TensorFlow与中文手写汉字识别 - Python - 伯乐在线
& TensorFlow与中文手写汉字识别
本文目标是利用TensorFlow做一个简单的图像分类器，在比较大的数据集上，尽可能高效地做图像相关处理，从Train，Validation到Inference，是一个比较基本的Example， 从一个基本的任务学习如果在TensorFlow下做高效地图像读取，基本的图像处理，整个项目很简单，但其中有一些trick，在实际项目当中有很大的好处， 比如绝对不要一次读入所有的 的数据到内存（尽管在Mnist这类级别的例子上经常出现)…
最开始看到是这篇blog里面的, 但是这篇文章只用了140训练与测试，试了下代码 很快，但是当扩展到所有的时，发现32g的内存都不够用，这才注意到原文中都是用numpy，会先把所有的数据放入到内存，但这个不必须的，无论在MXNet还是TensorFlow中都是不必 须的，MXNet使用的是DataIter，会在程序运行的过程中异步读取数据，TensorFlow也是这样的，TensorFlow封装了高级的api，用来做数据的读取，比如TFRecord，还有就是从filenames中读取， 来异步读取文件，然后做shuffle batch，再feed到模型的Graph中来做模型参数的更新。具体在tf如何做数据的读取可以看看
这里我会拿到所有的数据集来做训练与测试，算作是对斗大的熊猫上面那篇文章的一个扩展。
Batch Generate
数据集来自于，感谢分享精神！！！具体下载:
wget http://www.nlpr.ia.ac.cn/databases/download/feature_data/HWDB1.1trn_gnt.zip
wget http://www.nlpr.ia.ac.cn/databases/download/feature_data/HWDB1.1tst_gnt.zip
<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d
wget http://www.nlpr.ia.ac.cn/databases/download/feature_data/HWDB1.1trn_gnt.zipwget http://www.nlpr.ia.ac.cn/databases/download/feature_data/HWDB1.1tst_gnt.zip&
解压后发现是一些gnt文件，然后用了斗大的熊猫里面的代码，将所有文件都转化为对应label目录下的所有png的图片。（注意在HWDB1.1trn_gnt.zip解压后是alz文件，需要再次解压 我在mac没有找到合适的工具，windows上有alz的解压工具)。
import numpy as np
import struct
from PIL import Image
data_dir = '../data'
train_data_dir = os.path.join(data_dir, 'HWDB1.1trn_gnt')
test_data_dir = os.path.join(data_dir, 'HWDB1.1tst_gnt')
def read_from_gnt_dir(gnt_dir=train_data_dir):
def one_file(f):
header_size = 10
while True:
header = np.fromfile(f, dtype='uint8', count=header_size)
if not header.size: break
sample_size = header[0] + (header[1]&&8) + (header[2]&&16) + (header[3]&&24)
tagcode = header[5] + (header[4]&&8)
width = header[6] + (header[7]&&8)
height = header[8] + (header[9]&&8)
if header_size + width*height != sample_size:
image = np.fromfile(f, dtype='uint8', count=width*height).reshape((height, width))
yield image, tagcode
for file_name in os.listdir(gnt_dir):
if file_name.endswith('.gnt'):
file_path = os.path.join(gnt_dir, file_name)
with open(file_path, 'rb') as f:
for image, tagcode in one_file(f):
yield image, tagcode
char_set = set()
for _, tagcode in read_from_gnt_dir(gnt_dir=train_data_dir):
tagcode_unicode = struct.pack('&H', tagcode).decode('gb2312')
char_set.add(tagcode_unicode)
char_list = list(char_set)
char_dict = dict(zip(sorted(char_list), range(len(char_list))))
print len(char_dict)
import pickle
f = open('char_dict', 'wb')
pickle.dump(char_dict, f)
train_counter = 0
test_counter = 0
for image, tagcode in read_from_gnt_dir(gnt_dir=train_data_dir):
tagcode_unicode = struct.pack('&H', tagcode).decode('gb2312')
im = Image.fromarray(image)
dir_name = '../data/train/' + '%0.5d'%char_dict[tagcode_unicode]
if not os.path.exists(dir_name):
os.mkdir(dir_name)
im.convert('RGB').save(dir_name+'/' + str(train_counter) + '.png')
train_counter += 1
for image, tagcode in read_from_gnt_dir(gnt_dir=test_data_dir):
tagcode_unicode = struct.pack('&H', tagcode).decode('gb2312')
im = Image.fromarray(image)
dir_name = '../data/test/' + '%0.5d'%char_dict[tagcode_unicode]
if not os.path.exists(dir_name):
os.mkdir(dir_name)
