def predict(model, imgdata):
img_width, img_height = 223, 50
img_channels = 3
max_nb_cha = 6 # 文本最大长度
len_set = range(1, max_nb_cha + 1) # 文本可能的长度
cha_set = list("0123456789qwertyuioplkjhgfdsazxcvbnm") # 文本字符集
nb_classes = 36
X_test = load_data(imgdata, img_width, img_height, img_channels)
# print X_test
predictions = model.predict({'input': X_test})
pred_chas = [
one_hot_decoder(predictions['output%d' % j], cha_set)
for j in range(1, max_nb_cha + 1)
]
pred_nbs = one_hot_decoder(predictions['output_nb'], len_set)
length = min(max_nb_cha, pred_nbs[0])
ans = []
for i in range(length):
ans.append(pred_chas[i][0])
ans = ''.join(ans)
# print ans
# print predictions['output1']
res = {'valid': True, 'answer': ans, 'expr': ans}
return json.dumps(res)
def test_multi(model, len_set, cha_set, max_nb_cha, X_test, Y_test_nb, Y_test):
# 开始预测并验证准确率，需要先把预测结果从概率转到对应的标签
predictions = model.predict({'input': X_test})
# print predictions
pred_nbs = one_hot_decoder(predictions['output_nb'], len_set)
pred_chas = [
one_hot_decoder(predictions['output%d' % j], cha_set)
for j in range(1, max_nb_cha + 1)
]
Y_test_nb = one_hot_decoder(Y_test_nb, len_set)
Y_test = [one_hot_decoder(i, cha_set) for i in Y_test]
correct = 0
len_correct = 0
nb_sample = X_test.shape[0]
for i in range(nb_sample):
pred_nb = pred_nbs[i]
true_nb = Y_test_nb[i]
# print 'len(pred, true):', pred_nb, true_nb
allright = (pred_nb == true_nb)
if allright:
len_correct += 1
for j in range(true_nb):
# print pred_chas[j][i], Y_test[j][i]
allright = allright and (pred_chas[j][i] == Y_test[j][i])
if allright:
correct += 1
else:
print 'id:' + str(i + 1), pred_chas[0][i], Y_test[0][i]
print 'Length accuracy:', float(len_correct) / nb_sample
print 'Accuracy:', float(correct) / nb_sample
def predict(model, post_vals):
img_width, img_height = 223, 50
img_channels = 3
max_nb_cha = 6 # 文本最大长度
len_set = range(1, max_nb_cha + 1) # 文本可能的长度
cha_set = list("0123456789qwertyuioplkjhgfdsazxcvbnm") # 文本字符集
nb_classes = 36
keys = post_vals.keys()
img_vals = post_vals.values()
X_test = load_data(img_vals, img_width, img_height, img_channels)
predictions = model.predict({'input': X_test})
pred_chas = [
one_hot_decoder(predictions['output%d' % j], cha_set)
for j in range(1, max_nb_cha + 1)
]
pred_nbs = one_hot_decoder(predictions['output_nb'], len_set)
res = {}
for i, img_val in enumerate(img_vals):
length = min(max_nb_cha, pred_nbs[i])
ans = []
for j in range(length):
ans.append(pred_chas[j][i])
ans = ''.join(ans)
valid = True
if sys.getsizeof(img_val.stream) > 3000: # Atten tion!
valid = False
form = {'valid': valid, 'answer': ans, 'expr': ans}
res[keys[i]] = form
print res, len(res)
return json.dumps(res)
def predict(model, post_vals):
img_width, img_height = 223, 50
img_channels = 3
max_nb_cha = 6 # 文本最大长度
len_set = range(1, max_nb_cha+1) # 文本可能的长度
cha_set = list("0123456789qwertyuioplkjhgfdsazxcvbnm")# 文本字符集
nb_classes = 36
keys = post_vals.keys()
img_vals = post_vals.values()
X_test = load_data(img_vals, img_width, img_height, img_channels)
predictions = model.predict({'input':X_test})
pred_chas = [one_hot_decoder(predictions['output%d' % j], cha_set) for j in range(1, max_nb_cha+1)]
pred_nbs = one_hot_decoder(predictions['output_nb'], len_set)
res = {}
for i, img_val in enumerate(img_vals):
length = min(max_nb_cha, pred_nbs[i])
ans = []
for j in range(length):
ans.append(pred_chas[j][i])
ans = ''.join(ans)
valid = True
if sys.getsizeof(img_val.stream) > 3000: # Atten tion!
