def class_balancing(self):
"""
Repeat elements from smaller class until the class are balanced
"""
if len(self.params["distance"]) < 3:
patient_list = np.empty(shape=(0, 2))
for ID in self.list_IDs_original:
if self.params["partition"] == 'read':
ID = ID.replace("]", '').replace("[", '')
ID = ID.replace("'", '').split(", ")
ID = np.asarray(ID)
ID = ID.reshape((1, 2))
ID[0, 1] = int(ID[0, 1])
patient_list = np.append(patient_list, ID, axis=0)
labels = yReadFunction(patient_list, self.params)
malignant = patient_list[labels == 1]
tot_num = len(labels)
mal_num = int(np.sum(labels))
imbal = ((tot_num - mal_num) / mal_num) - 1
if imbal >= 1:
r = np.floor(imbal).astype(int)
for j in range(r):
patient_list = np.append(patient_list,
malignant, axis=0)
imbal = imbal - r
r = int(imbal * mal_num)
r = random.sample(range(mal_num), r)
patient_list = np.append(patient_list, malignant[r],
axis=0)
return patient_list
else:
labels = yReadFunction(self.list_IDs, self.params)
malignant = self.list_IDs[labels == 1]
tot_num = len(labels)
mal_num = int(np.sum(labels))
imbal = ((tot_num - mal_num) / mal_num) - 1
if imbal >= 1:
r = np.floor(imbal).astype(int)
for j in range(r):
new_patients = self.make_list()
new_labels = yReadFunction(new_patients,
self.params)
new_malignant = new_patients[new_labels == 1]
self.list_IDs = np.append(self.list_IDs,
new_malignant, axis=0)
imbal = imbal - r
r = int(imbal * mal_num)
r = random.sample(range(mal_num), r)
new_patients = self.make_list()
new_labels = yReadFunction(new_patients, self.params)
new_malignant = new_patients[new_labels == 1]
self.list_IDs = np.append(self.list_IDs,
new_malignant[r], axis=0)
new_labels = yReadFunction(self.list_IDs, self.params)
return self.list_IDs
def check_class_balance(partition, params, balance = (0.1,0.1)) :
label = params["label"]
partition["train"] = np.asarray(partition["train"])
partition["validation"] = np.asarray(partition["validation"])
train_label = yReadFunction(partition["train"],params)
val_label = yReadFunction(partition["validation"],params)
train = np.sum(train_label)/len(train_label)
val = np.sum(val_label)/len(val_label)
tot = (np.sum(val_label)+np.sum(train_label))/(len(val_label)+len(train_label))
if (abs((tot-train)/tot) > balance[2]) | (abs((tot-val)/tot) > balance[1]):
return False
else : return True
def prepare_batch(self, list_IDs):
"""
Prepare a batch of data:
creating a list of images and masks
after having possibly augmented them
saving a few examples to disk if required
"""
X = list()
Y = list()
S = list()
xReadFunction = self.params['xReadFunction']
yReadFunction = self.params['yReadFunction']
for ID in list_IDs:
if not isinstance(ID, np.ndarray):
ID = ID.replace("]", '').replace("[", '')
ID = ID.replace("'", '').split(", ")
ID[1] = int(ID[1])
ID = np.asarray(ID)
if len(self.params["distance"]) < 3:
x, seg = xReadFunction(ID, self.params, im_mask="both",
data=self.params["data"])
else :
x, seg, ID[1] = xReadFunction(ID, self.params,
im_mask="both", data=self.params["data"])
x, seg = self.imaugment(x, seg)
if self.params["only"] == "both":
x = np.concatenate([x, seg], axis=-1)
y = yReadFunction(ID, self.params)
X.append(x)
Y.append(y)
if self.params["scalars"] is True:
scalarsReadFunction = self.params['scalarsReadFunction']
scalars = scalarsReadFunction(ID, self.params["source"],
data=self.params["data"])
S.append(scalars)
S = np.asarray(S)
X = np.asarray(X)
Y = np.asarray(Y)
if len(self.params["distance"]) > 2:
dz = self.params["distance"][2]
else:
dz = 1
self.save_images(X[...,0:1], X[...,1:2], list_IDs,
overlay = False)
if self.params["scalars"] is False:
return X, Y
else:
return [X, S], Y
def prepare_batch(self, list_IDs):
"""
Prepare a batch of data:
creating a list of images and masks
after having possibly augmented them
saving a few examples to disk if required
"""
X = list()
Y = list()
S = list()
xReadFunction = self.params['xReadFunction']
yReadFunction = self.params['yReadFunction']
for ID in list_IDs:
if not isinstance(ID, np.ndarray):
ID = ID.replace("]", '').replace("[", '')
ID = ID.replace("'", '').split(", ")
ID[1] = int(ID[1])
ID = np.asarray(ID)
if len(self.params["distance"]) < 3:
x, seg = xReadFunction(ID,
self.params,
im_mask="both",
data=self.params["data"])
else:
x, seg, ID[1] = xReadFunction(ID,
self.params,
im_mask="both",
data=self.params["data"])
x, seg = self.imaugment(x, seg)
x = np.concatenate([x, seg], axis=-1)
y = yReadFunction(ID, self.params)
X.append(x)
Y.append(y)
X = np.asarray(X)
Y = np.asarray(Y)
return X, Y
def class_balancing(self):
"""
Repeat elements from smaller class until the class are balanced
"""
patient_list = self.list_IDs_original
for ID in self.list_IDs_original:
labels = yReadFunction(patient_list, self.params)
malignant = patient_list[labels == 1]
tot_num = len(labels)
mal_num = int(np.sum(labels))
imbal = ((tot_num - mal_num) / mal_num) - 1
if imbal >= 1:
r = np.floor(imbal).astype(int)
for j in range(r):
patient_list = np.append(patient_list, malignant, axis=0)
imbal = imbal - r
r = int(imbal * mal_num)
r = random.sample(range(mal_num), r)
patient_list = np.append(patient_list, malignant[r], axis=0)
return patient_list
