def __getitem__(self, idx):
img_path, gt = self.lines[idx]
img = cv2.imread(img_path, 0)
if img is None:
return None
if img.shape[0] != self.img_height:
if img.shape[0] < self.img_height and not self.warning:
self.warning = True
print("WARNING: upsampling image to fit size")
percent = float(self.img_height) / img.shape[0]
img = cv2.resize(img, (0, 0),
fx=percent,
fy=percent,
interpolation=cv2.INTER_CUBIC)
if img is None:
return None
if self.augmentation:
img = augmentation.apply_random_color_rotation(img)
img = augmentation.apply_tensmeyer_brightness(img)
img = grid_distortion.warp_image(img)
img = img.astype(np.float32)
img = img / 128.0 - 1.0
img = img[..., None]
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
return {"line_img": img, "gt": gt, "gt_label": gt_label}
def interpolate(model, style1, style2, text, char_to_idx, gpu, step=0.05):
if type(style1) is tuple:
batch_size = style1[0].size(0)
else:
batch_size = style1.size(0)
label = string_utils.str2label_single(text, char_to_idx)
label = torch.from_numpy(label.astype(np.int32))[:, None].expand(
-1, batch_size).to(gpu).long()
label_len = torch.IntTensor(batch_size).fill_(len(text))
results = []
styles = []
for alpha in np.arange(0, 1.0, step):
if type(style1) is tuple:
style = (style2[0] * alpha + (1 - alpha) * style1[0],
style2[1] * alpha + (1 - alpha) * style1[1],
style2[2] * alpha + (1 - alpha) * style1[2])
else:
style = style2 * alpha + (1 - alpha) * style1
gen = model(label, label_len, style, flat=True)
results.append(gen)
if type(style) is tuple:
styles.append((style[0].cpu().detach(), style[1].cpu().detach(),
style[2].cpu().detach()))
else:
styles.append(style.cpu().detach())
return results, styles
def generate(model, style, text, char_to_idx, gpu):
#print(style)
batch_size = 1 #style.size(0)
label = string_utils.str2label_single(text, char_to_idx)
label = torch.from_numpy(label.astype(np.int32))[:, None].expand(
-1, batch_size).to(gpu).long()
label_len = torch.IntTensor(batch_size).fill_(label.size(0))
results = []
styles = []
return model(label, label_len, style, flat=True)
def __getitem__(self, idx):
ids_idx, line_idx = self.detailed_ids[idx]
gt_json_path, img_path = self.ids[ids_idx]
gt_json = safe_load.json_state(gt_json_path)
if gt_json is None:
return None
if 'hw_path' not in gt_json[line_idx]:
return None
hw_path = gt_json[line_idx]['hw_path']
hw_path = hw_path.split("/")[-1:]
hw_path = "/".join(hw_path)
hw_folder = os.path.dirname(gt_json_path)
img = cv2.imread(os.path.join(hw_folder, hw_path))
if img is None:
return None
if img.shape[0] != self.img_height:
if img.shape[0] < self.img_height and not self.warning:
self.warning = True
print "WARNING: upsampling image to fit size"
percent = float(self.img_height) / img.shape[0]
img = cv2.resize(img, (0,0), fx=percent, fy=percent, interpolation = cv2.INTER_CUBIC)
if img is None:
return None
if self.augmentation:
img = augmentation.apply_random_color_rotation(img)
img = augmentation.apply_tensmeyer_brightness(img)
img = grid_distortion.warp_image(img)
img = img.astype(np.float32)
img = img / 128.0 - 1.0
gt = gt_json[line_idx]['gt']
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
return {
"line_img": img,
"gt": gt,
"gt_label": gt_label
}
def sample(self):
#ri = np.random.choice(self.num_styles,[self.gen_batch_size,2],replace=False)
#mix = np.random.random(self.gen_batch_size)
#if self.extrapolate>0:
# mix = (2*self.extrapolate+1)*mix - self.extrapolate
#style = self.styles[ri[:,0]]*mix + self.styles[ri[:,1]]*(1-mix)
if self.random_style:
style = torch.FloatTensor(self.gen_batch_size,
self.style_dim).normal_()
else:
authors = np.random.choice(self.num_authors,
[self.gen_batch_size, 2],
replace=True)
mix = np.random.random(self.gen_batch_size)
if self.extrapolate > 0:
mix = (2 * self.extrapolate + 1) * mix - self.extrapolate
style = []
for b in range(self.gen_batch_size):
style0_i = np.random.choice(len(self.styles[authors[b, 0]]))
style1_i = np.random.choice(len(self.styles[authors[b, 1]]))
style0 = self.styles[authors[b, 0]][style0_i]
style1 = self.styles[authors[b, 1]][style1_i]
style.append(style0 * mix[b] + style1 * (1 - mix[b]))
style = np.stack(style, axis=0)
style = torch.from_numpy(style).float()
all_labels = []
label_lengths = []
gt = []
for i in range(self.gen_batch_size):
idx = np.random.randint(0, len(self.text))
text = self.text[idx]
gt.append(text)
l = string_utils.str2label_single(text, self.char_to_idx)
all_labels.append(l)
label_lengths.append(len(l))