im.convert('RGB').save(dir_name+'/' + str(test_counter) + '.png')
test_counter += 1
<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d
import osimport numpy as npimport structfrom PIL import Image&&data_dir = '../data'train_data_dir = os.path.join(data_dir, 'HWDB1.1trn_gnt')test_data_dir = os.path.join(data_dir, 'HWDB1.1tst_gnt')&&def read_from_gnt_dir(gnt_dir=train_data_dir):&&&&def one_file(f):&&&&&&&&header_size = 10&&&&&&&&while True:&&&&&&&&&&&&header = np.fromfile(f, dtype='uint8', count=header_size)&&&&&&&&&&&&if not header.size: break&&&&&&&&&&&&sample_size = header[0] + (header[1]&&8) + (header[2]&&16) + (header[3]&&24)&&&&&&&&&&&&tagcode = header[5] + (header[4]&&8)&&&&&&&&&&&&width = header[6] + (header[7]&&8)&&&&&&&&&&&&height = header[8] + (header[9]&&8)&&&&&&&&&&&&if header_size + width*height != sample_size:&&&&&&&&&&&&&&&&break&&&&&&&&&&&&image = np.fromfile(f, dtype='uint8', count=width*height).reshape((height, width))&&&&&&&&&&&&yield image, tagcode&&&&for file_name in os.listdir(gnt_dir):&&&&&&&&if file_name.endswith('.gnt'):&&&&&&&&&&&&file_path = os.path.join(gnt_dir, file_name)&&&&&&&&&&&&with open(file_path, 'rb') as f:&&&&&&&&&&&&&&&&for image, tagcode in one_file(f):&&&&&&&&&&&&&&&&&&&&yield image, tagcodechar_set = set()for _, tagcode in read_from_gnt_dir(gnt_dir=train_data_dir):&&&&tagcode_unicode = struct.pack('&H', tagcode).decode('gb2312')&&&&char_set.add(tagcode_unicode)char_list = list(char_set)char_dict = dict(zip(sorted(char_list), range(len(char_list))))print len(char_dict)import picklef = open('char_dict', 'wb')pickle.dump(char_dict, f)f.close()train_counter = 0test_counter = 0for image, tagcode in read_from_gnt_dir(gnt_dir=train_data_dir):&&&&tagcode_unicode = struct.pack('&H', tagcode).decode('gb2312')&&&&im = Image.fromarray(image)&&&&dir_name = '../data/train/' + '%0.5d'%char_dict[tagcode_unicode]&&&&if not os.path.exists(dir_name):&&&&&&&&os.mkdir(dir_name)&&&&im.convert('RGB').save(dir_name+'/' + str(train_counter) + '.png')&&&&train_counter += 1for image, tagcode in read_from_gnt_dir(gnt_dir=test_data_dir):&&&&tagcode_unicode = struct.pack('&H', tagcode).decode('gb2312')&&&&im = Image.fromarray(image)&&&&dir_name = '../data/test/' + '%0.5d'%char_dict[tagcode_unicode]&&&&if not os.path.exists(dir_name):&&&&&&&&os.mkdir(dir_name)&&&&im.convert('RGB').save(dir_name+'/' + str(test_counter) + '.png')&&&&test_counter += 1
处理好的数据，放到了云盘，大家可以直接在我的云盘来下载处理好的数据集. 这里说明下，char_dict是汉字和对应的数字label的记录。
得到数据集后，就要考虑如何读取了，一次用numpy读入内存在很多小数据集上是可以行的，但是在稍微大点的数据集上内存就成了瓶颈，但是不要害怕，TensorFlow有自己的方法：
def batch_data(file_labels,sess, batch_size=128):
image_list = [file_label[0] for file_label in file_labels]
label_list = [int(file_label[1]) for file_label in file_labels]
print 'tag2 {0}'.format(len(image_list))
images_tensor = tf.convert_to_tensor(image_list, dtype=tf.string)
labels_tensor = tf.convert_to_tensor(label_list, dtype=tf.int64)
input_queue = tf.train.slice_input_producer([images_tensor, labels_tensor])
labels = input_queue[1]
images_content = tf.read_file(input_queue[0])
