锐客网

  1. 首页 > 睿知 > it技术 > >

TensorFlow中的CIFAR-10和CIFAR-100数据集

tensorflowIT知识数据集TensorFlow项目

本文概述

  • CIFAR-100数据集
  • 用例：使用TensorFlow在卷积神经网络的帮助下实现CIFAR10
  • 训练时间
CIFAR-10(加拿大高级研究所)和CIFAR-100被标记为8000万个微型图像数据集的子集。它们由Alex Krizhevsky, Geoffrey Hinton和Vinod Nair收集。数据集分为五个训练批次和一个测试批次, 每个批次具有10000张图像。
【TensorFlow中的CIFAR-10和CIFAR-100数据集】测试批次包含每个类别的1000张随机选择的图像。训练批次以随机顺序包含剩余图像, 但是一些训练批次以随机顺序包含剩余图像, 但是一些训练批次包含从一个类别到另一类别的更多图像。在它们之间, 每个批次的培训批次正好包含5000张图像。
TensorFlow中的CIFAR-10和CIFAR-100数据集

文章图片
这些类将完全互斥。汽车和卡车之间不会重叠。汽车包括类似于轿车和越野车的东西。卡车类仅包括大型卡车, 也不包括皮卡车。如果我们通过CIFAR数据集查看, 我们就会意识到这不仅仅是鸟或猫的一种。鸟和猫类包含许多不同类型的鸟和猫。鸟和猫类提供许多种类的鸟和猫, 它们的大小, 颜色, 放大率, 不同的角度和不同的姿势都不同。
对于无尽的数据集, 我们可以通过多种方式来编写第一和第二。它只是没有那么多样化, 最重要的是, 无尽的数据集是一个灰度标量。 CIFAR数据集包含32张乘32张彩色图像, 每张照片具有三个不同的色彩通道。现在我们最重要的问题是, 在无尽的数据集上表现如此出色的LeNet模型是否足以对CIFAR数据集进行分类？
CIFAR-100数据集 就像CIFAR-10数据集一样。唯一的区别是它有100个类, 每个类包含600个图像。每个班级有100张测试图像和500张训练图像。这100个类别分为20个超类, 每个图像带有一个” 粗” 标签(它所属的超类), 一个” 精细” 标签(它所属的类)和一个” 精细” 标签(该类)它所属的)。
CIFAR-100数据集中的以下类：
S.没有 Superclass Classes
1. Flowers 兰花, 罂粟, 玫瑰, 向日葵, 郁金香
2. Fish 水族馆鱼, 比目鱼, 雷, 鲨鱼, 鳟鱼
3. Aquatic mammals 海狸, 海豚, 水獭, 海豹, 鲸鱼
4. food containers 瓶子, 碗, 罐, 杯子, 盘子
5. 家用电器 时钟, 灯, 电话, 电视, 电脑键盘
6. 水果和蔬菜 苹果, 蘑菇, 橘子, 梨, 甜椒
7. Household furniture 桌子, 椅子, 沙发, 衣柜, 床,
8. 昆虫蜜蜂, 甲虫, 蝴蝶, 毛毛虫, 蟑螂
9. 大型自然户外场景 云, 森林, 山, 平原, 海洋
10. 大型人造户外用品 桥梁, 城堡, 房屋, 道路, 摩天大楼
11. 大食肉动物 熊, 豹, 狮子, 老虎, 狼
12. 中型哺乳动物 狐狸, 豪猪, 负鼠, 浣熊, 臭鼬
13. 大型杂食动物和食草动物 骆驼, 牛, 黑猩猩, 大象, 袋鼠
14. 非昆虫无脊椎动物 螃蟹, 龙虾, 蜗牛, 蜘蛛, 蠕虫
15. reptiles 鳄鱼, 恐龙, 蜥蜴, 蛇, 乌龟
16. trees 枫木, 橡木, 棕榈, 松木, 柳树
17. people 女孩, 男人, 女人, 婴儿, 男孩
18. Small mammals 仓鼠, 兔子, 老鼠, sh, 松鼠
19. 车辆1 自行车, 公共汽车, 摩托车, 皮卡车, 火车
20. Vehicles 2 割草机, 火箭, 电车, 拖拉机, 坦克
用例：使用TensorFlow在卷积神经网络的帮助下实现CIFAR10 现在, 训练一个网络, 以使用内置的卷积神经网络TensorFlow对来自CIFAR10数据集的图像进行分类。
TensorFlow中的CIFAR-10和CIFAR-100数据集