#all_labels = np.concatenate(all_labels)
label_lengths = torch.IntTensor(label_lengths)
max_len = label_lengths.max()
all_labels = [
np.pad(l, ((0, max_len - l.shape[0]), ), 'constant')
for l in all_labels
]
all_labels = np.stack(all_labels, axis=1)
label = torch.from_numpy(all_labels.astype(np.int32))
return style, label, label_lengths, gt
def __getitem__(self, idx):
img_path = os.path.join(self.directory, '{}.png'.format(idx))
img = cv2.imread(img_path, 0)
if len(img.shape) == 2:
img = img[..., None]
if self.augmentation is not None:
#img = augmentation.apply_random_color_rotation(img)
if 'brightness' in self.augmentation:
img = augmentation.apply_tensmeyer_brightness(img)
if 'warp' in self.augmentation:
img = grid_distortion.warp_image(img)
if len(img.shape) == 2:
img = img[..., None]
img = img.astype(np.float32)
img = 1.0 - img / 128.0
gt = self.labels[idx]
if gt is None:
#metadata = pyexiv2.ImageMetadata(img_path)
#metadata.read()
#metadata = piexif.load(img_path)
#if 'gt' in metadata:
# gt = metadata['gt']
#else:
print('Error unknown label for image: {}'.format(img_path))
return self.__getitem__((idx + 7) % self.set_size)
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
return {
"image": img,
"gt": gt,
"gt_label": gt_label,
#"author": author
}
def loss(preds, non_hw_sol, hw_sol, gt_lines, char_to_idx, criterion):
label_lengths = []
all_labels = []
for gt_str in gt_lines:
l = string_utils.str2label_single(gt_str, char_to_idx)
all_labels.append(l)
label_lengths.append(len(l))
all_labels = np.concatenate(all_labels)
label_lengths = np.array(label_lengths)
labels = torch.from_numpy(all_labels.astype(np.int32))
label_lengths = torch.from_numpy(label_lengths.astype(np.int32))
labels = Variable(labels, requires_grad=False)
label_lengths = Variable(label_lengths, requires_grad=False)
batch_size = preds.size(0)
preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
ctc_loss = 1e-2 * criterion(preds.cpu(), labels, preds_size, label_lengths)
log_one_minus_confidences = torch.log(1.0 - non_hw_sol[:, :, 0] + 1e-10)
log_confidences = torch.log(hw_sol[:, :, 0] + 1e-10)
selected_confidence = log_confidences.sum()
not_selected_confidence = log_one_minus_confidences.sum()
confidence_loss = -selected_confidence - not_selected_confidence
# print " - - - - Losses - - - - "
# print ctc_loss.data[0]
# print selected_confidence.data[0], log_confidences.size()
# print not_selected_confidence.data[0], log_one_minus_confidences.size()
# print ""
return ctc_loss + confidence_loss.cpu()
def __getitem__(self, idx):
syn_gen = random.choice(self.synthetic)
syn_img, gt, f_index = syn_gen.getSample()
w = round(syn_img.shape[1] * float(self.img_height) / syn_img.shape[0])
img = cv2.resize(syn_img, (w, self.img_height),
interpolation=cv2.INTER_CUBIC)
#syn_img=torch.from_numpy(syn_img).float()[None,None,...]
#if self.cuda:
# syn_img = syn_img.cuda()
#img = self.generator(syn_img)
#if img.shape[0] != self.img_height:
# if img.shape[0] < self.img_height and not self.warning:
# self.warning = True
# print "WARNING: upsampling image to fit size"
# percent = float(self.img_height) / img.shape[0]
# img = cv2.resize(img, (0,0), fx=percent, fy=percent, interpolation = cv2.INTER_CUBIC)
#img = (img.cpu().numpy()[0,0,...]+1)*128
img *= 255
img = img.astype(np.uint8)
if self.augmentation:
#img = augmentation.apply_random_color_rotation(img)
#img = augmentation.apply_tensmeyer_brightness(img) done in sythetic text generator
if random.random() < 0.5:
img = grid_distortion.warp_image(img)
img = img.astype(np.float32)
img = (img / 128.0 - 1.0)[..., None]
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
return {"line_img": img, "gt": gt, "gt_label": gt_label}
def __getitem__(self, idx):
inst = self.lineIndex[idx]
author = inst[0]
words = inst[1]
batch = []
for word in words:
if word >= len(self.w_authors[author]):
word = (word + 37) % len(self.w_authors[author])
img_path, lb, gt, id = self.w_authors[author][word]
img = cv2.imread(img_path,
0)[lb[0]:lb[1],
lb[2]:lb[3]] #read as grayscale, crop word
if img.shape[0] == 0 or img.shape[1] == 0:
return self.__getitem__((idx + 1) % self.__len__())
if img is None:
return None
if img.shape[0] != self.img_height:
if img.shape[0] < self.img_height and not self.warning:
self.warning = True
print("WARNING: upsampling image to fit size")
percent = float(self.img_height) / img.shape[0]
img = cv2.resize(img, (0, 0),
fx=percent,
fy=percent,
interpolation=cv2.INTER_CUBIC)
if img is None:
return None
