def __init__(self, dataset_dir, listfile=None, listfile_data=None):
""" Reader for phenotype classification task.
:param dataset_dir: Directory where timeseries files are stored.
:param listfile: Path to a listfile. If this parameter is left `None` then
`dataset_dir/listfile.csv` will be used.
"""
Reader.__init__(self, dataset_dir, listfile)
self._data = listfile_data
self._data = [line.split(',') for line in self._data]
self._data = [(mas[0], float(mas[1]), map(int, mas[2:])) for mas in self._data]
self._X = []
self._header = []
label_struct = utils.read_hierarchical_labels('../../data/phenotyping/label_list.txt', '../../data/phenotyping/label_struct.json')
label_mapper = {}
for super_label in label_struct.keys():
label_mapper[super_label] = set(label_struct[super_label])
for index in range(len(self._data)):
name = self._data[index][0]
# t = self._data[index][1]
# y = self._data[index][2]
(X, header) = self._read_timeseries(name)
self._X.append(X)
self._header.append(header)
self._data[index][2].append(0)
self._data[index][2].append(0)
for i in range(25):
if self._data[index][2][i] == 1:
if i in label_mapper[25]:
self._data[index][2][25] = 1
if i in label_mapper[26]:
self._data[index][2][26] = 1
def __init__(self, dim, batch_norm, dropout, rec_dropout, task,
target_repl=False, deep_supervision=False, num_classes=1,
depth=1, input_dim=76, **kwargs):
print "==> not used params in network class:", kwargs.keys()
self.dim = dim
self.batch_norm = batch_norm
self.dropout = dropout
self.rec_dropout = rec_dropout
self.depth = depth
if task in ['decomp', 'ihm', 'ph']:
final_activation = 'sigmoid'
elif task in ['los']:
if num_classes == 1:
final_activation = 'relu'
else:
final_activation = 'softmax'
else:
return ValueError("Wrong value for task")
# Input layers and masking
X = Input(shape=(None, input_dim), name='X')
inputs = [X]
mX = Masking()(X)
if deep_supervision:
M = Input(shape=(None,), name='M')
inputs.append(M)
# Configurations
is_bidirectional = True
if deep_supervision:
is_bidirectional = False
# Main part of the network
for i in range(depth - 1):
num_units = dim
if is_bidirectional:
num_units = num_units // 2
gru = GRU(units=num_units,
activation='tanh',
return_sequences=True,
recurrent_dropout=rec_dropout,
dropout=dropout)
if is_bidirectional:
mX = Bidirectional(gru)(mX)
else:
mX = gru(mX)
# Output module of the network
return_sequences = (target_repl or deep_supervision)
L_lv1 = GRU(units=dim,
activation='tanh',
return_sequences=True,
dropout=dropout,
recurrent_dropout=rec_dropout)(mX)
L = L_lv1
if dropout > 0:
L = Dropout(dropout)(L)
label_struct = utils.read_hierarchical_labels('../../data/phenotyping/label_list.txt', '../../data/phenotyping/label_struct.json')
# only support 2 levels
num_superclass = len(label_struct.keys())
y_lv1 = {}
y_lv2 = {}
for class_lv1 in label_struct.keys():
y_lv1[class_lv1] = Dense(1, activation=final_activation)(Lambda(lambda x: x[:,-1,:])(L))
L_lv2_gru = GRU(units=dim,
activation='tanh',
return_sequences=return_sequences,
dropout=dropout,
recurrent_dropout=rec_dropout)(L_lv1)
if dropout > 0:
L_lv2_gru = Dropout(dropout)(L_lv2_gru)
y_lv2[class_lv1] = {}
for class_lv2 in label_struct[class_lv1]:
y_lv2[class_lv1][class_lv2] = Dense(1, activation=final_activation)(L_lv2_gru)
label_mapper = {}
for super_label in label_struct.keys():
label_mapper[super_label] = set(label_struct[super_label])
y_final = []
for i in range(25):
if (i in label_mapper[25]) and (i not in label_mapper[26]):
y_final.append(Multiply()([y_lv1[25], y_lv2[25][i]]))
elif (i not in label_mapper[25]) and (i in label_mapper[26]):
y_final.append(Multiply()([y_lv1[26], y_lv2[26][i]]))
elif (i in label_mapper[25]) and (i in label_mapper[26]):
y_final.append(Add()([Multiply()([y_lv1[25], y_lv2[25][i]]), Multiply()([y_lv1[26], y_lv2[26][i]])]))
y_final.append(y_lv1[25])
y_final.append(y_lv1[26])
y = Concatenate()(y_final)
outputs = [y]
return super(Network, self).__init__(inputs=inputs,
outputs=outputs)
def __init__(self,
dim,
batch_norm,
dropout,
rec_dropout,
task,
target_repl=False,
deep_supervision=False,
num_classes=1,
depth=1,
input_dim=76,
**kwargs):
print "==> not used params in network class:", kwargs.keys()
self.dim = dim
self.batch_norm = batch_norm
self.dropout = dropout
self.rec_dropout = rec_dropout
self.depth = depth
if task in ['decomp', 'ihm', 'ph']:
final_activation = 'sigmoid'
elif task in ['los']:
if num_classes == 1:
final_activation = 'relu'
else:
final_activation = 'softmax'
else:
return ValueError("Wrong value for task")
# Input layers and masking
X = Input(shape=(None, input_dim), name='X')
inputs = [X]
mX = Masking()(X)
if deep_supervision:
M = Input(shape=(None, ), name='M')
inputs.append(M)
# Configurations
is_bidirectional = True
if deep_supervision:
is_bidirectional = False
# Output module of the network
return_sequences = (target_repl or deep_supervision)
L_lv1 = GRU(units=dim,
activation='tanh',
return_sequences=True,
dropout=dropout,
recurrent_dropout=rec_dropout,
name="L_lv1")(mX)
L = L_lv1
label_struct = utils.read_hierarchical_labels(
'../../data/phenotyping/label_list.txt',
'../../data/phenotyping/label_struct.json')
# only support 2 levels
num_superclass = len(label_struct.keys())
output_lv1 = Lambda(lambda x: x[:, -1, :])(L)
# if dropout > 0:
# output_lv1 = Dropout(dropout)(output_lv1)
output_lv1 = Dense(num_superclass,
activation=final_activation,
name="output_lv1")(output_lv1)
L_lv2_gru = GRU(units=dim,
activation='tanh',
return_sequences=return_sequences,
dropout=dropout,
recurrent_dropout=rec_dropout,
name="L_lv2_gru")(L_lv1)
# if dropout > 0:
# L_lv2_gru = Dropout(dropout)(L_lv2_gru)
L_lv2_2 = GRU(units=dim,
activation='tanh',
return_sequences=return_sequences,
dropout=dropout,
recurrent_dropout=rec_dropout,
name="L_lv2_2")(mX)
output_lv2 = Dense(25, activation=final_activation, name="output_lv2")(
Concatenate()([L_lv2_gru, L_lv2_2]))
y = Concatenate()([output_lv2, output_lv1])
outputs = [y]
return super(Network, self).__init__(inputs=inputs, outputs=outputs)
