tflearn 中文汉字识别模型试验汇总

def get_model(width, height, classes=40):

    # TODO, modify model

    # Building 'VGG Network'

    network = input_data(shape=[None, width, height, 1])  # if RGB, 224,224,3

    network = conv_2d(network, 64, 3, activation='relu')

    #network = conv_2d(network, 64, 3, activation='relu')

    network = max_pool_2d(network, 2, strides=2)

    network = conv_2d(network, 128, 3, activation='relu')

    #network = conv_2d(network, 128, 3, activation='relu')

    network = max_pool_2d(network, 2, strides=2)

    network = conv_2d(network, 256, 3, activation='relu')

    #network = conv_2d(network, 256, 3, activation='relu')

    #network = conv_2d(network, 256, 3, activation='relu')

    network = max_pool_2d(network, 2, strides=2)

    network = conv_2d(network, 512, 3, 2, activation='relu')

    # network = conv_2d(network, 512, 3, activation='relu')

    # network = conv_2d(network, 512, 3, activation='relu')

    # network = max_pool_2d(network, 2, strides=2)

    # network = conv_2d(network, 512, 3, activation='relu')

    # network = conv_2d(network, 512, 3, activation='relu')

    # network = conv_2d(network, 512, 3, activation='relu')

    network = max_pool_2d(network, 2, strides=2)

    # network = fully_connected(network, 4096, activation='relu')

    # network = dropout(network, 0.5)

    #network = fully_connected(network, 1024, activation='relu')

    network = fully_connected(network, 2048, activation='relu')

    network = dropout(network, 0.8)

    network = fully_connected(network, classes, activation='softmax')

    network = regression(network, optimizer='rmsprop',

                         loss='categorical_crossentropy',

                         learning_rate=0.0001)

    model = tflearn.DNN(network, checkpoint_path='checkpoint',

                        max_checkpoints=1, tensorboard_verbose=1)

    return model

if __name__ == "__main__":

    width, height = 32, 32

    X, Y, org_labels = load_data(dirname="data", resize_pics=(width, height))

    trainX, testX, trainY, testY = train_test_split(X, Y, test_size=0.2, random_state=666)

    print("sample data:")

    print(trainX[0])

    print(trainY[0])

    print(testX[-1])

    print(testY[-1])

    model = get_model(width, height, classes=100)

    filename = 'cnn_handwrite-acc0.8.tflearn'

    # try to load model and resume training

    #try:

    #    model.load(filename)

    #    print("Model loaded OK. Resume training!")

    #except:

    #    pass

    # Initialize our callback with desired accuracy threshold.

    early_stopping_cb = EarlyStoppingCallback(val_acc_thresh=0.9)

    try:

        model.fit(trainX, trainY, validation_set=(testX, testY), n_epoch=500, shuffle=True,

                  snapshot_epoch=True, # Snapshot (save & evaluate) model every epoch.

                  show_metric=True, batch_size=32, callbacks=early_stopping_cb, run_id='cnn_handwrite')

    except StopIteration as e:

        print("OK, stop iterate!Good!")

    model.save(filename)

    # predict all data and calculate confusion_matrix

    model.load(filename)

    pro_arr =model.predict(X)

    predict_labels = np.argmax(pro_arr, axis=1)

    print(classification_report(org_labels, predict_labels))

    print(confusion_matrix(org_labels, predict_labels))

上述模型效果：

。。。

试试mnist直接拿过来：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Real-time data preprocessing

    img_prep = tflearn.ImagePreprocessing()

    img_prep.add_featurewise_zero_center(per_channel=True)

    network = input_data(shape=[None, width, height, 1]) #, data_preprocessing=img_prep)  # if RGB, 224,224,3

    network = conv_2d(network, 32, 3, activation='relu', regularizer="L2")

    network = max_pool_2d(network, 2)

    network = local_response_normalization(network)

    network = conv_2d(network, 64, 3, activation='relu', regularizer="L2")

    network = max_pool_2d(network, 2)

    network = local_response_normalization(network)

    network = fully_connected(network, 128, activation='tanh')

    network = dropout(network, 0.8)

    network = fully_connected(network, 256, activation='tanh')

    network = dropout(network, 0.8)

    network = fully_connected(network, classes, activation='softmax')

    network = regression(network, optimizer='adam', learning_rate=0.01,

                         loss='categorical_crossentropy', name='target')