if len(img.shape) == 2:
img = img[..., None]
if self.augmentation is not None:
#img = augmentation.apply_random_color_rotation(img)
img = augmentation.apply_tensmeyer_brightness(img)
img = grid_distortion.warp_image(img)
img = img.astype(np.float32)
img = 1.0 - img / 128.0
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
if self.styles:
style_i = self.npr.choice(len(self.styles[author][id]))
style = self.styles[author][id][style_i]
else:
style = None
batch.append({
"image": img,
"gt": gt,
"style": style,
"gt_label": gt_label,
"name": '{}_{}'.format(author, word),
"author": author
})
#batch = [b for b in batch if b is not None]
#These all should be the same size or error
assert len(set([b['image'].shape[0] for b in batch])) == 1
assert len(set([b['image'].shape[2] for b in batch])) == 1
dim0 = batch[0]['image'].shape[0]
dim1 = max([b['image'].shape[1] for b in batch])
dim2 = batch[0]['image'].shape[2]
all_labels = []
label_lengths = []
input_batch = np.full((len(batch), dim0, dim1, dim2),
PADDING_CONSTANT).astype(np.float32)
for i in range(len(batch)):
b_img = batch[i]['image']
#toPad = (dim1-b_img.shape[1])
input_batch[i, :, 0:b_img.shape[1], :] = b_img
l = batch[i]['gt_label']
all_labels.append(l)
label_lengths.append(len(l))
#all_labels = np.concatenate(all_labels)
label_lengths = torch.IntTensor(label_lengths)
max_len = label_lengths.max()
all_labels = [
np.pad(l, ((0, max_len - l.shape[0]), ), 'constant')
for l in all_labels
]
all_labels = np.stack(all_labels, axis=1)
images = input_batch.transpose([0, 3, 1, 2])
images = torch.from_numpy(images)
labels = torch.from_numpy(all_labels.astype(np.int32))
#label_lengths = torch.from_numpy(label_lengths.astype(np.int32))
if batch[0]['style'] is not None:
styles = np.stack([b['style'] for b in batch], axis=0)
styles = torch.from_numpy(styles).float()
else:
styles = None
return {
"image": images,
"mask": makeMask(images, self.mask_post),
"label": labels,
"style": styles,
"label_lengths": label_lengths,
"gt": [b['gt'] for b in batch],
"name": [b['name'] for b in batch],
"author": [b['author'] for b in batch]
}
def __getitem__(self, idx):
if type( self.augmentation) is str and 'affine' in self.augmentation:
strech = (self.max_strech*2)*np.random.random() - self.max_strech +1
#self.max_rot_rad = self.max_rot_deg/180 * np.pi
skew = (self.max_rot_rad*2)*np.random.random() - self.max_rot_rad
if self.include_stroke_aug:
thickness_change= np.random.randint(-4,5)
fg_shade = np.random.random()*0.25 + 0.75
bg_shade = np.random.random()*0.2
blur_size = np.random.randint(2,4)
noise_sigma = np.random.random()*0.02
batch=[]
if self.triplet=='hard':
authors = random.sample(self.authors.keys(),self.triplet_author_size)
alines=[]
for author in authors:
if len(self.authors[author])>=self.triplet_sample_size*self.batch_size:
lines = random.sample(range(len(self.authors[author])),self.triplet_sample_size*self.batch_size)
else:
lines = list(range(len(self.authors[author])))
random.shuffle(lines)
dif = self.triplet_sample_size*self.batch_size-len(self.authors[author])
lines += lines[:dif]
alines += [(author,l) for l in lines]
else:
inst = self.lineIndex[idx]
author=inst[0]
lines=inst[1]
alines = [(author,l) for l in lines]
used_lines = set(lines)
if self.triplet:
if len(self.authors[author])<=2*self.batch_size:
for l in range(len(self.authors[author])):
if l not in used_lines:
alines.append((author,l))
if len(alines)<2*self.batch_size:
dif = 2*self.batch_size - len(alines)
for i in range(dif):
alines.append(alines[self.batch_size+i])
else:
unused_lines = set(range(len(self.authors[author])))-used_lines
for i in range(self.batch_size):
l = random.select(unused_lines)
unused_lines.remove(l)
alines.append((author,l))
other_authors = set(range(len(self.authors)))
other_authors.remove(author)
author = random.select(other_authors)
unused_lines = set(range(len(self.authors[author])))-used_lines
for i in range(self.batch_size):
l = random.select(unused_lines)
unused_lines.remove(l)
alines.append((author,l))
images=[]
for author,line in alines:
if line>=len(self.authors[author]):
line = (line+37)%len(self.authors[author])
img_path, lb, gt = self.authors[author][line]
img_path = os.path.join(self.dirPath,'images_gray',img_path)
if self.no_spaces:
gt = gt.replace(' ','')
if type(self.augmentation) is str and 'normalization' in self.augmentation and self.normalized_dir is not None and os.path.exists(os.path.join(self.normalized_dir,'{}_{}.png'.format(author,line))):