文章图片
考虑以下流程图, 以了解用例的工作原理：
图
安装必要的软件包：
pip3 install numpy tensorflow pickle

训练网络：
import numpy as npimport tensorflow as tffrom time import timeimport mathfrom include .data import get_data_setfrom include.model import model, lrtrain_x, train_y= get_data_set("train")test_x, test_y = get_data_set("test")tf. set_random_seed(21)x, y, output, y_pred_cls, global_step, learning_rate=model()global_accuracy =0epoch_start=0#PARAM_BATCH_SIZE=128_EPOCH=60_SAVE_PATH="./tensorboard/cifar-10-v1.0.0/"#LOSS AND OPTIMIZERloss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=output, labels=y))optimizer=tf.train.AdamOptimizer(learning_rate= learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-08). Minimize(loss, global_step=global_step)#PREDICTION AND ACCURACY CALCULATIONcorrect_prediction=tf.equal(y_pred_cls, tf.argmax(y, axis=1))accuracy = tf.reduce_mean(tf.cast(correct_predictiction, tf.float32))# SAVERmerged = tf.summary.merge_all()saver = tf.train.Saver()sess = tf.Session()train_writer= tf.summary.FileWriter(_SAVE_PATH, sess.graph)try:print(" Trying to restore last checkpoint?")last_chk_path= tf.train.latest_checkpoint(checkpoint_dir=SAVE_PATH)saver.restore(sess, save_path=last_chk_path)print("Restored checkpoint from:", last_chk_path)except ValueError:print("Failed to restore checkpoint. Initializing variable instead.")sess.run(tf.global_variables_initializer())def train(epoch):global epoch_startepoch_start= time()batch_size=int(math.ceil(len(train_x)/_BATCH_SIZE))i_global = 0for s in range(batch_size):batch_xs= train_x[s*_BATCH_SIZE: (s+1)*_BATCH_SIZE]batch_ys = train_y[s*_BATCH_SIZE: (s+1)*_BATCH_SIZE]start_time= time()i_global, _, batch_loss, batch_acc=sess.run( [global_step, optimizer, loss, accuracy], feed_dict={x: batch_xs, y: batch_ys, learning_rate: lr(epoch)})duration = time() - start_timeif s% 10== 0:percentage = int(round((s/batch_size)*100))bar_len=29filled_len= int ((bar_len*int(percentage))/100)bar='=' *filled_len + ?> ' + ?-? * (bar_len _filled_len)msg= "Global step: { :> 5} - [{}] {:> 3}% -acc: {:.{:> 4f} - loss: {:.4f} -{:.1f} sample/sec"print(msg.format(i_global, bar, percentage, batch_acc, batch_loss, _BATCH_SIZE/duration))test_and_save(i_global, epoch)def test_and_save(_global_step, epoch):global global_accuracyglobal epoch_starti=0predicted_class=np.zeroes(shape=len(test_x), dtype=np.int)while i< len (test_x) : j=min(i+_BATCH_SIZE, len(test_x)) batch_xs=test_x[I:j, :] batch_ys=test_y[i:j, :] predicted_class[i:j]=sess.run(y_pred_cls, feed_dict=x:batch_xs, y: batch_ys, learning_rate: lr(epoch)} ) i=j correct= (np.argmax(test_y, axis=1) == predicted_class) acc = correct.mean()*100 correct_numbers = correct.sum() hours, rem = divmod(time() - epoch_start, 3600) minutes, seconds = divmod(rem, 60) mes = "Epoch {} - accuracy: {: .2f}% ({}/{})- time: {:0> 2}:{:0.2}:{:05.2f}"print(mes.format((epoch+1), acc, correct_numbers, len(test_x), int(hours), int(minutes), seconds))if global_accuracy != 0 and global_accuracy < acc: summary = tf.Summary(value=http://www.srcmini.com/[ tf.Summary.Value(tag="Accuracy/test", simple_value=http://www.srcmini.com/acc), ]) train_writer.add_summary(summary, _global_step) saver.save(sess, save_path=_SAVE_PATH, global_step=_global_step) mes ="This epoch receive better accuracy: {:.2f} > {:.2f}. Saving session... print(mes.format((acc, global_accuracy))global_accuracy = accelif global_accuracy==0:global_accuracy=accprint("################################################################def main():train_start=time()for i in range(_EPOCH):print(" Epoch: {}/{}".format(( i+1), _EPOCH))train(i)hours, rem=divmod(time()-train_start, 3600 minutes, seconds=divmod(rem, 60)mes= "Best accuracy per session: {:.2f}, time: {:0> 2}:{:0> 2}:{:05.2f}"print(mes.format(global_accuracy, int(hours), int(minutes), seconds))if _name_ =="_main_":main()sess.close()