# images = tf.image.decode_png(images_content, channels=1)
images = tf.image.convert_image_dtype(tf.image.decode_png(images_content, channels=1), tf.float32)
# images = images / 256
pre_process(images)
# print images.get_shape()
labels = tf.one_hot(labels, 3755)
image_batch, label_batch = tf.train.shuffle_batch([images, labels], batch_size=batch_size, capacity=50000,min_after_dequeue=10000)
# print 'image_batch', image_batch.get_shape()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
return image_batch, label_batch, coord, threads
<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d
def batch_data(file_labels,sess, batch_size=128):&&&&image_list = [file_label[0] for file_label in file_labels]&&&&label_list = [int(file_label[1]) for file_label in file_labels]&&&&print 'tag2 {0}'.format(len(image_list))&&&&images_tensor = tf.convert_to_tensor(image_list, dtype=tf.string)&&&&labels_tensor = tf.convert_to_tensor(label_list, dtype=tf.int64)&&&&input_queue = tf.train.slice_input_producer([images_tensor, labels_tensor])&&&&&labels = input_queue[1]&&&&images_content = tf.read_file(input_queue[0])&&&&# images = tf.image.decode_png(images_content, channels=1)&&&&images = tf.image.convert_image_dtype(tf.image.decode_png(images_content, channels=1), tf.float32)&&&&# images = images / 256&&&&images =&&pre_process(images)&&&&# print images.get_shape()&&&&# one hot&&&&labels = tf.one_hot(labels, 3755)&&&&image_batch, label_batch = tf.train.shuffle_batch([images, labels], batch_size=batch_size, capacity=50000,min_after_dequeue=10000)&&&&# print 'image_batch', image_batch.get_shape()&&&&&coord = tf.train.Coordinator()&&&&threads = tf.train.start_queue_runners(sess=sess, coord=coord)&&&&return image_batch, label_batch, coord, threads&
简单介绍下，首先你需要得到所有的图像的path和对应的label的列表，利用tf.convert_to_tensor转换为对应的tensor， 利用tf.train.slice_input_producer将image_list ,label_list做一个slice处理，然后做图像的读取、预处理，以及label的one_hot表示，然后就是传到tf.train.shuffle_batch产生一个个shuffle batch，这些就可以feed到你的 模型。 slice_input_producer和shuffle_batch这类操作内部都是基于queue，是一种异步的处理方式,会在设备中开辟一段空间用作cache，不同的进程会分别一直往cache中塞数据 和取数据，保证内存或显存的占用以及每一个mini-batch不需要等待，直接可以从cache中获取。
Data Augmentation
由于图像场景不复杂，只是做了一些基本的处理，包括图像翻转，改变下亮度等等，这些在TensorFlow里面有现成的api，所以尽量使用TensorFlow来做相关的处理：
def pre_process(images):
if FLAGS.random_flip_up_down:
images = tf.image.random_flip_up_down(images)
if FLAGS.random_flip_left_right:
images = tf.image.random_flip_left_right(images)
if FLAGS.random_brightness:
images = tf.image.random_brightness(images, max_delta=0.3)
if FLAGS.random_contrast:
images = tf.image.random_contrast(images, 0.8, 1.2)
new_size = tf.constant([FLAGS.image_size,FLAGS.image_size], dtype=tf.int32)
images = tf.image.resize_images(images, new_size)
return images
<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d
def pre_process(images):&&&&if FLAGS.random_flip_up_down:&&&&&&&&images = tf.image.random_flip_up_down(images)&&&&if FLAGS.random_flip_left_right:&&&&&&&&images = tf.image.random_flip_left_right(images)&&&&if FLAGS.random_brightness:&&&&&&&&images = tf.image.random_brightness(images, max_delta=0.3)&&&&if FLAGS.random_contrast:&&&&&&&&images = tf.image.random_contrast(images, 0.8, 1.2)&&&&new_size = tf.constant([FLAGS.image_size,FLAGS.image_size], dtype=tf.int32)&&&&images = tf.image.resize_images(images, new_size)&&&&return images&
