer_state = args.normalizer_state
if normalizer_state is None:
normalizer_state = 'ihm_ts{}.input_str:{}.start_time:zero.normalizer'.format(
args.timestep, args.imputation)
normalizer_state = os.path.join(
os.path.dirname(__file__), normalizer_state)
normalizer.load_params(normalizer_state)
normalizer = None
train_raw = utils.load_data(
reader, discretizer, normalizer, args.small_part, return_names=True)
#val_raw = utils.load_data(val_reader, discretizer, normalizer, args.small_part)
print(len(train_raw['names']))
'''
1. Read text based on patient_id
2. merge text and time series based on patient_id
3. create numpy arrays
4. call Model functions.
5. train and evaluate.
'''
treader = TextReader(args.textdata)
train_text = treader.read_all_text_concat(train_raw['names'])
#val_raw = treader.read_all_text_concat(val_raw[''])
]
# choose here which columns to standardize
normalizer = Normalizer(fields=cont_channels)
normalizer_state = conf.norm_state
if normalizer_state is None:
normalizer_state = 'ihm_ts{}.input_str:{}.start_time:zero.normalizer'.format(
conf.timestep, conf.imputation)
normalizer_state = os.path.join(os.path.dirname(__file__),
normalizer_state)
normalizer.load_params(normalizer_state)
normalizer = None
train_raw = ihm_utils.load_data(train_reader,
discretizer,
normalizer,
conf.small_part,
return_names=True)
print("Number of train_raw_names: ", len(train_raw['names']))
text_reader = utils.TextReader(conf.textdata_fixed, conf.starttime_path)
train_text = text_reader.read_all_text_concat_json(train_raw['names'], 48)
data = utils.merge_text_raw(train_text, train_raw)
data_X = data[0]
data_y = data[1]
data_text = data[2]
del data
del train_raw
del train_text
loss_weights = None
model.compile(optimizer=optimizer_config, loss=loss, loss_weights=loss_weights)
## print model summary
model.summary()
# Load model weights
n_trained_chunks = 0
if args.load_state != "":
model.load_weights(args.load_state)
n_trained_chunks = int(
re.match(".*epoch([0-9]+).*", args.load_state).group(1))
# Read data
train_raw = utils.load_data(train_reader, discretizer, normalizer,
args.small_part)
val_raw = utils.load_data(val_reader, discretizer, normalizer, args.small_part)
if target_repl:
T = train_raw[0][0].shape[0]
def extend_labels(data):
data = list(data)
labels = np.array(data[1]) # (B,)
data[1] = [labels, None]
data[1][1] = np.expand_dims(labels, axis=-1).repeat(T,
axis=1) # (B, T)
data[1][1] = np.expand_dims(data[1][1], axis=-1) # (B, T, 1)
return data
train_raw = extend_labels(train_raw)
loss_weights = None
model.compile(optimizer=optimizer_config,
loss=loss,
loss_weights=loss_weights)
model.summary()
# Load model weights
n_trained_chunks = 0
if args.load_state != "":
model.load_weights(args.load_state)
n_trained_chunks = int(re.match(".*epoch([0-9]+).*", args.load_state).group(1))
# Read data
train_raw = utils.load_data(train_reader, discretizer, normalizer, args.small_part)
val_raw = utils.load_data(val_reader, discretizer, normalizer, args.small_part)
if target_repl:
T = train_raw[0][0].shape[0]
def extend_labels(data):
data = list(data)
labels = np.array(data[1]) # (B,)
data[1] = [labels, None]
data[1][1] = np.expand_dims(labels, axis=-1).repeat(T, axis=1) # (B, T)
data[1][1] = np.expand_dims(data[1][1], axis=-1) # (B, T, 1)
return data
train_raw = extend_labels(train_raw)
val_raw = extend_labels(val_raw)
args_dict['target_repl'] = target_repl
# Build the model
print("==> using model {}".format(args['network']))
K.clear_session()
model_module = imp.load_source(os.path.basename(args['network']),
args['network'])
model = model_module.Network(**args_dict)
model.load_weights(args['load_state'])
test_reader = InHospitalMortalityReader(
dataset_dir=os.path.join(args['data'], 'test'),
listfile=os.path.join(args['data'], 'test_listfile.csv'),
period_length=48.0)
ret = utils.load_data(test_reader,
discretizer,
normalizer,
args['small_part'],
return_names=True)
data = ret["data"][0]
labels = ret["data"][1]
names = ret["names"]
model = lrpify_model(model)
model.optimizer = optimizers.Adam(10**(-3), 0.9)
name = input("Enter target time series name")
print("Start cashing all activations")
index = names.index(name)
inp_data = data[[index]]
label = labels[index]
prediction = int(model.predict(inp_data)[0][0] > 0.5)