img = cv2.imread(os.path.join(self.normalized_dir,'{}_{}.png'.format(author,line)),0)
readNorm=True
else:
img = cv2.imread(img_path,0)
if img is None:
print('Error, could not read image: {}'.format(img_path))
return None
lb[0] = max(lb[0],0)
lb[2] = max(lb[2],0)
lb[1] = min(lb[1],img.shape[0])
lb[3] = min(lb[3],img.shape[1])
img = img[lb[0]:lb[1],lb[2]:lb[3]] #read as grayscale, crop line
readNorm=False
if img.shape[0] != self.img_height:
if img.shape[0] < self.img_height and not self.warning:
self.warning = True
print("WARNING: upsampling image to fit size")
percent = float(self.img_height) / img.shape[0]
if img.shape[1]*percent > self.max_width:
percent = self.max_width/img.shape[1]
img = cv2.resize(img, (0,0), fx=percent, fy=percent, interpolation = cv2.INTER_CUBIC)
if img.shape[0]<self.img_height:
diff = self.img_height-img.shape[0]
img = np.pad(img,((diff//2,diff//2+diff%2),(0,0)),'constant',constant_values=255)
elif img.shape[1]> self.max_width:
percent = self.max_width/img.shape[1]
img = cv2.resize(img, (0,0), fx=percent, fy=percent, interpolation = cv2.INTER_CUBIC)
if img.shape[0]<self.img_height:
diff = self.img_height-img.shape[0]
img = np.pad(img,((diff//2,diff//2+diff%2),(0,0)),'constant',constant_values=255)
if self.augmentation=='affine':
if img.shape[1]*strech > self.max_width:
strech = self.max_width/img.shape[1]
images.append( (line,gt,img,author) )
#we split the processing here so that strech will be adjusted for longest image in author batch
for line,gt,img,author in images:
if self.fg_masks_dir is not None:
fg_path = os.path.join(self.fg_masks_dir,'{}_{}.png'.format(author,line))
fg_mask = cv2.imread(fg_path,0)
fg_mask = fg_mask/255
if fg_mask.shape!=img[:,:].shape:
print('Error, fg_mask ({}, {}) not the same size as image ({})'.format(fg_path,fg_mask.shape,img[:,:,0].shape))
th,fg_mask = cv2.threshold(img,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
fg_mask = 255-fg_mask
ele = cv2.getStructuringElement( cv2.MORPH_ELLIPSE, (9,9) )
fg_mask = cv2.dilate(fg_mask,ele)
fg_mask = fg_mask/255
else:
fg_mask=None
if type(self.augmentation) is str and 'normalization' in self.augmentation and not readNorm:
img = normalize_line.deskew(img)
img = normalize_line.skeletonize(img)
if self.normalized_dir is not None:
cv2.imwrite(os.path.join(self.normalized_dir,'{}_{}.png'.format(author,line)),img)
if type(self.augmentation) is str and 'affine' in self.augmentation:
img,fg_mask = augmentation.affine_trans(img,fg_mask,skew,strech)
elif self.augmentation is not None and (type(self.augmentation) is not None or 'warp' in self.augmentation):
#img = augmentation.apply_random_color_rotation(img)
img = augmentation.apply_tensmeyer_brightness(img)
img = grid_distortion.warp_image(img)
assert(fg_mask is None)
if self.include_stroke_aug:
new_img = augmentation.change_thickness(img,thickness_change,fg_shade,bg_shade,blur_size,noise_sigma)
if len(new_img.shape)==2:
new_img = new_img[...,None]
new_img = new_img*2 -1.0
if len(img.shape)==2:
img = img[...,None]
img = img.astype(np.float32)
if self.remove_bg:
img = 1.0 - img / 256.0
#kernel = torch.FloatTensor(7,7).fill_(1/49)
#blurred_mask = F.conv2d(fg_mask,kernel,padding=3)
blurred_mask = cv2.blur(fg_mask,(7,7))
img *= blurred_mask[...,None]
img = 2*img -1
else:
img = 1.0 - img / 128.0
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
if self.styles:
style_i = self.npr.choice(len(self.styles[author][id]))
style = self.styles[author][id][style_i]
else:
style=None
name = '{}_{}'.format(author,line)
if self.identity_spaced:
spaced_label = gt_label[:,None].astype(np.long)
else:
spaced_label = None if self.spaced_by_name is None else self.spaced_by_name[name]
if spaced_label is not None:
assert(spaced_label.shape[1]==1)
toAppend= {
"image": img,
"gt": gt,
"style": style,
"gt_label": gt_label,
"spaced_label": spaced_label,
"name": name,
"center": self.center,
"author": author,
"author_idx": self.author_list.index(author)
}
if self.fg_masks_dir is not None:
toAppend['fg_mask'] = fg_mask
if self.include_stroke_aug:
toAppend['changed_image'] = new_img
batch.append(toAppend)
#batch = [b for b in batch if b is not None]
#These all should be the same size or error
assert len(set([b['image'].shape[0] for b in batch])) == 1
assert len(set([b['image'].shape[2] for b in batch])) == 1
dim0 = batch[0]['image'].shape[0]
dim1 = max([b['image'].shape[1] for b in batch])
dim2 = batch[0]['image'].shape[2]
all_labels = []
label_lengths = []
if self.spaced_by_name is not None or self.identity_spaced:
spaced_labels = []
else:
spaced_labels = None
max_spaced_len=0
input_batch = np.full((len(batch), dim0, dim1, dim2), PADDING_CONSTANT).astype(np.float32)
if self.fg_masks_dir is not None:
fg_masks = np.full((len(batch), dim0, dim1, 1), 0).astype(np.float32)
if self.include_stroke_aug:
changed_batch = np.full((len(batch), dim0, dim1, dim2), PADDING_CONSTANT).astype(np.float32)
for i in range(len(batch)):
b_img = batch[i]['image']
toPad = (dim1-b_img.shape[1])
if 'center' in batch[0] and batch[0]['center']:
toPad //=2
else:
toPad = 0
input_batch[i,:,toPad:toPad+b_img.shape[1],:] = b_img
if self.fg_masks_dir is not None:
fg_masks[i,:,toPad:toPad+b_img.shape[1],0] = batch[i]['fg_mask']
if self.include_stroke_aug:
changed_batch[i,:,toPad:toPad+b_img.shape[1],:] = batch[i]['changed_image']
l = batch[i]['gt_label']
all_labels.append(l)
label_lengths.append(len(l))
if spaced_labels is not None:
sl = batch[i]['spaced_label']
spaced_labels.append(sl)
max_spaced_len = max(max_spaced_len,sl.shape[0])
#all_labels = np.concatenate(all_labels)
label_lengths = torch.IntTensor(label_lengths)
max_len = label_lengths.max()
all_labels = [np.pad(l,((0,max_len-l.shape[0]),),'constant') for l in all_labels]
all_labels = np.stack(all_labels,axis=1)
if self.spaced_by_name is not None or self.identity_spaced:
spaced_labels = [np.pad(l,((0,max_spaced_len-l.shape[0]),(0,0)),'constant') for l in spaced_labels]
ddd = spaced_labels
spaced_labels = np.concatenate(spaced_labels,axis=1)
spaced_labels = torch.from_numpy(spaced_labels)
assert(spaced_labels.size(1) == len(batch))
images = input_batch.transpose([0,3,1,2])
images = torch.from_numpy(images)
labels = torch.from_numpy(all_labels.astype(np.int32))
#label_lengths = torch.from_numpy(label_lengths.astype(np.int32))
if self.fg_masks_dir is not None:
fg_masks = fg_masks.transpose([0,3,1,2])
fg_masks = torch.from_numpy(fg_masks)
if batch[0]['style'] is not None:
styles = np.stack([b['style'] for b in batch], axis=0)
styles = torch.from_numpy(styles).float()
else:
styles=None
mask, top_and_bottom, center_line = makeMask(images,self.mask_post, self.mask_random)
##DEBUG
#for i in range(5):
# mask2, top_and_bottom2 = makeMask(images,self.mask_post, self.mask_random)
# #extra_masks.append(mask2)
# mask2 = ((mask2[0,0]+1)/2).numpy().astype(np.uint8)*255
# cv2.imshow('mask{}'.format(i),mask2)
#mask = ((mask[0,0]+1)/2).numpy().astype(np.uint8)*255
#cv2.imshow('mask'.format(i),mask)
#cv2.waitKey()
toRet= {
"image": images,
"mask": mask,
"top_and_bottom": top_and_bottom,
"center_line": center_line,
"label": labels,
"style": styles,
"label_lengths": label_lengths,
"gt": [b['gt'] for b in batch],
"spaced_label": spaced_labels,
"name": [b['name'] for b in batch],
"author": [b['author'] for b in batch],
"author_idx": [b['author_idx'] for b in batch],
}
if self.fg_masks_dir is not None:
toRet['fg_mask'] = fg_masks
if self.include_stroke_aug:
changed_images = changed_batch.transpose([0,3,1,2])
changed_images = torch.from_numpy(changed_images)
toRet['changed_image']=changed_images
return toRet
def __getitem__(self, idx):
author, line = self.lineIndex[idx]
img_path, lb, gt = self.authors[author][line]
if self.add_spaces:
gt = ' ' + gt + ' '
if type(
self.augmentation
) is str and 'normalization' in self.augmentation and self.normalized_dir is not None and os.path.exists(
os.path.join(self.normalized_dir, '{}_{}.png'.format(
author, line))):
img = cv2.imread(
os.path.join(self.normalized_dir,
'{}_{}.png'.format(author, line)), 0)
readNorm = True
else:
img = cv2.imread(img_path,
0)[lb[0]:lb[1],
lb[2]:lb[3]] #read as grayscale, crop line
readNorm = False
if img is None:
return None
if img.shape[0] != self.img_height:
if img.shape[0] < self.img_height and not self.warning:
self.warning = True
print("WARNING: upsampling image to fit size")
percent = float(self.img_height) / img.shape[0]
img = cv2.resize(img, (0, 0),
fx=percent,
fy=percent,
interpolation=cv2.INTER_CUBIC)
if img is None:
return None
if len(img.shape) == 2:
img = img[..., None]
if type(
self.augmentation
) is str and 'normalization' in self.augmentation and not readNorm:
img = normalize_line.deskew(img)
img = normalize_line.skeletonize(img)
if self.normalized_dir is not None:
cv2.imwrite(
os.path.join(self.normalized_dir,
'{}_{}.png'.format(author, line)), img)
elif self.augmentation is not None and (type(
self.augmentation) is not str or 'warp' in self.augmentation):
#img = augmentation.apply_random_color_rotation(img)
