def forward(self, input, target, labels):
_assert_no_grad(target)
loss = torch.abs(input - target)
lptnums = [lptnum_of(label) for label in labels]
img_chars_l = []
for i, lptnum in enumerate(lptnums):
for j, _ in enumerate(loss[i, :, self.hb:self.he, :]):
img_chars_l.append(
self._get_img_chars(lptnum, loss[i, j,
self.hb:self.he, :]))
# test only
# for i in range(len(img_chars_l)):
# k=0
# for im, c in zip(img_chars_l[i][0],img_chars_l[i][1]):
# utils.save_image(im,"{}_{}_{}.png".format(i,c,k),normalize=True, range=(0,255))
# k += 1
# raise SystemExit
# Split img_char by char
splited_chars_l = {}
for i in range(len(img_chars_l)):
for im, c in zip(img_chars_l[i][0], img_chars_l[i][1]):
splited_chars_l[c] = splited_chars_l.get(c, [])
splited_chars_l[c].append(im)
min = 1000000
max = 0
hist = [[c, len(splited_chars_l[c])] for c in splited_chars_l.keys()]
for h in hist:
if min > h[1]:
min = h[1]
if max < h[1]:
max = h[1]
for i, h in enumerate(hist):
if max - min == 0:
hist[i][1] = 2
else:
hist[i][1] = 2 - (h[1] - min) / (max - min)
for c in splited_chars_l.keys():
l = splited_chars_l[c]
for i, _ in enumerate(l):
for h in hist:
if h[0] == c:
l[i] *= h[1]
break
return loss.mean()
def forward(self, input, target, labels):
_assert_no_grad(target)
loss = torch.abs(input - target)
lptnums = [lptnum_of(label) for label in labels]
img_chars_l = []
for i, lptnum in enumerate(lptnums):
for j, _ in enumerate(loss[i, :, self.hb:self.he, :]):
img_chars_l.append(
self._get_img_chars(lptnum, loss[i, j,
self.hb:self.he, :]))
# test only
# for i in range(len(img_chars_l)):
# k=0
# for im, c in zip(img_chars_l[i][0],img_chars_l[i][1]):
# utils.save_image(im,"{}_{}_{}.png".format(i,c,k),normalize=True, range=(0,255))
# k += 1
# raise SystemExit
# Split img_char by char
splited_chars_l = {}
for i in range(len(img_chars_l)):
for im, c in zip(img_chars_l[i][0], img_chars_l[i][1]):
splited_chars_l[c] = splited_chars_l.get(c, [])
splited_chars_l[c].append(im)
# Calculate mean loss
mean_char_loss = []
total_char_area = 0
for k in splited_chars_l.keys():
sum_k = 0
sum_s = 0
for i in range(len(splited_chars_l[k])):
sum_k += splited_chars_l[k][i].sum()
sum_s += product(splited_chars_l[k][i].size())
splited_chars_l[k][i] = 0
mean_char_loss.append(sum_k / sum_s)
total_char_area += sum_s
bg_loss = loss.sum() / (product(loss.size()) - total_char_area)
mean_char_loss.append(bg_loss)
return sum(mean_char_loss) / len(mean_char_loss)
L1 = [(Cx[0],Cy[0]-1)]
L1.append((L1[0][0] + char_size[0] + Cx[1], Cy[0]-1))
L1.append((L1[1][0] + char_size[0] + Cx[2]+1, Cy[0]-1))
L1.append((L1[2][0] + char_size[0] + Cx[1], Cy[0]-1))
L1.append((Cx[0], Cy[0] + char_size[1] + Cy[1]))
L1.append((L1[4][0] + char_size[0] + Cx[3], L1[4][1]))
L1.append((L1[5][0] + char_size[0] + Cx[3], L1[4][1]))
L1.append((L1[6][0] + char_size[0] + Cx[3], L1[4][1]))
L2 = [(Dx[0], Cy[0] + char_size[1] + Cy[1])]
L2.append((L2[0][0] + char_size[0] + Dx[1], L2[0][1]))
L2.append((L2[1][0] + char_size[0] + Dx[1], L2[0][1]))
L2.append((L2[2][0] + char_size[0] + Dx[2], L2[0][1]))
L2.append((L2[3][0] + char_size[0] + Dx[1], L2[0][1]))
# print(path.basename(filename)) # debug
lpt_num = lptnum_of(path.basename(filename), sep=opt.sep)
lptnum = lpt_num.split("_")
lptnum[1] = lptnum[1].replace(".","")
if len(lptnum[0]) == 5: # e.g. 59MĐ1_123.45
raise NotImplementedError
elif len(lptnum[0]) == 3: # e.g. 59F_123.45
raise NotImplementedError
elif len(lptnum[0]) == 4: # e.g. 59Z1_123.45
img_chars = [img[L1[i][1]:char_size[1]+L1[i][1], L1[i][0]:char_size[0]+L1[i][0]] for i in range(len(lptnum[0]))]
compareimg_chars = [compareimg[L1[i][1]:char_size[1]+L1[i][1], L1[i][0]:char_size[0]+L1[i][0]] for i in range(len(lptnum[0]))]
else:
raise ValueError("len(lptnum[0]) != 3 or 4 or 5")
if len(lptnum[1]) == 4: # e.g. 59Z1_1234
img_chars.extend([img[L1[i+len(lptnum[0])][1]:L1[i+len(lptnum[0])][1]+char_size[1], L1[i+len(lptnum[0])][0]:L1[i+len(lptnum[0])][0]+char_size[0]] for i in range(len(lptnum[1]))])
with open(opt.textfile) as f:
lines = f.read().splitlines()
alphabet = [
'0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E',
'F', 'G', 'H', 'K', 'L', 'M', 'N', 'P', 'R', 'S', 'T', 'U', 'V', 'X', 'Y',
'Z'
]
char_count = {}
hist = {}
for line in lines:
linesplt = line.split(":")
try:
truth = lptnum_of(linesplt[0])
except IndexError:
truth = None
if truth == None:
print("The string \"{}\" is wrong lptnum format.".format(linesplt[0]))
continue
recover = linesplt[-1].strip(' ')
for t, r in zip(truth, recover):
if t == "_" or t == ".":
continue
if t != r:
hist[(t, r)] = hist.get((t, r), 0) + 1
char_count[t] = char_count.get(t, 0) + 1
for k in hist.keys():
hist[k] /= char_count[k[0]]