    # Training

    model = tflearn.DNN(network, tensorboard_verbose=0)

    return model

模型效果：很难收敛！！！

接下来看看其他模型：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Building 'VGG Network'

    network = input_data(shape=[None, width, height, 1])  # if RGB, 224,224,3

    network = conv_2d(network, 64, 3, activation='relu')

    #network = conv_2d(network, 64, 3, activation='relu')

    network = max_pool_2d(network, 2, strides=2)

    network = conv_2d(network, 128, 3, activation='relu')

    #network = conv_2d(network, 128, 3, activation='relu')

    network = max_pool_2d(network, 2, strides=2)

    netword = tflearn.batch_normalization(network)

    network = fully_connected(network, 1024, activation='relu')

    network = dropout(network, 0.8)

    network = fully_connected(network, classes, activation='softmax')

    network = regression(network, optimizer='rmsprop',

                         loss='categorical_crossentropy',

                         learning_rate=0.0001)

    model = tflearn.DNN(network, checkpoint_path='checkpoint',

                        max_checkpoints=1, tensorboard_verbose=1)

    return model

上述模型效果：

Inception模型：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Building 'VGG Network'

    network = input_data(shape=[None, width, height, 1])  # if RGB, 224,224,3

    network = conv_2d(network, 64, 3, activation='relu')

    inception_3b_1_1 = conv_2d(network, 64, filter_size=1, activation='relu', name='inception_3b_1_1')

    inception_3b_3_3 = conv_2d(network, 64, filter_size=3, activation='relu', name='inception_3b_3_3')

    inception_3b_5_5 = conv_2d(network, 64, filter_size=5, activation='relu', name='inception_3b_5_5')

    inception_3b_output = merge([inception_3b_1_1, inception_3b_3_3, inception_3b_5_5], mode='concat', axis=3, name='inception_3b_output')

    network = max_pool_2d(inception_3b_output, kernel_size=3, strides=2, name='pool3_3_3')

    network = dropout(network, 0.4)

    network = fully_connected(network, classes, activation='softmax')

    network = regression(network, optimizer='momentum',

                     loss='categorical_crossentropy',

                     learning_rate=0.001)

    #network = regression(network, optimizer='rmsprop',

    #                     loss='categorical_crossentropy',

    #                     learning_rate=0.0001)

    model = tflearn.DNN(network, checkpoint_path='checkpoint',

                        max_checkpoints=1, tensorboard_verbose=1)

    return model

上述模型效果：

--

Training Step: 597  | total loss: 4.36442 | time: 342.271ss

| Momentum | epoch: 001 | loss: 4.36442 - acc: 0.0578 | val_loss: 4.30726 - val_acc: 0.1274 -- iter: 19094/19094

--

Terminating training at the end of epoch 1

Training Step: 1193  | total loss: 3.02893 | time: 322.366ss

Training Step: 1194  | total loss: 3.00916 | time: 339.206ser: 19072/19094

| Momentum | epoch: 002 | loss: 3.00916 - acc: 0.2988 | val_loss: 2.71907 - val_acc: 0.4845 -- iter: 19094/19094

--

Terminating training at the end of epoch 2

Training Step: 1791  | total loss: 2.23406 | time: 347.633ss

| Momentum | epoch: 003 | loss: 2.23406 - acc: 0.4559 | val_loss: 1.84004 - val_acc: 0.5888 -- iter: 19094/19094

换成avg pool跑起来很慢：

    #network = max_pool_2d(inception_3b_output, kernel_size=3, strides=2, name='pool3_3_3')

    network = avg_pool_2d(inception_3b_output, kernel_size=7, strides=1) # acc: 0.0217 | val_loss: 4.50712 - val_acc: 0.0630 -- iter: 19094/19094

花费时间长，而且看不到什么效果：

resnet结构：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Building 'VGG Network'

    network = input_data(shape=[None, width, height, 1])  # if RGB, 224,224,3

    # Residual blocks

    # 32 layers: n=5, 56 layers: n=9, 110 layers: n=18

    n = 2

    net = tflearn.conv_2d(network, 16, 3, regularizer='L2', weight_decay=0.0001)

    net = tflearn.residual_block(net, n, 16)

    net = tflearn.residual_block(net, 1, 32, downsample=True)

    net = tflearn.residual_block(net, n-1, 32)

    net = tflearn.residual_block(net, 1, 64, downsample=True)

    net = tflearn.residual_block(net, n-1, 64)

    net = tflearn.batch_normalization(net)

    net = tflearn.activation(net, 'relu')

    net = tflearn.global_avg_pool(net)

