def __call__(self, batch):
images, labels = zip(*batch)
new_images = []
imgH = self.imgH
imgW = self.imgW
if self.keep_ratio:
ratios = []
for img in images:
nheight = self.imgH
nwidth = self.imgW
if nheight is None:
nheight = img.shape[0]
if nwidth is None:
nwidth = int(
np.random.uniform(.8, 1.2) * img.shape[1] * nheight /
img.shape[0])
img = image_resize(img, height=nheight - 16, width=nwidth)
cv2.imwrite(
'/HOME/pondenka/manuel/CycleGANRD/HTR_ctc/data/test.png',
img)
img = img[np.newaxis, ...]
# img = centered(img, (nheight, int(1.2 * nwidth) + 32))
#img = Variable(torch.Tensor(img).float().unsqueeze(0))
new_images.append(img)
return Variable(torch.Tensor(new_images)), labels
def __getitem__(self, index):
img = self.data[index][0]
#img = cv2.Canny((255 * img).astype(np.uint8),100,200) / 255.0
transcr = " " + self.data[index][1] + " "
#transcr = self.data[index][1]
fheight, fwidth = self.fixed_size[0], self.fixed_size[1]
if self.set == 'train':
# random resize at training !!!
nwidth = int(np.random.uniform(.75, 1.25) * img.shape[1])
nheight = int(
(np.random.uniform(.9, 1.1) * img.shape[0] / img.shape[1]) *
nwidth)
else:
nheight, nwidth = img.shape[0], img.shape[1]
#nheight, nwidth = img.shape[0], img.shape[1]
#nheight, nwidth = int(.85 * img.shape[0]), int(.85 * img.shape[1])
# small pads!!
nheight, nwidth = max(4, min(fheight - 32,
nheight)), max(8,
min(fwidth - 64, nwidth))
#nheight, nwidth = fheight-32, fwidth-64
img = image_resize(img,
height=int(1.0 * nheight),
width=int(1.0 * nwidth))
img = centered(img, (fheight, fwidth), border_value=0.0)
if self.transforms is not None:
for tr in self.transforms:
if np.random.rand() < .5:
img = tr(img)
# pad with zeroes
#img = centered(img, (fheight, fwidth), np.random.uniform(.2, .8, size=2), border_value=0.0)
img = torch.Tensor(img).float().unsqueeze(0)
#if self.transforms is not None:
# for tr in self.transforms:
# if np.random.rand() < .5:
# img = tr(img.unsqueeze(0))[0]
return img, transcr
def __getitem__(self, index):
img = self.data[index][0].squeeze()
transcr = self.data[index][1]
# add augmentation
# ?? elastic, morphology, resizing
nheight = self.fixed_size[0]
nwidth = self.fixed_size[1]
if nheight is None:
nheight = img.shape[0]
if nwidth is None:
nwidth = int(np.random.uniform(.8, 1.2) * img.shape[1] * nheight / img.shape[0])
#augmentation
noise = np.random.uniform(.1, 0.25) * self.augment_factor
blur = np.random.uniform(.5, 2.0) * self.augment_factor
#img = image_resize(img, height=2000, width=(int(2000/nheight)*nwidth))
img = rotate(img, angle=np.random.random_integers(-2, 2), mode='constant', cval= 1, resize=True) # rotating
img = random_noise(img, var=noise ** 2) # adding noise
img = gaussian(img, sigma=blur, multichannel=True) # blurring
#end augmentation
# img = img.resize(int((nheight-16)/2), int(nwidth/2), PIL.Image.NEAREST)
#
# img = img.resize((nheight-16, nwidth), PIL.Image.NEAREST)
if self.set == 'test':
img = image_resize(img, height=nheight-16, width=nwidth)
# binarization
#img = img[:, :, np.newaxis]
#_, img = cv2.threshold(img, 0, 255, cv2.THRESH_OTSU)
#img = torch.Tensor(img).float().unsqueeze(0)
return img, transcr
def __getitem__(self, index):
img = self.data[index][0]
transcr = self.data[index][1]
# add augmentation
# ?? elastic, morphology, resizing
if (self.set == 'test2'):
nheight = self.fixed_size[0]
nwidth = self.fixed_size[1]
if nheight is None:
nheight = img.shape[0]
if nwidth is None:
nwidth = int(img.shape[1] * nheight / img.shape[0])
img = image_resize(img, height=nheight - 16, width=nwidth)
# img = centered(img, (nheight, int(1.2 * nwidth) + 32))
#img = torch.Tensor(img).float()
return img, transcr
def main_loader(set, level):
info = gather_iam_info(set, level)
data = []
for i, (img_path, transcr) in enumerate(info):
if i % 1000 == 0:
print('imgs: [{}/{} ({:.0f}%)]'.format(i, len(info),
100. * i / len(info)))
try:
img = img_io.imread(img_path + '.png')
img = 1 - img.astype(np.float32) / 255.0
img = image_resize(img, height=img.shape[0] // 2)
except:
print('main_loader error: Could not load {}'.format(img_path))
data += [(img, transcr.replace("|", " "))]
return data
def __getitem__(self, index):
img = self.data[index][0]
transcr = self.data[index][1]
# add augmentation
# ?? elastic, morphology, resizing
nheight = self.fixed_size[0]
nwidth = self.fixed_size[1]
if nheight is None:
nheight = img.shape[0]
if nwidth is None:
nwidth = int(np.random.uniform(.8, 1.2) * img.shape[1] * nheight / img.shape[0])
img = image_resize(img, height=nheight-16, width=nwidth)
img = centered(img, (nheight, int(1.2 * nwidth) + 32))
img = torch.Tensor(img).float().unsqueeze(0)
return img, transcr