Build Graph
这里很简单的构造了一个两个卷积+一个全连接层的网络，没有做什么更深的设计，感觉意义不大，设计了一个dict，用来返回后面要用的所有op，还有就是为了方便再训练中查看loss和accuracy， 没有什么特别的，很容易理解, labels 为None时 方便做inference。
def network(images, labels=None):
endpoints = {}
conv_1 = slim.conv2d(images, 32, [3,3],1, padding='SAME')
max_pool_1 = slim.max_pool2d(conv_1, [2,2],[2,2], padding='SAME')
conv_2 = slim.conv2d(max_pool_1, 64, [3,3],padding='SAME')
max_pool_2 = slim.max_pool2d(conv_2, [2,2],[2,2], padding='SAME')
flatten = slim.flatten(max_pool_2)
out = slim.fully_connected(flatten,3755, activation_fn=None)
global_step = tf.Variable(initial_value=0)
if labels is not None:
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(out, labels))
train_op = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(loss, global_step=global_step)
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(out, 1), tf.argmax(labels, 1)), tf.float32))
tf.summary.scalar('loss', loss)
tf.summary.scalar('accuracy', accuracy)
merged_summary_op = tf.summary.merge_all()
output_score = tf.nn.softmax(out)
predict_val_top3, predict_index_top3 = tf.nn.top_k(output_score, k=3)
endpoints['global_step'] = global_step
if labels is not None:
endpoints['labels'] = labels
endpoints['train_op'] = train_op
endpoints['loss'] = loss
endpoints['accuracy'] = accuracy
endpoints['merged_summary_op'] = merged_summary_op
endpoints['output_score'] = output_score
endpoints['predict_val_top3'] = predict_val_top3
endpoints['predict_index_top3'] = predict_index_top3
return endpoints
<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d
def network(images, labels=None):&&&&endpoints = {}&&&&conv_1 = slim.conv2d(images, 32, [3,3],1, padding='SAME')&&&&max_pool_1 = slim.max_pool2d(conv_1, [2,2],[2,2], padding='SAME')&&&&conv_2 = slim.conv2d(max_pool_1, 64, [3,3],padding='SAME')&&&&max_pool_2 = slim.max_pool2d(conv_2, [2,2],[2,2], padding='SAME')&&&&flatten = slim.flatten(max_pool_2)&&&&out = slim.fully_connected(flatten,3755, activation_fn=None)&&&&global_step = tf.Variable(initial_value=0)&&&&if labels is not None:&&&&&&&&loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(out, labels))&&&&&&&&train_op = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(loss, global_step=global_step)&&&&&&&&accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(out, 1), tf.argmax(labels, 1)), tf.float32))&&&&&&&&tf.summary.scalar('loss', loss)&&&&&&&&tf.summary.scalar('accuracy', accuracy)&&&&&&&&merged_summary_op = tf.summary.merge_all()&&&&output_score = tf.nn.softmax(out)&&&&predict_val_top3, predict_index_top3 = tf.nn.top_k(output_score, k=3)&&&&&endpoints['global_step'] = global_step&&&&if labels is not None:&&&&&&&&endpoints['labels'] = labels&&&&&&&&endpoints['train_op'] = train_op&&&&&&&&endpoints['loss'] = loss&&&&&&&&endpoints['accuracy'] = accuracy&&&&&&&&endpoints['merged_summary_op'] = merged_summary_op&&&&endpoints['output_score'] = output_score&&&&endpoints['predict_val_top3'] = predict_val_top3&&&&endpoints['predict_index_top3'] = predict_index_top3&&&&return endpoints&