    # Regression

    net = tflearn.fully_connected(net, classes, activation='softmax')

    mom = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)

    net = tflearn.regression(net, optimizer=mom,

                             loss='categorical_crossentropy')

    # Training

    model = tflearn.DNN(net, checkpoint_path='model_resnet_cifar10',

                        max_checkpoints=10, tensorboard_verbose=0,

                        clip_gradients=0.)

    return model

--
Terminating training at the end of epoch 7
Training Step: 4776 | total loss: 0.13311 | time: 132.182ss
| Momentum | epoch: 008 | loss: 0.13311 - acc: 0.9561 | val_loss: 0.22734 - val_acc: 0.9370 -- iter: 19094/19094
--
Terminating training at the end of epoch 8
Successfully left training! Final model accuracy: 0.95614439249
OK, stop iterate!Good!
avg / total 0.97 0.96 0.96 23868

resnet加深结构：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Building 'VGG Network'

    network = input_data(shape=[None, width, height, 1])  # if RGB, 224,224,3

    # Building Residual Network

    net = tflearn.conv_2d(network, 64, 3, activation='relu', bias=False)

    # Residual blocks

    net = tflearn.residual_bottleneck(net, 3, 16, 64)

    net = tflearn.residual_bottleneck(net, 1, 32, 128, downsample=True)

    net = tflearn.residual_bottleneck(net, 2, 32, 128)

    net = tflearn.residual_bottleneck(net, 1, 64, 256, downsample=True)

    net = tflearn.residual_bottleneck(net, 2, 64, 256)

    net = tflearn.batch_normalization(net)

    net = tflearn.activation(net, 'relu')

    net = tflearn.global_avg_pool(net)

    # Regression

    net = tflearn.fully_connected(net, classes, activation='softmax')

    net = tflearn.regression(net, optimizer='momentum',

                             loss='categorical_crossentropy',

                             learning_rate=0.1)

    # Training

    model = tflearn.DNN(net, checkpoint_path='model_resnet_mnist',

                        max_checkpoints=10, tensorboard_verbose=0)

    return model

结果是训练的时间更久了。

--
Terminating training at the end of epoch 5
Training Step: 3582 | total loss: 0.14701 | time: 313.084s
| Momentum | epoch: 006 | loss: 0.14701 - acc: 0.9516 | val_loss: 0.30464 - val_acc: 0.9103 -- iter: 19094/19094
--
Terminating training at the end of epoch 6
Successfully left training! Final model accuracy: 0.951571881771
OK, stop iterate!Good!

avg / total 0.94 0.93 0.93 23868

resnet加入预处理：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Real-time data preprocessing

    img_prep = tflearn.ImagePreprocessing()

    img_prep.add_featurewise_zero_center(per_channel=True)

    network = input_data(shape=[None, width, height, 1], data_preprocessing=img_prep)  # if RGB, 224,224,3

    ...

效果：也还是很不错！

--

 Training Step: 597  | total loss: 1.12591 | time: 312.814s

| Momentum | epoch: 001 | loss: 1.12591 - acc: 0.6664 | val_loss: 1.86609 - val_acc: 0.5209 -- iter: 19094/19094

--

Terminating training at the end of epoch 1

 Training Step: 1194  | total loss: 0.61108 | time: 312.415s

| Momentum | epoch: 002 | loss: 0.61108 - acc: 0.8291 | val_loss: 0.56165 - val_acc: 0.8395 -- iter: 19094/19094

highway模型：又快又好！

def get_model(width, height, classes=40):

    # TODO, modify model

    network = input_data(shape=[None, width, height, 1])  # if RGB, 224,224,3

    # Building convolutional network

    #highway convolutions with pooling and dropout

    for i in range(3):

        for j in [3, 2, 1]:

            network = highway_conv_2d(network, 16, j, activation='elu')

        network = max_pool_2d(network, 2)

        network = batch_normalization(network)

    network = fully_connected(network, 128, activation='elu')

    network = fully_connected(network, 256, activation='elu')

    network = fully_connected(network, classes, activation='softmax')

    network = regression(network, optimizer='adam', learning_rate=0.01,

                         loss='categorical_crossentropy', name='target')

    model = tflearn.DNN(network, tensorboard_verbose=0)

    return model

最后看看cifar10模型效果：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Real-time data preprocessing

    img_prep = tflearn.ImagePreprocessing()

