def result_analysis(self):
prediction = self.decode_batch()
batch_size = self.softmax.size(1)
delete_total = 0
replace_total = 0
insert_total = 0
len_total = 0
word_total = 0
for i in range(batch_size):
pre_list = prediction[i]
label_list = self.label[i][self.label[i] != -1].tolist()
label_list = [int(ele) + 1 for ele in label_list]
distance, (delete, replace,
insert) = cal_distance(label_list, pre_list)
delete_total += delete
replace_total += replace
insert_total += insert
len_total += len(label_list)
if distance == 0:
word_total += 1
result = [
delete_total, replace_total, insert_total, len_total, word_total
]
return prediction, result
def result_analysis_recall(self):
prediction = self.decode_batch()
batch_size = self.softmax.size(1)
delete_total = 0
replace_total = 0
insert_total = 0
len_total = 0
correct_count = 0
pre_total = 0
word_total = 0
all_total = 0
for i in range(batch_size):
pre_list = prediction[i]
label_list = self.label[i][self.label[i] != -1].tolist()
label_list = [int(ele) + 1 for ele in label_list]
temp_label_list = list()
for k in range(len(label_list)):
temp_label_list.append(label_list[k])
distance, (delete, replace,
insert) = cal_distance(label_list, pre_list)
delete_total += delete
replace_total += replace
insert_total += insert
len_total += len(label_list)
pre_total += len(pre_list)
correct_count += len(label_list) - delete - replace
if distance == 0:
word_total += 1
# else:
# alphabet = u'_深秦京海成南杭苏松0123456789ABCDEFGHJKLMNPQRSTUVWXYZ'
# label_list.append(0)
# label_list.extend(pre_list)
# file_name = 'wrong_img/%d_%s.jpg'%(self.count_n, ''.join([alphabet[j] for j in label_list]))
# cv2.imwrite(file_name, self.image[i]*127.5+127.5)
# self.count_n += 1
all_total += 1
result = [
delete_total, replace_total, insert_total, len_total,
correct_count, len_total, pre_total, word_total, all_total
]
return prediction, result
def result_analysis_recall(self):
prediction = self.decode_batch()
batch_size = self.softmax.size(1)
delete_total = 0
replace_total = 0
insert_total = 0
len_total = 0
correct_count = 0
pre_total = 0
word_total = 0
all_total = 0
for i in range(batch_size):
pre_list = prediction[i]
label_list = self.label[i][self.label[i] != -1].tolist()
label_list = [int(ele) + 1 for ele in label_list]
distance, (delete, replace,
insert) = cal_distance(label_list, pre_list)
delete_total += delete
replace_total += replace
insert_total += insert
len_total += len(label_list)
pre_total += len(pre_list)
correct_count += len(label_list) - delete - replace
if distance == 0:
word_total += 1
all_total += 1
# print (correct_count , len_total, pre_total)
# recall = float(correct_count) / len_total
# precision = correct_count / (pre_total+0.000001)
result = [
delete_total, replace_total, insert_total, len_total,
correct_count, len_total, pre_total, word_total, all_total
]
return prediction, result
def result_analysis_recall(self):
result = self.softmax.data.topk(1)[1]
result = result.squeeze();
result = result.type(torch.LongTensor)
batch_size,_ = self.label.size()
label_batch = torch.transpose(self.label,0,1)
# label_batch = label_batch.data
delete_total = 0
replace_total = 0
insert_total = 0
len_total = 0
word_total = 0
all_total = 0
pre_total = 0
correct_count = 0
# print(self.processed_batches)
# if self.processed_batches > 5000:
show_result_pred = ''
show_result_label = ''
for i in range(batch_size):
pre_list = []
# pdb.set_trace()
for ln in result[:,i]:
ln = ln.item()
if ln != self.EOS_token:
pre_list.append(ln);
else:
break;
label_list = []
for ln in label_batch[:,i]:
if ln != -1:
label_list.append(int(ln));
else:
break;
# if i == 0:
# print('label_list:', label_list)
# print('pre_list:', pre_list)
# if (self.training and i == 0) or np.random.rand() < 0.0001:
# alphabet = '_$!$#$"$\'$&$)$($+$*$-$,$/$.$1$0$3$2$5$4$7$6$9$8$;$:$?$A$C$B$E$D$G$F$I$H$K$J$M$L$O$N$Q$P$S$R$U$T$W$V$Y$X$Z$a$c$b$e$d$g$f$i$h$k$j$m$l$o$n$q$p$s$r$u$t$w$v$y$x$z$|'.split('$')
# print(''.join([alphabet[i+1] for i in label_list]))
# print(''.join([alphabet[i+1] for i in pre_list]))
# if self.training and i <= debug[1]:
# if i < self.debug[0]:
if i == 0:
alphabet = '_$!$#$"$\'$&$)$($+$*$-$,$/$.$1$0$3$2$5$4$7$6$9$8$;$:$?$A$C$B$E$D$G$F$I$H$K$J$M$L$O$N$Q$P$S$R$U$T$W$V$Y$X$Z$a$c$b$e$d$g$f$i$h$k$j$m$l$o$n$q$p$s$r$u$t$w$v$y$x$z$|'.split('$')
show_result_pred = show_result_pred + ' ' + ''.join([alphabet[i+1] for i in pre_list])
show_result_label = show_result_label + ' ' + ''.join([alphabet[i+1] for i in label_list])
distance, (delete, replace, insert) = cal_distance(label_list, pre_list)
delete_total += delete
replace_total += replace
insert_total += insert
len_total += len(label_list)
pre_total += len(pre_list)
correct_count += len(label_list) - delete - replace;
if distance == 0:
word_total += 1
all_total += 1
if self.training or np.random.rand() < 0.01:
# print('')
# print(self.image_count)
print('label:', show_result_label)
print('pred: ',show_result_pred)
#print(np.array(self.debug[2]))
# from PIL import Image, ImageDraw
# im = Image.fromarray(self.debug[1][0,0].cpu().data.numpy()*127.5+127.5)
# draw = ImageDraw.Draw(im)
# for pos in self.debug[2]:
# x = (pos-1)*16+16
# draw.ellipse((x-8,160-8, x+8, 160+8), fill = 128)
# im = im.convert('RGB')
# if self.training:
# im.save('/media/pci/256c7be3-9106-49d7-a195-2dbc53b60824/zecheng/output/tmp_output001/image/00train_%05d.jpg'% self.image_count)
# else:
# im.save('/media/pci/256c7be3-9106-49d7-a195-2dbc53b60824/zecheng/output/tmp_output001/image/01test_%05d.jpg'% self.image_count)
# self.image_count += 1
rec_result = [delete_total, replace_total, insert_total, len_total, correct_count, len_total, pre_total, word_total, all_total]
return result, rec_result