train函数包括从已有checkpoint中restore，得到step，快速恢复训练过程，训练主要是每一次得到mini-batch，更新参数，每隔eval_steps后做一次train batch的eval，每隔save_steps 后保存一次checkpoint。
def train():
sess = tf.Session()
file_labels = get_imagesfile(FLAGS.train_data_dir)
images, labels, coord, threads = batch_data(file_labels, sess)
endpoints = network(images, labels)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
train_writer = tf.train.SummaryWriter('./log' + '/train',sess.graph)
test_writer = tf.train.SummaryWriter('./log' + '/val')
start_step = 0
if FLAGS.restore:
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
saver.restore(sess, ckpt)
print "restore from the checkpoint {0}".format(ckpt)
start_step += int(ckpt.split('-')[-1])
logger.info(':::Training Start:::')
while not coord.should_stop():
# logger.info('step {0} start'.format(i))
start_time = time.time()
_, loss_val, train_summary, step = sess.run([endpoints['train_op'], endpoints['loss'], endpoints['merged_summary_op'], endpoints['global_step']])
train_writer.add_summary(train_summary, step)
end_time = time.time()
logger.info("the step {0} takes {1} loss {2}".format(step, end_time-start_time, loss_val))
if step & FLAGS.max_steps:
# logger.info("the step {0} takes {1} loss {2}".format(i, end_time-start_time, loss_val))
if step % FLAGS.eval_steps == 1:
accuracy_val,test_summary, step = sess.run([endpoints['accuracy'], endpoints['merged_summary_op'], endpoints['global_step']])
test_writer.add_summary(test_summary, step)
logger.info('===============Eval a batch in Train data=======================')
logger.info( 'the step {0} accuracy {1}'.format(step, accuracy_val))
logger.info('===============Eval a batch in Train data=======================')
if step % FLAGS.save_steps == 1:
logger.info('Save the ckpt of {0}'.format(step))
saver.save(sess, os.path.join(FLAGS.checkpoint_dir, 'my-model'), global_step=endpoints['global_step'])
except tf.errors.OutOfRangeError:
# print "============train finished========="
logger.info('==================Train Finished================')
saver.save(sess, os.path.join(FLAGS.checkpoint_dir, 'my-model'), global_step=endpoints['global_step'])
coord.request_stop()
coord.join(threads)
sess.close()
12345678910111213141516171819202122232425262728293031323334353637383940414243444546
def train():&&&&sess = tf.Session()&&&&file_labels = get_imagesfile(FLAGS.train_data_dir)&&&&images, labels, coord, threads = batch_data(file_labels, sess)&&&&endpoints = network(images, labels)&&&&saver = tf.train.Saver()&&&&sess.run(tf.global_variables_initializer())&&&&train_writer = tf.train.SummaryWriter('./log' + '/train',sess.graph)&&&&test_writer = tf.train.SummaryWriter('./log' + '/val')&&&&start_step = 0&&&&if FLAGS.restore:&&&&&&&&ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)&&&&&&&&if ckpt:&&&&&&&&&&&&saver.restore(sess, ckpt)&&&&&&&&&&&&print "restore from the checkpoint {0}".format(ckpt)&&&&&&&&&&&&start_step += int(ckpt.split('-')[-1])&&&&logger.info(':::Training Start:::')&&&&try:&&&&&&&&while not coord.should_stop():&&&&&&&&# logger.info('step {0} start'.format(i))&&&&&&&&&&&&start_time = time.time()&&&&&&&&&&&&_, loss_val, train_summary, step = sess.run([endpoints['train_op'], endpoints['loss'], endpoints['merged_summary_op'], endpoints['global_step']])&&&&&&&&&&&&train_writer.add_summary(train_summary, step)&&&&&&&&&&&&end_time = time.time()&&&&&&&&&&&&logger.info("the step {0} takes {1} loss {2}".format(step, end_time-start_time, loss_val))&&&&&&&&&&&&if step & FLAGS.max_steps:&&&&&&&&&&&&&&&&break&&&&&&&&&&&&# logger.info("the step {0} takes {1} loss {2}".format(i, end_time-start_time, loss_val))&&&&&&&&&&&&if step % FLAGS.eval_steps == 1:&&&&&&&&&&&&&&&&accuracy_val,test_summary, step = sess.run([endpoints['accuracy'], endpoints['merged_summary_op'], endpoints['global_step']])&&&&&&&&&&&&&&&&test_writer.add_summary(test_summary, step)&&&&&&&&&&&&&&&&logger.info('===============Eval a batch in Train data=======================')&&&&&&&&&&&&&&&&logger.info( 'the step {0} accuracy {1}'.format(step, accuracy_val))&&&&&&&&&&&&&&&&logger.info('===============Eval a batch in Train data=======================')&&&&&&&&&&&&if step % FLAGS.save_steps == 1:&&&&&&&&&&&&&&&&logger.info('Save the ckpt of {0}'.format(step))&&&&&&&&&&&&&&&&saver.save(sess, os.path.join(FLAGS.checkpoint_dir, 'my-model'), global_step=endpoints['global_step'])&&&&except tf.errors.OutOfRangeError:&&&&&&&&# print "============train finished========="&&&&&&&&logger.info('==================Train Finished================')&&&&&&&&saver.save(sess, os.path.join(FLAGS.checkpoint_dir, 'my-model'), global_step=endpoints['global_step'])&&&&finally:&&&&&&&&coord.request_stop()&&&&coord.join(threads)&&&&sess.close()&
Loss and Accuracy
Validation
训练完成之后，想对最终的模型在测试数据集上做一个评估，这里我也曾经尝试利用batch_data，将slice_input_producer中epoch设置为1，来做相关的工作，但是发现这里无法和train 共用，会出现epoch无初始化值的问题（train中传epoch为None），所以这里自己写了shuffle batch的逻辑，将测试集的images和labels通过feed_dict传进到网络，得到模型的输出， 然后做相关指标的计算：
def validation():
# it should be fixed by using placeholder with epoch num in train stage
sess = tf.Session()
file_labels = get_imagesfile(FLAGS.test_data_dir)
test_size = len(file_labels)
print test_size
val_batch_size = FLAGS.val_batch_size
test_steps = test_size / val_batch_size
print test_steps
# images, labels, coord, threads= batch_data(file_labels, sess)
images = tf.placeholder(dtype=tf.float32, shape=[None, 64, 64, 1])
labels = tf.placeholder(dtype=tf.int32, shape=[None,3755])
# read batch images from file_labels
# images_batch = np.zeros([128,64,64,1])
# labels_batch = np.zeros([128,3755])
# labels_batch[0][20] = 1
endpoints = network(images, labels)
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
saver.restore(sess, ckpt)
# logger.info("restore from the checkpoint {0}".format(ckpt))
# logger.info('Start validation')
final_predict_val = []
final_predict_index = []
groundtruth = []
for i in range(test_steps):
start = i* val_batch_size
end = (i+1)*val_batch_size
images_batch = []
labels_batch = []
labels_max_batch = []
logger.info('=======start validation on {0}/{1} batch========='.format(i, test_steps))
for j in range(start,end):
image_path = file_labels[j][0]
