if not os.path.exists(training_options['model_path']):
os.mkdir(training_options['model_path'])
# 定义神经网络
net, train = model.build_network(training_options=training_options, image=image,
drop_rate=training_options['drop_rate'], labels=labels)
# 获得saver对象，可以保存model以及读取model
saver = tf.train.Saver()
saver.restore(sess, training_options['model_path'] + training_options['model_name'])
# 初始化 imageUtils类，获得所有训练，测试数据
imageUtils = ImageUtils()
test_data = imageUtils.test_data
test_label = imageUtils.test_label
total_result = np.zeros((200, 2))
for i in range(200):
result = sess.run([net['digit1'], net['digit2'], net['digit3'], net['digit4']],
feed_dict={image: imageUtils.trainstion_data(test_data, start=i * 1000,end=i * 1000 + 1000)})
result = code_utils.batch_out_transition(result)
predicted = [result[0][index] + result[1][index] + result[2][index] + result[3][index]
for index in range(len(result[0]))]
label = code_utils.batch_out_transition(test_label[i * 1000:i * 1000 + 1000])
four_right_count = np.count_nonzero([predicted[index] == label[index] for index in range(len(predicted))])
one_right_count = np.count_nonzero(
[predicted[index][s_index] == label[index][s_index] for index in range(len(predicted)) for s_index in
range(len(predicted[index]))])
total_result[i, 0] = four_right_count / 1000 * 100
total_result[i, 1] = one_right_count / 4000 * 100
print('一千个测试数据： 四个字符同时正确率: {0:.2f}%\t\t单个字符正确率: {1:.2f}%'.format(total_result[i, 0],
total_result[i, 1]))
print('总结果： 四个字符同时正确率: {0:.2f}%\t\t单个字符正确率: {1:.2f}%'.format(np.mean(total_result[:, 0]),
np.mean(total_result[:, 1])))
labels['digit3']: sample_labels[:, 2],
labels['digit4']: sample_labels[:, 3]
})
writer.add_summary(result, global_step)
if episode != 0 and episode % training_options['save_step'] == 0:
saver.save(
sess, training_options['model_path'] +
training_options['model_name'])
# 指定回合数时 打印预测值、真实值、正确率
if episode % training_options['show_test'] == 0:
t_sample_datas, t_sample_labels = imageUtils.sample(
len(test_data), 10, test_data, test_label)
result = sess.run(
[net['digit1'], net['digit2'], net['digit3'], net['digit4']],
feed_dict={image: t_sample_datas})
result = code_utils.batch_out_transition(result)
predicted = [
result[0][index] + result[1][index] + result[2][index] +
result[3][index] for index in range(len(result[0]))
]
label = code_utils.batch_out_transition(t_sample_labels)
four_right_count = np.count_nonzero([
predicted[index] == label[index]
for index in range(len(predicted))
])
one_right_count = np.count_nonzero([
predicted[index][s_index] == label[index][s_index]
for index in range(len(predicted))
for s_index in range(len(predicted[index]))
])
print('predicted:\t{} \nlabel:\t\t{}'.format(predicted, label))