if type(self.augmentation) is str and "low" in self.augmentation:
if random.random() > 0.1:
img = augmentation.apply_tensmeyer_brightness(img)
if random.random() > 0.01:
img = grid_distortion.warp_image(img,
w_mesh_std=0.7,
h_mesh_std=0.7)
else:
img = augmentation.apply_tensmeyer_brightness(img)
img = grid_distortion.warp_image(img)
if len(img.shape) == 2:
img = img[..., None]
img = img.astype(np.float32)
img = 1.0 - img / 128.0
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
return {
"image": img,
"gt": gt,
"gt_label": gt_label,
"name": '{}_{}'.format(author, line),
"center": self.center,
"author": author
}
def __getitem__(self, idx):
if self.train and self.refresh_self:
if self.used_instances >= self.set_size:
self.refresh_data(None, None, logged=False)
self.used_instances = 0
self.used_instances += self.batch_size
if self.augmentation is not None and 'affine' in self.augmentation:
strech = (self.max_strech *
2) * np.random.random() - self.max_strech + 1
skew = (self.max_rot_rad *
2) * np.random.random() - self.max_rot_rad
if self.include_stroke_aug:
thickness_change = np.random.randint(-4, 5)
fg_shade = np.random.random() * 0.25 + 0.75
bg_shade = np.random.random() * 0.2
blur_size = np.random.randint(2, 4)
noise_sigma = np.random.random() * 0.02
batch = []
for b in range(self.batch_size):
img_path = os.path.join(self.directory, '{}.png'.format(idx + b))
img = cv2.imread(img_path, 0)
if img is None:
print('Error, could not read {}'.format(img_path))
return self[(idx + 1) % len(self)]
if self.augmentation == 'affine':
if img.shape[1] * strech > self.max_width:
strech = self.max_width / img.shape[1]
if img.shape[1] > self.max_width:
percent = float(self.max_width) / img.shape[1]
img = cv2.resize(img, (0, 0),
fx=percent,
fy=1,
interpolation=cv2.INTER_CUBIC)
if img.shape[1] > self.clip_width:
img = img[:, :self.clip_width]
if self.use_fg_mask:
th, fg_mask = cv2.threshold(
img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
fg_mask = 255 - fg_mask
ele = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (9, 9))
fg_mask = cv2.dilate(fg_mask, ele)
fg_mask = fg_mask / 255
else:
fg_mask = None
if len(img.shape) == 2:
img = img[..., None]
if self.augmentation is not None:
#img = augmentation.apply_random_color_rotation(img)
if 'affine' in self.augmentation:
img, fg_mask = augmentation.affine_trans(
img, fg_mask, skew, strech)
if 'brightness' in self.augmentation:
img = augmentation.apply_tensmeyer_brightness(img)
assert (fg_mask is None)
if 'warp' in self.augmentation and random.random(
) < self.warp_freq:
try:
img = grid_distortion.warp_image(img)
except cv2.error as e:
print(e)
print(img.shape)
assert (fg_mask is None)
if 'invert' in self.augmentation and random.random() < 0.25:
img = 1 - img
if self.include_stroke_aug:
new_img = augmentation.change_thickness(
img, thickness_change, fg_shade, bg_shade, blur_size,
noise_sigma)
new_img = new_img * 2 - 1.0
if len(img.shape) == 2:
img = img[..., None]
img = img.astype(np.float32)
img = 1.0 - img / 128.0
if self.train:
gt = self.labels[idx]
else:
with open(self.gt_filename) as f:
for i in range(0, idx + 1):
gt = f.readline()
gt = gt.strip()
if gt is None:
#metadata = pyexiv2.ImageMetadata(img_path)
#metadata.read()
#metadata = piexif.load(img_path)
#if 'gt' in metadata:
# gt = metadata['gt']
#else:
print('Error unknown label for image: {}'.format(img_path))
return self.__getitem__((idx + 7) % self.set_size)
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
font_idx = '?'
toRet = {
"image": img,
"gt": gt,
"gt_label": gt_label,
"author": font_idx,
"name": '{}_{}'.format(idx + b, font_idx),
"style": None,
"spaced_label": None
}
if self.use_fg_mask:
toRet['fg_mask'] = fg_mask
if self.include_stroke_aug:
toRet['changed_image'] = new_img
batch.append(toRet)
dim0 = batch[0]['image'].shape[0]
dim1 = max([b['image'].shape[1] for b in batch])
dim2 = batch[0]['image'].shape[2]
all_labels = []
label_lengths = []
if self.spaced_by_name is not None:
spaced_labels = []
else:
spaced_labels = None
max_spaced_len = 0
input_batch = np.full((len(batch), dim0, dim1, dim2),
PADDING_CONSTANT).astype(np.float32)
if self.use_fg_mask:
fg_masks = np.full((len(batch), dim0, dim1, 1),
0).astype(np.float32)
if self.include_stroke_aug:
changed_batch = np.full((len(batch), dim0, dim1, dim2),
PADDING_CONSTANT).astype(np.float32)
for i in range(len(batch)):
b_img = batch[i]['image']
toPad = (dim1 - b_img.shape[1])
if 'center' in batch[0] and batch[0]['center']:
toPad //= 2
else:
toPad = 0