temp_image = Image.open(image_path).convert('L')
temp_image = temp_image.resize((FLAGS.image_size, FLAGS.image_size),Image.ANTIALIAS)
temp_label = np.zeros([3755])
label = int(file_labels[j][1])
# print label
temp_label[label] = 1
# print "====",np.asarray(temp_image).shape
labels_batch.append(temp_label)
# print "====",np.asarray(temp_image).shape
images_batch.append(np.asarray(temp_image)/255.0)
labels_max_batch.append(label)
# print images_batch
images_batch = np.array(images_batch).reshape([-1, 64, 64, 1])
labels_batch = np.array(labels_batch)
batch_predict_val, batch_predict_index = sess.run([endpoints['predict_val_top3'],
endpoints['predict_index_top3']], feed_dict={images:images_batch, labels:labels_batch})
logger.info('=======validation on {0}/{1} batch end========='.format(i, test_steps))
final_predict_val += batch_predict_val.tolist()
final_predict_index += batch_predict_index.tolist()
groundtruth += labels_max_batch
sess.close()
return final_predict_val, final_predict_index, groundtruth
<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d<div class="crayon-num" data-line="crayon-5b06d<div class="crayon-num crayon-striped-num" data-line="crayon-5b06d
def validation():&&&&# it should be fixed by using placeholder with epoch num in train stage&&&&sess = tf.Session()&&&&&file_labels = get_imagesfile(FLAGS.test_data_dir)&&&&test_size = len(file_labels)&&&&print test_size&&&&val_batch_size = FLAGS.val_batch_size&&&&test_steps = test_size / val_batch_size&&&&print test_steps&&&&# images, labels, coord, threads= batch_data(file_labels, sess)&&&&images = tf.placeholder(dtype=tf.float32, shape=[None, 64, 64, 1])&&&&labels = tf.placeholder(dtype=tf.int32, shape=[None,3755])&&&&# read batch images from file_labels&&&&# images_batch = np.zeros([128,64,64,1])&&&&# labels_batch = np.zeros([128,3755])&&&&# labels_batch[0][20] = 1&&&&#&&&&endpoints = network(images, labels)&&&&saver = tf.train.Saver()&&&&ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)&&&&if ckpt:&&&&&&&&saver.restore(sess, ckpt)&&&&&&&&# logger.info("restore from the checkpoint {0}".format(ckpt))&&&&# logger.info('Start validation')&&&&final_predict_val = []&&&&final_predict_index = []&&&&groundtruth = []&&&&for i in range(test_steps):&&&&&&&&start = i* val_batch_size&&&&&&&&end = (i+1)*val_batch_size&&&&&&&&images_batch = []&&&&&&&&labels_batch = []&&&&&&&&labels_max_batch = []&&&&&&&&logger.info('=======start validation on {0}/{1} batch========='.format(i, test_steps))&&&&&&&&for j in range(start,end):&&&&&&&&&&&&image_path = file_labels[j][0]&&&&&&&&&&&&temp_image = Image.open(image_path).convert('L')&&&&&&&&&&&&temp_image = temp_image.resize((FLAGS.image_size, FLAGS.image_size),Image.ANTIALIAS)&&&&&&&&&&&&temp_label = np.zeros([3755])&&&&&&&&&&&&label = int(file_labels[j][1])&&&&&&&&&&&&# print label&&&&&&&&&&&&temp_label[label] = 1&&&&&&&&&&&&# print "====",np.asarray(temp_image).shape&&&&&&&&&&&&labels_batch.append(temp_label)&&&&&&&&&&&&# print "====",np.asarray(temp_image).shape&&&&&&&&&&&&images_batch.append(np.asarray(temp_image)/255.0)&&&&&&&&&&&&labels_max_batch.append(label)&&&&&&&&# print