    img_prep.add_featurewise_zero_center(per_channel=True)

    img_prep.add_featurewise_stdnorm()

    network = input_data(shape=[None, width, height, 1], data_preprocessing=img_prep)  # if RGB, 224,224,3

    network = conv_2d(network, 32, 3, activation='relu')

    network = max_pool_2d(network, 2)

    network = conv_2d(network, 64, 3, activation='relu')

    network = conv_2d(network, 64, 3, activation='relu')

    network = max_pool_2d(network, 2)

    network = fully_connected(network, 512, activation='relu')

    network = dropout(network, 0.5)

    network = fully_connected(network, classes, activation='softmax')

    network = regression(network, optimizer='adam',

                         loss='categorical_crossentropy',

                         learning_rate=0.001)

    model = tflearn.DNN(network, tensorboard_verbose=0)

    return model

效果也很不错！又快又好：

densenet效果

def get_model(width, height, classes=40):

    # Growth Rate (12, 16, 32, ...)

    k = 12

    # Depth (40, 100, ...)

    L = 40

    nb_layers = int((L - 4) / 3)

    network = input_data(shape=[None, width, height, 1])

    net = tflearn.conv_2d(network, 16, 3, regularizer='L2', weight_decay=0.0001)

    net = tflearn.densenet_block(net, nb_layers, k)

    net = tflearn.densenet_block(net, nb_layers, k)

    net = tflearn.densenet_block(net, nb_layers, k)

    net = tflearn.global_avg_pool(net)

    # Regression

    net = tflearn.fully_connected(net, classes, activation='softmax')

    opt = tflearn.Nesterov(0.1, lr_decay=0.1, decay_step=32000, staircase=True)

    net = tflearn.regression(net, optimizer=opt,

                             loss='categorical_crossentropy')

    # Training

    model = tflearn.DNN(net, checkpoint_path='model_densenet_cifar10',

                        max_checkpoints=10, tensorboard_verbose=0, clip_gradients=0.)

    return model

resnext 模型：

def get_model(width, height, classes=40):

    # TODO, modify model

    # Real-time data preprocessing

    img_prep = tflearn.ImagePreprocessing()

    img_prep.add_featurewise_zero_center(per_channel=True)

    img_prep.add_featurewise_stdnorm()

    network = input_data(shape=[None, width, height, 1], data_preprocessing=img_prep)  # if RGB, 224,224,3

    n = 2

    net = tflearn.conv_2d(network, 16, 3, regularizer='L2', weight_decay=0.0001)

    net = tflearn.resnext_block(net, n, 16, 32)

    net = tflearn.resnext_block(net, 1, 32, 32, downsample=True)

    net = tflearn.resnext_block(net, n-1, 32, 32)

    net = tflearn.resnext_block(net, 1, 64, 32, downsample=True)

    net = tflearn.resnext_block(net, n-1, 64, 32)

    net = tflearn.batch_normalization(net)

    net = tflearn.activation(net, 'relu')

    net = tflearn.global_avg_pool(net)

    # Regression

    net = tflearn.fully_connected(net, classes, activation='softmax')

    opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)

    net = tflearn.regression(net, optimizer=opt,

                             loss='categorical_crossentropy')

    # Training

    model = tflearn.DNN(net, checkpoint_path='model_resnext_cifar10',

                        max_checkpoints=10, tensorboard_verbose=0, clip_gradients=0.)