valid = False
form = {'valid':valid, 'answer':ans, 'expr':ans}
res[keys[i]] = form
print res, len(res)
return json.dumps(res)
def pred(model, X, char_set, label_set, post_correction): pred_res = model.predict(X) pred_res = [one_hot_decoder(i, char_set) for i in pred_res] pred_res = [list2str(i) for i in pred_res] # post correction if post_correction: pred_res = correction(pred_res, label_set) return pred_res
def pred(model, X, char_set, label_set, post_correction):
pred_res = model.predict(X)
pred_res = [one_hot_decoder(i, char_set) for i in pred_res]
pred_res = [list2str(i) for i in pred_res]
# post correction
if post_correction:
pred_res = correction(pred_res, label_set)
return pred_res
def pred(self, X):
pred_res = self.model.predict(X, batch_size=256)
self.pred_probs = pred_res
pred_res = [one_hot_decoder(i, self.char_set) for i in pred_res]
pred_res = [list2str(i) for i in pred_res]
# post correction
if self.post_correction:
pred_res = correction(pred_res, self.label_set)
return pred_res
def pred(self, X):
pred_res = self.model.predict(X, batch_size=256)
self.pred_probs = pred_res
pred_res = [one_hot_decoder(i, self.char_set) for i in pred_res]
pred_res = [list2str(i) for i in pred_res]
# post correction
if self.post_correction:
pred_res = correction(pred_res, self.label_set)
return pred_res
def test_single(model, len_set, cha_set, max_nb_cha, X_test, Y_test_nb,
Y_test):
# 开始预测并验证准确率，需要先把预测结果从概率转到对应的标签
predictions = model.predict(X_test)
print model.predict_proba(X_test)[86]
# print predictions
pred_chas = one_hot_decoder(predictions, cha_set)
Y_test = one_hot_decoder(Y_test[0], cha_set)
correct = 0
nb_sample = X_test.shape[0]
for i in range(nb_sample):
allright = pred_chas[i] == Y_test[i]
if allright:
correct += 1
else:
print 'id:' + str(i + 1), pred_chas[i], Y_test[i]
pass
print 'Accuracy:', float(correct) / nb_sample
def test(model, test_data, char_set, post_correction):
test_X, test_y = test_data[0], test_data[1]
test_y = [one_hot_decoder(i, char_set) for i in test_y]
test_y = [list2str(i) for i in test_y]
pred_res = pred(model, test_X, char_set, post_correction)
nb_correct = sum(pred_res[i]==test_y[i] for i in range(len(pred_res)))
for i in range(len(pred_res)):
if test_y[i] != pred_res[i]:
print ('test:', test_y[i])
print ('pred:', pred_res[i])
pass
print ('nb_correct: ', nb_correct)
print ('Acurracy: ', float(nb_correct) / len(pred_res))
def predict(model, imgdata):
img_width, img_height = 223, 50
img_channels = 3
max_nb_cha = 6 # 文本最大长度
len_set = range(1, max_nb_cha+1) # 文本可能的长度
cha_set = list("0123456789qwertyuioplkjhgfdsazxcvbnm")# 文本字符集
nb_classes = 36
X_test = load_data(imgdata, img_width, img_height, img_channels)
# print X_test
predictions = model.predict({'input':X_test})
pred_chas = [one_hot_decoder(predictions['output%d' % j], cha_set) for j in range(1, max_nb_cha+1)]
pred_nbs = one_hot_decoder(predictions['output_nb'], len_set)
length = min(max_nb_cha, pred_nbs[0])
ans = []
for i in range(length):
ans.append(pred_chas[i][0])
ans = ''.join(ans)
# print ans
# print predictions['output1']
res = {'valid':True, 'answer':ans, 'expr':ans}
return json.dumps(res)
def test(model, test_data, char_set, label_set, post_correction): test_X, test_y = test_data[0], test_data[1] test_y = [one_hot_decoder(i, char_set) for i in test_y] test_y = [list2str(i) for i in test_y] pred_res = pred(model, test_X, char_set, label_set, post_correction) # for i in pred_res: # print i nb_correct = sum(pred_res[i]==test_y[i] for i in range(len(pred_res))) for i in range(len(pred_res)): if test_y[i] != pred_res[i]: print 'test:', test_y[i] print 'pred:', pred_res[i] pass print 'nb_correct: ', nb_correct print 'Acurracy: ', float(nb_correct) / len(pred_res)
