def _perform_augmentation_segmentation(self,
img,
label,
mask_image,
augmenter,
event=None,
events=False):
seq_image, seq_label, seq_mask, seq_event = get_augmenter(
name=augmenter, c_val=255)
# apply some contrast to de rgb image
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image.augment_images(img)
img = img.reshape(img.shape[1:])
label = label.reshape(sum(((1, ), label.shape), ()))
label = seq_label.augment_images(label)
label = label.reshape(label.shape[1:])
mask_image = mask_image.reshape(sum(((1, ), mask_image.shape), ()))
mask_image = seq_mask.augment_images(mask_image)
mask_image = mask_image.reshape(mask_image.shape[1:])
if events:
event = event.reshape(sum(((1, ), event.shape), ()))
# event = self.augment_event(event)
event = seq_event.augment_images(event)
event = event.reshape(event.shape[1:])
return img, label, mask_image, event
return img, label, mask_image
def _perform_augmentation_segmentation(self, img, label, mask_image,
augmenter):
seq_image_contrast, seq_image_translation, seq_label, seq_mask = get_augmenter(
name=augmenter, c_val=self.ignore_label)
# apply some contrast to de rgb image
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image_contrast.augment_images(img)
img = img.reshape(img.shape[1:])
# Apply shifts and rotations to the mask, labels and image
# Reshapes for the AUGMENTER framework
# the loops are due to the external library failures
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image_translation.augment_images(img)
img = img.reshape(img.shape[1:])
label = label.reshape(sum(((1, ), label.shape), ()))
label = seq_label.augment_images(label)
label = label.reshape(label.shape[1:])
mask_image = mask_image.reshape(sum(((1, ), mask_image.shape), ()))
mask_image = seq_mask.augment_images(mask_image)
mask_image = mask_image.reshape(mask_image.shape[1:])
return img, label, mask_image
def _get_batch_rgb(self, size=32, train=True, augmenter=None):
x = np.zeros([size, self.height, self.width, self.dim],
dtype=np.float32)
y = np.zeros([size], dtype=np.uint8)
if train:
file_list = self.train_list
folder = '/train/'
# Get [size] random numbers
indexes = [
i % len(file_list)
for i in range(self.index_train, self.index_train + size)
]
self.index_train = indexes[-1] + 1
else:
file_list = self.test_list
folder = '/test/'
# Get [size] random numbers
indexes = [
i % len(file_list)
for i in range(self.index_test, self.index_test + size)
]
self.index_test = indexes[-1] + 1
random_files = [file_list[number] for number in indexes]
classes = [
self.classes[file.split(folder)[1].split('/')[0]]
for file in random_files
]
for index in range(size):
img = cv2.imread(random_files[index])
if img is None:
print(random_files[index])
print(indexes[index])
if img.shape[1] != self.width or img.shape[0] != self.height:
img = cv2.resize(img, (self.width, self.height),
interpolation=cv2.INTER_AREA)
x[index, :, :, :] = img
y[index] = classes[index]
# the labeling to categorical (if 5 classes and value is 2: 2 -> [0,0,1,0,0])
y = to_categorical(y, num_classes=self.n_classes)
# augmentation
if augmenter:
augmenter_seq = get_augmenter(name=augmenter)
x = augmenter_seq.augment_images(x)
# x = x.astype(np.float32)
# tf.keras.applications.imagenet_utils.preprocess_input(x, mode='tf')
# x = tf.keras.applications.xception.preprocess_input(x)
x = x.astype(np.float32) / 255.0 - 0.5
return x, y
def _perform_augmentation_segmentation(self, img, label, mask_image,
augmenter):
seq_image, seq_label, seq_mask = get_augmenter(name=augmenter,
c_val=255)
#apply some contrast to de rgb image
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image.augment_images(img)
img = img.reshape(img.shape[1:])
label = label.reshape(sum(((1, ), label.shape), ()))
label = seq_label.augment_images(label)
label = label.reshape(label.shape[1:])
mask_image = mask_image.reshape(sum(((1, ), mask_image.shape), ()))
mask_image = seq_mask.augment_images(mask_image)
mask_image = mask_image.reshape(mask_image.shape[1:])
return img, label, mask_image
def _get_batch_rgb(self, size=32, train=True, augmenter=None):
x = np.zeros([size, self.height, self.width, self.dim],
dtype=np.float32)
y = np.zeros([size], dtype=np.uint8)
file_list = self.test_list
folder = '/test/'
if train:
file_list = self.train_list
folder = '/train/'
# Get [size] random numbers
random_files = [
file_list[random.randint(0,
len(file_list) - 1)]