输出
Epoch: 60/60Global step: 23070 - [> -----------------------------]0% - acc: 0.9531 - loss: 1.5081 - 7045.4 sample/secGlobal step: 23080 - [> -----------------------------]3% - acc: 0.9453 - loss: 1.5159 - 7147.6 sample/secGlobal step: 23090 - [=> ----------------------------]5% - acc: 0.9844 - loss: 1.4764 - 7154.6 sample/secGlobal step: 23100 - [==> ---------------------------]8% - acc: 0.9297 - loss: 1.5307 - 7104.4 sample/secGlobal step: 23110 - [==> ---------------------------]10% - acc: 0.9141 - loss: 1.5462 - 7091.4 sample/secGlobal step: 23120 - [===> --------------------------]13% - acc: 0.9297 - loss: 1.5314 - 7162.9 sample/secGlobal step: 23130 - [====> -------------------------]15% - acc: 0.9297 - loss: 1.5307 - 7174.8 sample/secGlobal step: 23140 - [=====> ------------------------]18% - acc: 0.9375 - loss: 1.5231 - 7140.0 sample/secGlobal step: 23150 - [=====> ------------------------]20% - acc: 0.9297 - loss: 1.5301 - 7152.8 sample/secGlobal step: 23160 - [======> -----------------------]23% - acc: 0.9531 - loss: 1.5080 - 7112.3 sample/secGlobal step: 23170 - [=======> ----------------------]26% - acc: 0.9609 - loss: 1.5000 - 7154.0 sample/secGlobal step: 23180 - [========> ---------------------]28% - acc: 0.9531 - loss: 1.5074 - 6862.2 sample/secGlobal step: 23190 - [========> ---------------------]31% - acc: 0.9609 - loss: 1.4993 - 7134.5 sample/secGlobal step: 23200 - [=========> --------------------]33% - acc: 0.9609 - loss: 1.4995 - 7166.0 sample/secGlobal step: 23210 - [==========> -------------------]36% - acc: 0.9375 - loss: 1.5231 - 7116.7 sample/secGlobal step: 23220 - [===========> ------------------]38% - acc: 0.9453 - loss: 1.5153 - 7134.1 sample/secGlobal step: 23230 - [===========> ------------------]41% - acc: 0.9375 - loss: 1.5233 - 7074.5 sample/secGlobal step: 23240 - [============> -----------------]43% - acc: 0.9219 - loss: 1.5387 - 7176.9 sample/secGlobal step: 23250 - [=============> ----------------]46% - acc: 0.8828 - loss: 1.5769 - 7144.1 sample/secGlobal step: 23260 - [==============> ---------------]49% - acc: 0.9219 - loss: 1.5383 - 7059.7 sample/secGlobal step: 23270 - [==============> ---------------]51% - acc: 0.8984 - loss: 1.5618 - 6638.6 sample/secGlobal step: 23280 - [===============> --------------]54% - acc: 0.9453 - loss: 1.5151 - 7035.7 sample/secGlobal step: 23290 - [================> -------------]56% - acc: 0.9609 - loss: 1.4996 - 7129.0 sample/secGlobal step: 23300 - [=================> ------------]59% - acc: 0.9609 - loss: 1.4997 - 7075.4 sample/secGlobal step: 23310 - [=================> ------------]61% - acc: 0.8750 - loss:1.5842 - 7117.8 sample/secGlobal step: 23320 - [==================> -----------]64% - acc: 0.9141 - loss:1.5463 - 7157.2 sample/secGlobal step: 23330 - [===================> ----------]66% - acc: 0.9062 - loss: 1.5549 - 7169.3 sample/secGlobal step: 23340 - [====================> ---------]69% - acc: 0.9219 - loss: 1.5389 - 7164.4 sample/secGlobal step: 23350 - [====================> ---------]72% - acc: 0.9609 - loss: 1.5002 - 7135.4 sample/secGlobal step: 23360 - [=====================> --------]74% - acc: 0.9766 - loss: 1.4842 - 7124.2 sample/secGlobal step: 23370 - [======================> -------]77% - acc: 0.9375 - loss: 1.5231 - 7168.5 sample/secGlobal step: 23380 - [======================> -------]79% - acc: 0.8906 - loss: 1.5695 - 7175.2 sample/secGlobal step: 23390 - [=======================> ------]82% - acc: 0.9375 - loss: 1.5225 - 7132.1 sample/secGlobal step: 23400 - [========================> -----]84% - acc: 0.9844 - loss: 1.4768 - 7100.1 sample/secGlobal step: 23410 - [=========================> ----]87% - acc: 0.9766 - loss: 1.4840 - 7172.0 sample/secGlobal step: 23420 - [==========================> ---]90% - acc: 0.9062 - loss: 1.5542 - 7122.1 sample/secGlobal step: 23430 - [==========================> ---]92% - acc: 0.9297 - loss: 1.5313 - 7145.3 sample/secGlobal step: 23440 - [===========================> --]95% - acc: 0.9297 - loss: 1.5301 - 7133.3 sample/secGlobal step: 23450 - [============================> -]97% - acc: 0.9375 - loss: 1.5231 - 7135.7 sample/secGlobal step: 23460 - [=============================> ] 100% - acc: 0.9250 - loss: 1.5362 - 10297.5 sample/secEpoch 60 - accuracy: 78.81% (7881/10000)This epoch receive better accuracy: 78.81 > 78.78. Saving session...##################################################################################################