input_batch[i, :, toPad:toPad + b_img.shape[1], :] = b_img
if self.use_fg_mask:
fg_masks[i, :, toPad:toPad + b_img.shape[1],
0] = batch[i]['fg_mask']
if self.include_stroke_aug:
changed_batch[i, :, toPad:toPad + b_img.shape[1],
0] = batch[i]['changed_image']
l = batch[i]['gt_label']
all_labels.append(l)
label_lengths.append(len(l))
if spaced_labels is not None:
sl = batch[i]['spaced_label']
spaced_labels.append(sl)
max_spaced_len = max(max_spaced_len, sl.shape[0])
#all_labels = np.concatenate(all_labels)
label_lengths = torch.IntTensor(label_lengths)
max_len = label_lengths.max()
all_labels = [
np.pad(l, ((0, max_len - l.shape[0]), ), 'constant')
for l in all_labels
]
all_labels = np.stack(all_labels, axis=1)
if self.spaced_by_name is not None:
spaced_labels = [
np.pad(l, ((0, max_spaced_len - l.shape[0]), (0, 0)),
'constant') for l in spaced_labels
]
ddd = spaced_labels
spaced_labels = np.concatenate(spaced_labels, axis=1)
spaced_labels = torch.from_numpy(spaced_labels)
assert (spaced_labels.size(1) == len(batch))
images = input_batch.transpose([0, 3, 1, 2])
images = torch.from_numpy(images)
labels = torch.from_numpy(all_labels.astype(np.int32))
#label_lengths = torch.from_numpy(label_lengths.astype(np.int32))
if self.use_fg_mask:
fg_masks = fg_masks.transpose([0, 3, 1, 2])
fg_masks = torch.from_numpy(fg_masks)
if batch[0]['style'] is not None:
styles = np.stack([b['style'] for b in batch], axis=0)
styles = torch.from_numpy(styles).float()
else:
styles = None
mask, top_and_bottom, center_line = makeMask(images, self.mask_post,
self.mask_random)
toRet = {
"image": images,
"mask": mask,
"top_and_bottom": top_and_bottom,
"center_line": center_line,
"label": labels,
"style": styles,
"label_lengths": label_lengths,
"gt": [b['gt'] for b in batch],
"spaced_label": spaced_labels,
"name": [b['name'] for b in batch],
"author": [b['author'] for b in batch],
}
if self.use_fg_mask:
toRet['fg_mask'] = fg_masks
if self.include_stroke_aug:
changed_images = changed_batch.transpose([0, 3, 1, 2])
changed_images = torch.from_numpy(changed_images)
toRet['changed_image'] = changed_images
return toRet
def __getitem__(self, idx):
inst = self.lineIndex[idx]
author = inst[0]
lines = inst[1]
batch = []
for line in lines:
if line >= len(self.authors[author]):
line = (line + 37) % len(self.authors[author])
img_path, gt, pad_above, pad_below = self.authors[author][line]
img = cv2.imread(img_path, 0) #read as grayscale
if img is None:
return None
if pad_above < 0:
img = img[-pad_above:, :]
pad_above = 0
if pad_below < 0:
img = img[:pad_below, :]
pad_below = 0
#if pad_above>0 or pad_below>0:
img = img = np.pad(img, ((pad_above, pad_below), (10, 10)),
'constant',
constant_values=255)
#we also pad a bit on the sides
#print('{}, {} {}'.format(img_path,pad_above,pad_below))
if img.shape[0] != self.img_height:
if img.shape[0] < self.img_height and not self.warning:
self.warning = True
print("WARNING: upsampling image to fit size")
percent = float(self.img_height) / img.shape[0]
if img.shape[1] * percent > self.max_width:
percent = self.max_width / img.shape[1]
img = cv2.resize(img, (0, 0),
fx=percent,
fy=percent,
interpolation=cv2.INTER_CUBIC)
if img.shape[0] < self.img_height:
diff = self.img_height - img.shape[0]
img = np.pad(img,
((diff // 2, diff // 2 + diff % 2), (0, 0)),
'constant',
constant_values=255)
if len(img.shape) == 2:
img = img[..., None]
if self.fg_masks_dir is not None:
fg_path = os.path.join(self.fg_masks_dir,
'{}_{}.png'.format(author, line))
fg_mask = cv2.imread(fg_path, 0)
fg_mask = fg_mask / 255
if fg_mask.shape != img[:, :, 0].shape:
print(
'Error, fg_mask ({}, {}) not the same size as image ({})'
.format(fg_path, fg_mask.shape, img[:, :, 0].shape))
th, fg_mask = cv2.threshold(
img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
fg_mask = 255 - fg_mask
ele = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (9, 9))
fg_mask = cv2.dilate(fg_mask, ele)
fg_mask = fg_mask / 255
if self.augmentation is not None:
#img = augmentation.apply_random_color_rotation(img)
img = augmentation.apply_tensmeyer_brightness(img)
img = grid_distortion.warp_image(img)
if len(img.shape) == 2:
img = img[..., None]
img = img.astype(np.float32)
img = 1.0 - img / 128.0
if len(gt) == 0:
return None
gt_label = string_utils.str2label_single(gt, self.char_to_idx)
if self.styles:
style_i = self.npr.choice(len(self.styles[author][id]))
style = self.styles[author][id][style_i]
else:
style = None
name = img_path[img_path.rfind('/') + 1:img_path.rfind('.')]