for file in range(size)
]
classes = [
self.classes[file.split(folder)[1].split('/')[0]]
for file in random_files
]
for index in range(size):
img = cv2.imread(random_files[index])
if img.shape[1] != self.width and img.shape[0] != self.height:
img = cv2.resize(img, (self.width, self.height),
interpolation=cv2.INTER_AREA)
x[index, :, :, :] = img
y[index] = classes[index]
# the labeling to categorical (if 5 classes and value is 2: 2 -> [0,0,1,0,0])
y = to_categorical(y, num_classes=len(self.classes))
# augmentation
if augmenter:
augmenter_seq = get_augmenter(name=augmenter)
x = augmenter_seq.augment_images(x)
x = x.astype(np.float32) / 255.0 - 0.5
return x, y
def _get_batch_segmentation(self,
size=32,
train=True,
augmenter=None,
index=None,
validation=False):
x = np.zeros([size, self.height, self.width, self.dim],
dtype=np.float32)
y = np.zeros([size, self.height, self.width], dtype=np.uint8)
mask_expanded = np.ones(
[size, self.height, self.width, self.n_classes], dtype=np.uint8)
image_list = self.image_test_list
label_list = self.label_test_list
folder = '/test/'
if train:
image_list = self.image_train_list
label_list = self.label_train_list
folder = '/train/'
# Get [size] random numbers
indexes = [
random.randint(0,
len(image_list) - 1) for file in range(size)
]
if index:
indexes = [i for i in range(index, index + size)]
random_images = [image_list[number] for number in indexes]
random_labels = [label_list[number] for number in indexes]
# for every random image, get the image, label and mask.
# the augmentation has to be done separately due to augmentation
for index in range(size):
img = cv2.imread(random_images[index])
label = cv2.imread(random_labels[index], 0)
if img.shape[1] != self.width and img.shape[0] != self.height:
img = cv2.resize(img, (self.width, self.height),
interpolation=cv2.INTER_AREA)
if label.shape[1] != self.width and label.shape[0] != self.height:
label = cv2.resize(label, (self.width, self.height),
interpolation=cv2.INTER_NEAREST)
macara = mask_expanded[index, :, :, 0]
if train and augmenter and random.random() < 0.90:
seq_image2, seq_image, seq_label, seq_mask = get_augmenter(
name=augmenter, c_val=self.ignore_label)
#apply some contrast to de rgb image
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image2.augment_images(img)
img = img.reshape(img.shape[1:])
if random.random() < 0.90:
#Apply shifts and rotations to the mask, labels and image
# Reshapes for the AUGMENTER framework
# the loops are due to the external library failures
cuenta_ignore = sum(sum(sum(img == self.ignore_label)))
cuenta_ignore2 = cuenta_ignore
i = 0
while abs(cuenta_ignore2 - cuenta_ignore) < 5 and i < 15:
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image.augment_images(img)
img = img.reshape(img.shape[1:])
cuenta_ignore2 = sum(sum(
sum(img == self.ignore_label)))
i = i + 1
cuenta_ignore = sum(sum(label == self.ignore_label))
cuenta_ignore2 = cuenta_ignore
i = 0
while cuenta_ignore2 == cuenta_ignore and i < 15:
label = label.reshape(sum(((1, ), label.shape), ()))
label = seq_label.augment_images(label)
label = label.reshape(label.shape[1:])
cuenta_ignore2 = sum(sum(label == self.ignore_label))
i = i + 1
cuenta_ignore = sum(sum(macara == self.ignore_label))
cuenta_ignore2 = cuenta_ignore
i = 0
while cuenta_ignore2 == cuenta_ignore and i < 15:
macara = macara.reshape(sum(((1, ), macara.shape), ()))
macara = seq_mask.augment_images(macara)
macara = macara.reshape(macara.shape[1:])
cuenta_ignore2 = sum(sum(macara == self.ignore_label))
i = i + 1
if self.ignore_label and not validation:
#ignore_label to value 0-n_classes and add it to mask
mask_ignore = label == self.ignore_label
macara[mask_ignore] = 0
label[mask_ignore] = 0
x[index, :, :, :] = img
y[index, :, :] = label
for i in xrange(mask_expanded.shape[3]):
mask_expanded[index, :, :, i] = macara
# the labeling to categorical (if 5 classes and value is 2: 2 -> [0,0,1,0,0])
a, b, c = y.shape
y = y.reshape((a * b * c))
if self.ignore_label and validation:
y = to_categorical(y, num_classes=self.n_classes + 1)
else:
y = to_categorical(y, num_classes=self.n_classes)
y = y.reshape((a, b, c, self.n_classes)).astype(np.uint8)
x = x.astype(np.float32) / 255.0 - 0.5
return x, y, mask_expanded
except:
pass
# all the trainable ops
tf.summary.histogram(op.name, op)
total_parameters = 0
for variable in tf.trainable_variables():