    return model

tflearn 中文汉字识别模型试验汇总的更多相关文章

tflearn 中文汉字识别，训练后模型存为pb给TensorFlow使用——模型层次太深，或者太复杂训练时候都不会收敛
tflearn 中文汉字识别,训练后模型存为pb给TensorFlow使用. 数据目录在data,data下放了汉字识别图片: data$ ls0 1 10 11 12 13 14 15 ...
Eclipse中文乱码解决汇总（应该比较全）：
Eclipse中文乱码解决汇总(应该比较全,欢迎补充): 方法一: 把GBK改成utf-8. 方法二: Window->preference->general->content ty ...
ST官方翻译的中文应用笔记汇总
ST官方翻译的中文应用笔记汇总 http://www.51hei.com/stm32/3382.html 官方中文AN:AN3116:STM32? 的 ADC 模式及其应用AN1015:用于提高微控制 ...
PTA数据结构与算法题目集(中文) 7-41PAT排名汇总 (25 分)
PTA数据结构与算法题目集(中文) 7-41PAT排名汇总 (25 分) 7-41 PAT排名汇总 (25 分) 计算机程序设计能力考试(Programming Ability Test,简称P ...
苹果Swift语言中文教程资源汇总
苹果swift语言中文教程(零)搭配环境以及代码执行成功http://vjiazhi.com/kaifa/1014.html 苹果Swift语言中文教程(一)基础数据类型 http://vjiazhi ...
一周学会Mootools 1.4中文教程:(7)汇总收尾
转眼之间已经第七课了,这也将成为最后一课,如果这7课下来您感觉水平没有达到预想的水平,没关系您可以继续关注本站的博文,我会陆续发一些类似的文章帮您提升水平,另外我最近打算和群里的几个Mootools爱 ...
转：git windows中文乱码问题解决汇总
it的Windows版本Msysgit对中文的支持不够好 .当使用时,会出现以下三种情况的中文乱码: 下面的几个文件都在git安装目录下文件夹etc内.1.ls不能显示中文目录解决办法:在git/g ...
tomcat7.0在centos7下中文乱码问题解决汇总
1. 系统级的中文设置 Centos7跟之前的6和5安装的中文字符命令都不一样 [root@iZ25bdzgev8Z ~]# locale #查看当前系统的语言环境 [root@iZ25bdzgev8 ...
CloudNotes之桌面客户端篇：插件系统的实现
[CloudNotes版本更新历史与各版本下载地址请点击此处] [CloudNotes中文系列文章汇总列表请点击此处] [查看CloudNotes源代码请点击此处] 有时候,同一个名词,针对不同的人群 ...

随机推荐

【Todo】Java学习路线（方向指导）
在网上搜了下Java学习路线(关键词:学习,因为众所周知,实践出牛人,在平时工作不怎么深入的情况下,才强调学习的方向的重要性 ^_^) 发现下面知乎这个回答写的真好.mark如下: https://w ...
Linux 主机被入侵后的处理案例
Linux主机被入侵后的处理案例提交我的留言加载中已留言一次Linux被入侵后的分析下面通过一个案例介绍下当一个服务器被rootkit入侵后的处理思路和处理过程,rootkit攻击是Lin ...
页面登陆框老是乱乱的？banner跨页图片缩小之后总是在側面不能显示主要部分？哈哈~我来帮你忙~~
有banner背景图片和登陆框的html.css排布目的:无论页面大小,背景图片都要居中(显示图片中间主要内容,而不是側面的一些东西),登陆框基本能在页面内显示. 盒子的排列应该是这种: <d ...
Golang Map Addressability
http://wangzhezhe.github.io/blog/2016/01/22/golangmapaddressability-dot-md/ 在golang中关于map可达性的问题(addr ...
《C程序猿：从校园到职场》出版预告(4)：从“散兵游勇”到“正规部队”
看过电视剧<楚汉传奇>的朋友应该对这个场景还有印象:当刘邦第一次去找项羽帮忙的时候.他们一行人看到了项羽军营是怎样练兵的.想到自己练兵的方法,当时就震惊了."刘家军"就 ...
轻松搞定RabbitMQ（四）——发布/订阅
转自 http://blog.csdn.net/xiaoxian8023/article/details/48729479 翻译地址:http://www.rabbitmq.com/tutorials ...
windows下搭建hadoopproject（一）
这里是接着之前的一篇 <hadoop在windows下的环境搭建 >来的~~~ 一.安装文件准备 1:下载好hadoop-1.0.0.tar.gz, 下载地址是https://archiv ...
JS 模板引擎 Handlebars.js 中文说明
Handlebars 为你提供了一个可以毫无挫折感的高效率书写语义化的模板所必需的一切. Mustache 模板和 Handlebars 是兼容的,所以你可以把Mustache模板拿来导入到Han ...
从头认识java-15.1 填充容器（2）-利用Collection的addAll方式
接着上一章节,我们继续介绍填充容器. 这一章节我们结束还有一种填充容器的方式:addAll 样例: package com.ray.ch15; import java.util.ArrayList; ...
linux命令详解：file命令
前言 file命令可以获取多种文件类型,包括文本文件.脚本文件.源码文件.多媒体文件(音频视频)等.file是通过查看文件的头部内容,来获取文件的类型,而不像Window那样是通过扩展名来确定文件类型 ...

tflearn 中文汉字识别模型试验汇总

tflearn 中文汉字识别模型试验汇总的更多相关文章

随机推荐

热门专题