images_batch&&&&&&&&images_batch = np.array(images_batch).reshape([-1, 64, 64, 1])&&&&&&&&labels_batch = np.array(labels_batch)&&&&&&&&batch_predict_val, batch_predict_index = sess.run([endpoints['predict_val_top3'],&&&&&&&&&&&&&&&&&&&&&&&&endpoints['predict_index_top3']], feed_dict={images:images_batch, labels:labels_batch})&&&&&&&&logger.info('=======validation on {0}/{1} batch end========='.format(i, test_steps))&&&&&&&&final_predict_val += batch_predict_val.tolist()&&&&&&&&final_predict_index += batch_predict_index.tolist()&&&&&&&&groundtruth += labels_max_batch&&&&sess.close()&&&&return final_predict_val, final_predict_index, groundtruth&
在训练20w个step之后，大概能达到在测试集上能够达到：
相信如果在网络设计上多花点时间能够在一定程度上提升accuracy和top 3 accuracy.有兴趣的小伙伴们可以玩玩这个数据集。
def inference(image):
temp_image = Image.open(image).convert('L')
temp_image = temp_image.resize((FLAGS.image_size, FLAGS.image_size),Image.ANTIALIAS)
sess = tf.Session()
logger.info('========start inference============')
images = tf.placeholder(dtype=tf.float32, shape=[None, 64, 64, 1])
endpoints = network(images)
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
saver.restore(sess, ckpt)
predict_val, predict_index = sess.run([endpoints['predict_val_top3'],endpoints['predict_index_top3']], feed_dict={images:temp_image})
sess.close()
return final_predict_val, final_predict_index
123456789101112131415
def inference(image):&&&&temp_image = Image.open(image).convert('L')&&&&temp_image = temp_image.resize((FLAGS.image_size, FLAGS.image_size),Image.ANTIALIAS)&&&&sess = tf.Session()&&&&logger.info('========start inference============')&&&&images = tf.placeholder(dtype=tf.float32, shape=[None, 64, 64, 1])&&&&endpoints = network(images)&&&&saver = tf.train.Saver()&&&&ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)&&&&if ckpt:&&&&&&&&saver.restore(sess, ckpt)&&&&predict_val, predict_index = sess.run([endpoints['predict_val_top3'],endpoints['predict_index_top3']], feed_dict={images:temp_image})&&&&sess.close()&&&&return final_predict_val, final_predict_index&
运气挺好，随便找了张图片就能准确识别出来
综上，就是利用tensorflow做中文手写识别的全部，从如何使用tensorflow内部的queue来有效读入数据，到如何设计network， 到如何做train，validation，inference，珍格格流程比较清晰， 美中不足的是，原本打算是在训练过程中，来对测试集做评估，但是在使用queue读test_data_dir下的filenames，和train本身的好像有点问题，不过应该是可以解决的，我这里就pass了。另外可能 还有一些可以改善的地方，比如感觉可以把batch data one hot的部分写入到network，这样，减缓在validation时内存会因为onehot的sparse开销比较大。
感觉这个中文手写汉字数据集价值很大，后面感觉会有好多可以玩的，比如
可以参考项亮大神的这篇文章做定长的字符识别和不定长的字符识别，定长的基本原理是说，可以把最终输出扩展为k个输出， 每个值表示对应的字符label，这样cnn模型在feature extract之后就可以自己去识别对应字符而无需人工切割；而LSTM+CTC来解决不定长的验证码，类似于将音频解码为汉字
最近GAN特别火，感觉可以考虑用这个数据来做某个字的生成，和text2img那个项目
这部分的代码都在我的github上，有遇到相关功能,想参考代码的可以去上面找找，没准就能解决你们遇到的一些小问题.
感谢的PR，使得代码更简洁， 并且修改了网络的结构，使得模型准确率上升很高, 最后top1和top3的结果：
可能感兴趣的话题
感谢楼主，以前练习过MNIST数据集，一直没找到中文数据集，这次终于找到了。
关于 Python 频道
Python频道分享 Python 开发技术、相关的行业动态。
新浪微博：
推荐微信号
（加好友请注明来意）
&#8211; 好的话题、有启发的回复、值得信赖的圈子
&#8211; 分享和发现有价值的内容与观点
&#8211; 为IT单身男女服务的征婚传播平台
&#8211; 优秀的工具资源导航
&#8211; 翻译传播优秀的外文文章
&#8211; 国内外的精选文章
&#8211; UI,网页，交互和用户体验
&#8211; 专注iOS技术分享
&#8211; 专注Android技术分享
&#8211; JavaScript, HTML5, CSS
&#8211; 专注Java技术分享
&#8211; 专注Python技术分享
& 2018 伯乐在线}