# shape is an array of tf.Dimension
shape = variable.get_shape()
variable_parameters = 1
for dim in shape:
variable_parameters *= dim.value
total_parameters += variable_parameters
print("Total parameters of the net: " + str(total_parameters))
augmenter_seq = get_augmenter(name='caltech')
show_each_steps = 100
saver = tf.train.Saver(tf.global_variables())
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state('./model') # './model/best'
ckpt_best = tf.train.get_checkpoint_state('./model/best') # './model/best'
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
saver.restore(sess, ckpt.model_checkpoint_path)
else:
sess.run(tf.global_variables_initializer())
merged = tf.summary.merge_all()
writer_train = tf.summary.FileWriter('./logs/train', sess.graph)
writer_test = tf.summary.FileWriter('./logs/test', sess.graph)
def _get_batch_segmentation(self, size=32, train=True, augmenter=None):
# init numpy arrays
x = np.zeros([size, self.height, self.width, self.dim],
dtype=np.float32)
y = np.zeros([size, self.height, self.width], dtype=np.uint8)
mask = np.ones([size, self.height, self.width], dtype=np.float32)
if train:
image_list = self.image_train_list
label_list = self.label_train_list
index_files = self.index_train
folder = '/train/'
# Get [size] random numbers
indexes = [
i % len(image_list)
for i in range(self.index_train, self.index_train + size)
]
self.index_train = indexes[-1] + 1
else:
image_list = self.image_test_list
label_list = self.label_test_list
folder = '/test/'
# Get [size] random numbers
indexes = [
i % len(image_list)
for i in range(self.index_test, self.index_test + size)
]
self.index_test = indexes[-1] + 1
random_images = [image_list[number] for number in indexes]
random_labels = [label_list[number] for number in indexes]
# for every random image, get the image, label and mask.
# the augmentation has to be done separately due to augmentation
for index in range(size):
if self.dim == 1:
img = cv2.imread(random_images[index], 0)
else:
img = cv2.imread(random_images[index])
label = cv2.imread(random_labels[index], 0)
# check if error
if img is None or label is None:
print(random_images[index])
print(random_labels[index])
print(indexes[index])
if img.shape[1] != self.width or img.shape[0] != self.height:
img = cv2.resize(img, (self.width, self.height),
interpolation=cv2.INTER_AREA)
if label.shape[1] != self.width or label.shape[0] != self.height:
label = cv2.resize(label, (self.width, self.height),
interpolation=cv2.INTER_NEAREST)
mask_image = mask[index, :, :]
if train and augmenter:
seq_image_contrast, seq_image_translation, seq_label, seq_mask = get_augmenter(
name=augmenter, c_val=self.ignore_label)
#apply some contrast to de rgb image
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image_contrast.augment_images(img)
img = img.reshape(img.shape[1:])
#Apply shifts and rotations to the mask, labels and image
# Reshapes for the AUGMENTER framework
# the loops are due to the external library failures
img = img.reshape(sum(((1, ), img.shape), ()))
img = seq_image_translation.augment_images(img)
img = img.reshape(img.shape[1:])
label = label.reshape(sum(((1, ), label.shape), ()))
label = seq_label.augment_images(label)
label = label.reshape(label.shape[1:])
mask_image = mask_image.reshape(
sum(((1, ), mask_image.shape), ()))
mask_image = seq_mask.augment_images(mask_image)
mask_image = mask_image.reshape(mask_image.shape[1:])
# modify the mask and the labels. Mask
mask_ignore = label == self.ignore_label
mask_image[
mask_ignore] = 0 # The ignore pixels will have a value o 0 in the mask
label[
mask_ignore] = self.n_classes # The ignore label will be n_classes
if self.dim == 1:
img = np.reshape(img, (img.shape[0], img.shape[1], self.dim))
x[index, :, :, :] = img
y[index, :, :] = label
mask[index, :, :] = mask_image
# Apply weights to the mask
mask = self._from_binarymask_to_weighted_mask(y, mask)
# the labeling to categorical (if 5 classes and value is 2: 2 -> [0,0,1,0,0])
a, b, c = y.shape
y = y.reshape((a * b * c))
# Convert to categorical. Add one class for ignored pixels
y = to_categorical(y, num_classes=self.n_classes + 1)
y = y.reshape((a, b, c, self.n_classes + 1)).astype(np.uint8)
#tf.keras.applications.imagenet_utils.preprocess_input(x, mode='tf')
#x = tf.keras.applications.xception.preprocess_input(x)
x = x.astype(np.float32) / 255.0 - 0.5
return x, y, mask