def Load(self,
path: str = 'data',
patterns: List[str] = ['jpg', 'jpeg', 'png'],
name="images",
input=None,
in_memory=False):
'''load data from folders, where each folder represents separate class'''
global res
if self.model.mode == 'train':
return
if self.model.mode == "design":
if input is None:
input = name
folders = glob.glob(path + '/*/')
#print(folders)
image_names = []
image_names = reduce(lambda x, y: x + y,
[[(glob.glob(x + '*.' + _type), x)
for x in folders] for _type in patterns])
image_list: NodeList = NodeList(self.graph)
for iname in image_names:
images, image_class = iname
image_class = image_class.split('/')[-2]
image_list = image_list + [
Node(
self.graph, {
"image": mvImage(x, in_memory=in_memory),
"target_class": image_class.lower()
}) for x in images
]
random.shuffle(image_list)
self.model.metadic[name] = NodeList.from_list(
image_list, self.graph)
self.model.capactity = len(self.model.metadic)
print("Loaded " + str(len(image_list)) + " images")
res = image_list
self.model.class_source = name
self.model.last_call = name
self.model.mtype = "Normal"
self.model.start_point = name
self.model.record('Image.Load', ['name', 'input'], [name, input])
if self.model.mode == "predict":
if type(self.model.metadic[input]) is str:
path = self.model.metadic[input]
images = (glob.glob(path))
image_list = [
Node(
self.graph, {
"image": mvImage(x, in_memory=in_memory),
"target_class": "unknown"
}) for x in images
]
self.model.metadic[name] = NodeList.from_list(
image_list, self.graph)
self.model.capactity = len(self.model.metadic)
return self
def LoadPDF(path, in_memory=False):
''' this is special function for loading images from pdf file
it actuall creates png images of pages on disk! requires WAND and ImageMagik'''
try:
from wand.image import Image #another local import to avoid proliferation of auxiliary dependencies
except:
print(
"ERROR:Can not import Wand library, that is needed for reading PDF. Please make sure wand http://docs.wand-py.org/ is installed"
)
return
import neuthink.metagraph as m # can't import before, will get circular import
diag = path
graph = Graph()
image_list: m.dNodeList = m.dNodeList(graph)
#if you see not authorized error, run this from console: sudo sed -i '/PDF/s/none/read|write/' /etc/ImageMagick-6/policy.xml
#also see here https://stackoverflow.com/questions/42928765/convertnot-authorized-aaaa-error-constitute-c-readimage-453
with (Image(filename=diag, resolution=300)) as source:
images = source.sequence
pages = len(images)
for i in range(pages):
Image(images[i]).save(filename=str(i) + '.png')
image = mvImage(path=str(i) + '.png', in_memory=in_memory)
imnode = Node(graph, {'image': image, 'type': 'image'})
image_list.append(imnode)
return image_list
def UnVectorize(self, source=None, target=None):
if self.model.mode == "train":
return
if source in self.model.metatensor:
data = self.model.metatensor[source]
if self.model.device is not None:
data = data.cpu()
data = data.detach()
if len(data.shape) == 4:
data = data.squeeze(1)
data = data.numpy()
minrange = min(self.model.batch_size,
len(self.model) - self.model.index)
for i in range(self.model.index, self.model.index + minrange):
self.model[i][target] = mvImage(
"",
source=Image.fromarray(
(data[i - self.model.index] * 255).astype('float32')),
in_memory=True)
self.res = [
x[target]
for x in self.model[self.model.index:(self.model.index +
minrange)]
]
self.model.result = self.res
if self.model.mode == 'design':
self.model.record("Image.UnVectorize", ['source', 'target'],
[source, target])
return self.model
def MakeSubimages(image: mvImage, stride: int, size: int) -> List[mvImage]:
'''takes one big image and splits it into subimages of given size using specified stride'''
newimages = []
# print(image.width)
for i in range(0, int(image.width), stride):
for j in range(0, int(image.height), stride):
im2 = mvImage("",
in_memory=True,
source=image.content.crop(
(i, j, i + size, j + size)))
newimages.append(im2)
return newimages
def make_list(image_names, image_list, patterns):
for iname in image_names:
# print(iname)
images, image_class = iname
if len(image_class) > 0:
image_class = image_class.split('/')[-2]
#image_name = image_class.
image_list = image_list + [
Node(
graph, {
"image": mvImage(x, in_memory=in_memory),
target_class: image_class.lower(),
'type': 'image'
}) for x in images
]
return image_list
def JoinBoundingBoxes(scene: NodeList, min_distance=None, comp_class=None):
'''joins all components of type comp_class that are closer then mindistance or if they bounding boxes intersect '''
#TODO: Code really needs cleanup and simplification
#TODO: Need to have an ability to specify vertical and horizontal distance separtely
def center(component1):
cx = (component1[0][0] + component1[1][0]) / 2
cy = (component1[0][1] + component1[1][1]) / 2
return (cx, cy)
def distance(component1, component2):
c1 = center(component1)
c2 = center(component2)
return math.sqrt((c1[0] - c2[0])**2 + (c1[1] - c2[1])**2)
def boundary_distance(x, y):
left_right = sorted([x, y], key=lambda x: x[0][0])
top_down = sorted([x, y], key=lambda x: x[1][1])
left_d = left_right[1][0][0] - left_right[0][1][0]
top_d = top_down[1][1][1] - top_down[0][1][1]
return (left_d, top_d)
is_finished = False
while not is_finished:
is_finished = True
newnodes = NodeList(scene.parent_graph)
for x in scene:
if not "_state_delete" in x:
for y in scene:
if x != y and not "_state_delete" in y:
left_d, top_d = boundary_distance(
x['location'], y['location'])
print(left_d, top_d)
if (left_d < min_distance and top_d < min_distance
) and x['target_class'] == comp_class and y[
'target_class'] == comp_class and not "_state_delete" in x and not "_state_delete" in y:
p1 = x.Parents({})
p2 = y.Parents({})
if len(p1) > 0 and len(p2) > 0:
if p1[0] != p2[0]:
continue
left_right = sorted(
[x, y], key=lambda x: x['location'][0][0])
top_down = sorted(
[x, y], key=lambda x: x['location'][0][1])
print("---------------")
print(x, y)
print(left_right, top_down)
newbox = (left_right[0]['location'][0][0],
top_down[0]['location'][0][1]), (
(left_right[1]['location'][1][0],
top_down[1]['location'][1][1]))
print(newbox)
print("---------------")
new_image = mvImage(
"",
in_memory=True,
source=x['base_image']['image'].content.crop(
(newbox[0][0], newbox[0][1], newbox[1][0],
newbox[1][1])))
new_image_node = Node(
scene.parent_graph, {
'image': new_image,
'location': newbox,
'base_image': x['base_image'],
'target_class': comp_class,
'type': 'image'
})
newnodes.append(new_image_node)
x.Parents({}).First().Connect(new_image_node)
x['_state_delete'] = 'yes'
y['_state_delete'] = 'yes'
is_finished = False
scene = scene + newnodes
#remove
scene.Match({"_state_delete": 'yes'}).Delete()
return