def batch_yielder(self):
"""
batch generator for the training data
"""
while True:
indexes = list(range(len(self.label_list)))
shuffle(indexes) #shuffling the indexes to have a variety of classes in a batch
nbr_of_batches = len(indexes) // self.batch_size
for i in range(nbr_of_batches):
start = i * self.batch_size
end = (i + 1) * self.batch_size
batch = []
outputs = []
for j in range(start, end):
img_path = (
root
+ "data/train_images/train_images/"
+ self.label_list[indexes[j]][0]
+ ".png"
)
img = cv2.imread(img_path)
if not self.mobile:
img = colour_to_grey(img)
mod_type =randint(0, 7) #randomly choose a transformation to apply to the image at hand
img=dataGenerator(img, mod_type, self.datagen)
batch.append(img)
outputs.append(np.array(self.label_list[indexes[j]][1]))
batch = np.array(batch)
outputs = np.array(outputs)
yield (batch, outputs)
# A last loop to get the last of the data
batch = []
outputs = []
for j in range(nbr_of_batches * self.batch_size, len(indexes)):
img_path = (
root
+ "data/train_images/train_images/"
+ self.label_list[indexes[j]][0]
+ ".png"
)
img = cv2.imread(img_path)
if not self.mobile:
img = colour_to_grey(img)
mod_type =randint(0, 7) #randomly choose a transformation to apply to the image at hand
img=dataGenerator(img, mod_type, self.datagen)
batch.append(img)
outputs.append(np.array(self.label_list[indexes[j]][1]))
batch = np.array(batch)
outputs = np.array(outputs)
yield (batch, outputs)
def validation_batch_yielder(self):
"""
batch generator for validation data
This function has practically the same structure as the test_generator only we don't shuffle the indexes nor do we augment the data
"""
while True:
indexes = list(range(len(self.validation_list)))
nbr_of_batches = len(indexes) // self.batch_size
for i in range(nbr_of_batches):
start = i * self.batch_size
end = (i + 1) * self.batch_size
batch = []
outputs = []
for j in range(start, end):
img_path = (
root
+ "data/train_images/train_images/"
+ self.validation_list[j][0]
+ ".png"
)
img = cv2.imread(img_path)
if not self.mobile:
img = colour_to_grey(img)
batch.append(img)
outputs.append(np.array(self.validation_list[j][1]))
batch = np.array(batch)
outputs = np.array(outputs)
yield (batch, outputs)
batch = []
outputs = []
for j in range(nbr_of_batches * self.batch_size, len(indexes)):
img_path = (
root
+ "data/train_images/train_images/"
+ self.validation_list[j][0]
+ ".png"
)
img = cv2.imread(img_path)
if not self.mobile:
img = colour_to_grey(img)
batch.append(img)
outputs.append(np.array(self.validation_list[j][1]))
batch = np.array(batch)
outputs = np.array(outputs)
yield (batch, outputs)
def test_generator(test_data, network):
for i in range(len(test_data)):
img_path = (root + "data/train_images/train_images/" +
test_data[i][0] + ".png")
img = cv2.imread(img_path)
if network != "mobilenet":
img = colour_to_grey(img)
yield (np.array([img]))
def generator(data, network, outputs_list):
for i in range(len(data)):
img_path = (root + "data/train_images/train_images/" + data[i][0] +
".png")
outputs_list.append([data[i][1]])
img = cv2.imread(img_path)
if network != "mobilenet":
img = colour_to_grey(img)
yield (np.array([img]))
outputs_list = np.array(outputs_list)
def test_batch_yielder(self):
"""
generator for test data
"""
test_img_paths = os.listdir(root + "data/test_images/test_images")
for j in range(len(test_img_paths)):
batch = []
img_path = root + "data/test_images/test_images/" + test_img_paths[j]
img = cv2.imread(img_path)
if not self.mobile:
img = colour_to_grey(img) # we have to turn it to grey img since tht is what our network was trained on
batch.append(img)
batch = np.array(batch)
yield (batch)
def __getitem__(self, idx):
start = idx * self.batch_size
end = (idx + 1) * self.batch_size
batch = []
outputs = []
nbr_pos = 0
nbr_neg = 0
for index in self.indexes[start:end]:
img_path = (root + "data/train_images/train_images/" +
self.data[index][0] + ".png")
img = cv2.imread(img_path)
if self.network != "mobilenet":
img = colour_to_grey(img)
if self.data_type == "train":
mod_type = randint(
0, 7
) #randomly choose a transformation to apply to the image at hand
img = dataGenerator(img, mod_type, self.datagen)
if self.data[index][1] == 1:
nbr_pos += 1
else:
nbr_neg += 1
batch.append(img)
outputs.append(self.data[index][1])
# #***try to balance batches*******
# if self.data_type=="train":
# print(self.nbr_pos_in_train)
# for k in range (2*int(nbr_neg/max(1,nbr_pos))):
# position=randint(0,self.nbr_pos_in_train-1)
# outputs.append(self.data[index][1])
# img_path = (
# root
# + "data/train_images/train_images/"
# + self.data[index][0]
# + ".png"
# )
# mod_type =randint(0, 7) #randomly choose a transformation to apply to the image at hand
# img=dataGenerator(img, mod_type, self.datagen)
# batch.append(img)
# print (outputs)
batch = np.array(batch)
outputs = np.array(outputs)
return (batch, outputs)