def execute(self, image_path):
"""
Show the result of the network on the given image
:param image_path: The path of the image
"""
is_cuda = next(self.model.parameters()).is_cuda
image = Image.open(image_path).convert('RGB')
width, height = image.size
new_width = math.sqrt(6 * (10**5) * width / height)
new_width = int(new_width)
new_height = height * new_width // width
resized = image.resize((new_width, new_height), Image.ANTIALIAS)
image = torch.from_numpy(image_pillow_to_numpy(image))
resized = torch.from_numpy(image_pillow_to_numpy(resized))
if is_cuda:
result = self.model(
torch.autograd.Variable(
resized.unsqueeze(0).float().cuda()))[0]
else:
result = self.model(
torch.autograd.Variable(resized.unsqueeze(0).float().cpu()))[0]
self.loss.show_ytrue(resized.cpu().numpy(),
result.cpu().detach().numpy())
lines = self.loss.ytrue_to_lines(image.cpu().numpy(),
result.cpu().detach().numpy())
for line, pos in lines:
show_numpy_image(line, invert_axes=True)
def __getitem__(self, index):
image = Image.open(self.list[index]).convert("RGB")
width, height = image.size
new_width = math.sqrt(6 * (10**5) * width / height)
new_width = int(new_width)
new_height = height * new_width // width
if new_width < width:
resized = image.resize((new_width, new_height), Image.ANTIALIAS)
else:
resized = image
return image_pillow_to_numpy(resized), image_pillow_to_numpy(
image), self.list[index]
def __getitem__(self, index):
image_path, regions = self.labels[index % len(self.labels)]
image = Image.open(image_path).convert('RGB')
width, height = image.size
new_width = math.sqrt(6 * (10**5) * width / height)
# new_width = new_width * uniform(0.8, 1.2)
new_width = int(new_width)
new_height = height * new_width // width
resized = image.resize((new_width, new_height), Image.ANTIALIAS)
new_regions = []
for region in regions:
x0, y0, x1, y1, h = region
x0 = x0 * new_height // height
y0 = y0 * new_height // height
x1 = x1 * new_height // height
y1 = y1 * new_height // height
h = h * new_height // height
new_regions.append([x0, y0, x1, y1, h])
if len(new_regions) == 0:
label = np.zeros((new_height, new_width, 2), dtype=np.float32)
else:
label = self.loss.document_to_ytrue(
np.array([new_width, new_height], dtype='int32'),
np.array(new_regions, dtype='int32'))
# angle = randint(-45, 45)
# label = rotate(label, angle, order=0)
# image = rotate(image, angle, order=0)
# image = np.swapaxes(image, 0, 2)
# image = np.swapaxes(image, 1, 2)
label = np.swapaxes(label, 0, 2)
label = np.swapaxes(label, 1, 2)
# if self.transform:
# image = self.helper.augment(image, distort=False)
# c = gauss_distort([image[0], image[1], image[2], label[0], label[1]])
# image = np.stack((c[0], c[1], c[2]), axis=0)
# label = np.stack((c[3], c[4]), axis=0)
return image_pillow_to_numpy(resized), image_pillow_to_numpy(
image), image_path, label
def __getitem__(self, index):
image_pillow, label = self.generate_image_with_label(index)
image = image_pillow_to_numpy(image_pillow)
try:
return torch.from_numpy(image), (self.loss.preprocess_label(
label, image.shape[2]), label, image.shape[2])
except:
return self.__getitem__(index)
def __getitem__(self, index):
id, text = self.labels[index]
ids = id.split("-")
image_path = os.path.join(self.images_path, ids[0] + "/" + ids[0] + "-" + ids[1] + "/" + ids[0] + "-" + ids[1] + "-" + ids[2] + ".png")
# Load the image
image = Image.open(image_path).convert('RGB')
width, height = image.size
image = image.resize((width * self.height // height, self.height), Image.ANTIALIAS)
if self.transform:
image = self.document_helper.paster_into_random_background(image)
image = image_pillow_to_numpy(image)
return torch.from_numpy(image), (self.loss.preprocess_label(text, width * self.height // height), text, image.shape[2])
def __getitem__(self, index):
image_path, text = self.labels[index]
# Load the image
image = Image.open(image_path).convert('RGB')
width, height = image.size
if self.height is not None:
image = image.resize((width * self.height // height, self.height),
Image.ANTIALIAS)
if self.transform:
image = self.document_helper.paster_into_random_background(image)
image = image_pillow_to_numpy(image)
return torch.from_numpy(image), (self.loss.preprocess_label(
text, width * self.height // height), text, image.shape[2])