在测试数据集上运行网络：
import numpy as npimport tensorflow as tffrom include.data import get_data_setfrom include.model import modeltest_x, test_y= get_data_set("test")x, y, output, y_pred_cls, global_step, learning_rate =model()_BATCH_SIZE = 128_CLASS_SIZE = 10_SAVE_PATH = "./tensorboard/cifar-10-v1.0.0/"saver= tf.train.Saver()Sess=tf.Session()try; print(" Trying to restore last checkpoint ...")last_chk_path = tf.train.latest_checkpoint(checkpoint_dir=_SAVE_PATHsaver.restore(sess, save_path=last_chk_path)print("Restored checkpoint from:", last_chk_path)expect ValueError:print("Failed to restore checkpoint. Initializing variables instead.")sess.run(tf.global_variables_initializer())def main():i=0predicted_class= np.zeros(shape=len(test_x), dtype=np.int)while i< lens(test_x):j=min(i+_BATCH_SIZE, len(test_x))batch_xs=test_x[i:j, :]batch_xs=test_y[i:j, :]pre dicted_class[i:j] = sess.run(y_pred_cls, feed_dict={x: batch_xs, y: batch_ys})i=jcorr ect = (np.argmax(test_y, axis=1) == predicted_class)acc=correct.mean()*100correct_numbers=correct.sum()print()print("Accuracy is on Test-Set: {0:.2f}% ({1} / {2})".format(acc, correct_numbers, len(test_x)))if__name__=="__main__":main()sess.close()

简单输出
Trying to restore last checkpoint ...Restored checkpoint from: ./tensorboard/cifar-10-v1.0.0/-23460Accuracy on Test-Set: 78.81% (7881 / 10000)

TensorFlow中的CIFAR-10和CIFAR-100数据集

文章图片
训练时间 在这里, 我们看到60个纪元需要多少时间：
Device Batch Time Accuracy[%]
NVidia 128 8m4s 79.12
Inteli77700HQ 128 3h30m 78.91

    推荐阅读


    上一篇：融云 x DSPORT（拿下游戏社交「实时社区」第一滴血）

    下一篇：在TensorFlow中引入卷积神经网络图解