spaced_label = None if self.spaced_by_name is None else self.spaced_by_name[
img_path]
if spaced_label is not None:
assert (spaced_label.shape[1] == 1)
toAppend = {
"image": img,
"gt": gt,
"style": style,
"gt_label": gt_label,
"spaced_label": spaced_label,
"name": name,
"center": self.center,
"author": author
}
if self.fg_masks_dir is not None:
toAppend['fg_mask'] = fg_mask
batch.append(toAppend)
#batch = [b for b in batch if b is not None]
#These all should be the same size or error
assert len(set([b['image'].shape[0] for b in batch])) == 1
assert len(set([b['image'].shape[2] for b in batch])) == 1
dim0 = batch[0]['image'].shape[0]
dim1 = max([b['image'].shape[1] for b in batch])
dim2 = batch[0]['image'].shape[2]
all_labels = []
label_lengths = []
if self.spaced_by_name is not None:
spaced_labels = []
else:
spaced_labels = None
max_spaced_len = 0
input_batch = np.full((len(batch), dim0, dim1, dim2),
PADDING_CONSTANT).astype(np.float32)
if self.fg_masks_dir is not None:
fg_masks = np.full((len(batch), dim0, dim1, 1),
0).astype(np.float32)
for i in range(len(batch)):
b_img = batch[i]['image']
toPad = (dim1 - b_img.shape[1])
if 'center' in batch[0] and batch[0]['center']:
toPad //= 2
else:
toPad = 0
input_batch[i, :, toPad:toPad + b_img.shape[1], :] = b_img
if self.fg_masks_dir is not None:
fg_masks[i, :, toPad:toPad + b_img.shape[1],
0] = batch[i]['fg_mask']
l = batch[i]['gt_label']
all_labels.append(l)
label_lengths.append(len(l))
if spaced_labels is not None:
sl = batch[i]['spaced_label']
spaced_labels.append(sl)
max_spaced_len = max(max_spaced_len, sl.shape[0])
#all_labels = np.concatenate(all_labels)
label_lengths = torch.IntTensor(label_lengths)
max_len = label_lengths.max()
all_labels = [
np.pad(l, ((0, max_len - l.shape[0]), ), 'constant')
for l in all_labels
]
all_labels = np.stack(all_labels, axis=1)
if self.spaced_by_name is not None:
spaced_labels = [
np.pad(l, ((0, max_spaced_len - l.shape[0]), (0, 0)),
'constant') for l in spaced_labels
]
ddd = spaced_labels
spaced_labels = np.concatenate(spaced_labels, axis=1)
spaced_labels = torch.from_numpy(spaced_labels)
assert (spaced_labels.size(1) == len(batch))
images = input_batch.transpose([0, 3, 1, 2])
images = torch.from_numpy(images)
labels = torch.from_numpy(all_labels.astype(np.int32))
#label_lengths = torch.from_numpy(label_lengths.astype(np.int32))
if self.fg_masks_dir is not None:
fg_masks = fg_masks.transpose([0, 3, 1, 2])
fg_masks = torch.from_numpy(fg_masks)
if batch[0]['style'] is not None:
styles = np.stack([b['style'] for b in batch], axis=0)
styles = torch.from_numpy(styles).float()
else:
styles = None
mask, top_and_bottom, center_line = makeMask(images, self.mask_post,
self.mask_random)
##DEBUG
#for i in range(5):
# mask2, top_and_bottom2 = makeMask(images,self.mask_post, self.mask_random)
# #extra_masks.append(mask2)
# mask2 = ((mask2[0,0]+1)/2).numpy().astype(np.uint8)*255
# cv2.imshow('mask{}'.format(i),mask2)
#mask = ((mask[0,0]+1)/2).numpy().astype(np.uint8)*255
#cv2.imshow('mask'.format(i),mask)
#cv2.waitKey()
toRet = {
"image": images,
"mask": mask,
"top_and_bottom": top_and_bottom,
"center_line": center_line,
"label": labels,
"style": styles,
"label_lengths": label_lengths,
"gt": [b['gt'] for b in batch],
"spaced_label": spaced_labels,
"name": [b['name'] for b in batch],
"author": [b['author'] for b in batch],
}
if self.fg_masks_dir is not None:
toRet['fg_mask'] = fg_masks
return toRet
def getInstance(self):
all_labels = []
label_lengths = []
gt = []
for i in range(self.batch_size):
if self.words:
idx = np.random.randint(0, len(self.text))
text = self.text[idx]
if len(self.text) > self.max_len:
diff = len(self.text) - self.max_len
start = random.randint(0, diff)
text[start:start + self.max_len]
else:
length = random.randint(self.min_len, self.max_len)
idx = np.random.randint(0, len(self.text) - length)
if self.characterBalance:
startIdx = idx
flipped = False
goalChar = random.choice(self.chars)
while True:
text = self.text[idx:idx + length]
if goalChar in text:
break
idx += length
if idx >= len(self.text) - length:
flipped = True
idx = 0
if flipped and idx >= startIdx:
#this char is not in the text set, so we'll just add it somewhere random
r = random.randint(0, len(text))
text = text[:r] + goalChar + text[r + 1:]
break
else:
text = self.text[idx:idx + length]
assert (len(text) > 0)
if text == ' ':
text = self.text[idx + 1]
gt.append(text)
if self.char_to_idx is not None:
l = string_utils.str2label_single(text, self.char_to_idx)
all_labels.append(l)
label_lengths.append(len(l))
if self.char_to_idx is not None:
#all_labels = np.concatenate(all_labels)
label_lengths = torch.IntTensor(label_lengths)
max_len = label_lengths.max()
all_labels = [
np.pad(l, ((0, max_len - l.shape[0]), ), 'constant')
for l in all_labels
]
all_labels = np.stack(all_labels, axis=1)
return {
'label': torch.from_numpy(all_labels.astype(np.int32)),
'label_lengths': label_lengths,
'gt': gt,
'image': None
}
else:
return {'gt': gt, 'image': None}
