一，准备数据

cifar2数据集为cifar10数据集的子集，只包括前两种类别airplane和automobile。
训练集有airplane和automobile图片各5000张，测试集有airplane和automobile图片各1000张。
cifar2任务的目标是训练一个模型来对飞机airplane和机动车automobile两种图片进行分类。
我们准备的Cifar2数据集的文件结构如下所示。
image.png

在tensorflow中准备图片数据的常用方案有两种，第一种是使用tf.keras中的ImageDataGenerator工具构建图片数据生成器。

第二种是使用tf.data.Dataset搭配tf.image中的一些图片处理方法构建数据管道。
第一种方法更为简单，其使用范例可以参考以下文章。
https://zhuanlan.zhihu.com/p/67466552

第二种方法是TensorFlow的原生方法，更加灵活，使用得当的话也可以获得更好的性能。
我们此处介绍第二种方法。

  1. import tensorflow as tf 
  2. from tensorflow.keras import datasets,layers,models
  4. BATCH_SIZE = 100
  6. def load_image(img_path,size = (32,32)):
  7.     label = tf.constant(1,tf.int8) if tf.strings.regex_full_match(img_path,".*automobile.*") \
  8.             else tf.constant(0,tf.int8)
  9.     img = tf.io.read_file(img_path)
  10.     img = tf.image.decode_jpeg(img) #注意此处为jpeg格式
  11.     img = tf.image.resize(img,size)/255.0
  12.     return(img,label)
  1. #使用并行化预处理num_parallel_calls 和预存数据prefetch来提升性能
  2. ds_train = tf.data.Dataset.list_files("./data/cifar2/train/*/*.jpg") \
  3.            .map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE) \
  4.            .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
  5.            .prefetch(tf.data.experimental.AUTOTUNE)  
  7. ds_test = tf.data.Dataset.list_files("./data/cifar2/test/*/*.jpg") \
  8.            .map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE) \
  9.            .batch(BATCH_SIZE) \
  10.            .prefetch(tf.data.experimental.AUTOTUNE)
  1. %matplotlib inline
  2. %config InlineBackend.figure_format = 'svg'
  4. #查看部分样本
  5. from matplotlib import pyplot as plt 
  7. plt.figure(figsize=(8,8)) 
  8. for i,(img,label) in enumerate(ds_train.unbatch().take(9)):
  9.     ax=plt.subplot(3,3,i+1)
  10.     ax.imshow(img.numpy())
  11.     ax.set_title("label = %d"%label)
  12.     ax.set_xticks([])
  13.     ax.set_yticks([]) 
  14. plt.show()

image.png

  1. for x,y in ds_train.take(1):
  2.     print(x.shape,y.shape)
  1. (100, 32, 32, 3) (100,)

二，定义模型

使用Keras接口有以下3种方式构建模型：使用Sequential按层顺序构建模型，使用函数式API构建任意结构模型，继承Model基类构建自定义模型。

此处选择使用函数式API构建模型。

  1. tf.keras.backend.clear_session() #清空会话
  3. inputs = layers.Input(shape=(32,32,3))
  4. x = layers.Conv2D(32,kernel_size=(3,3))(inputs)
  5. x = layers.MaxPool2D()(x)
  6. x = layers.Conv2D(64,kernel_size=(5,5))(x)
  7. x = layers.MaxPool2D()(x)
  8. x = layers.Dropout(rate=0.1)(x)
  9. x = layers.Flatten()(x)
  10. x = layers.Dense(32,activation='relu')(x)
  11. outputs = layers.Dense(1,activation = 'sigmoid')(x)
  13. model = models.Model(inputs = inputs,outputs = outputs)
  15. model.summary()
  1. Model: "model"
  2. _________________________________________________________________
  3. Layer (type)                 Output Shape              Param #   
  4. =================================================================
  5. input_1 (InputLayer)         [(None, 32, 32, 3)]       0         
  6. _________________________________________________________________
  7. conv2d (Conv2D)              (None, 30, 30, 32)        896       
  8. _________________________________________________________________
  9. max_pooling2d (MaxPooling2D) (None, 15, 15, 32)        0         
  10. _________________________________________________________________
  11. conv2d_1 (Conv2D)            (None, 11, 11, 64)        51264     
  12. _________________________________________________________________
  13. max_pooling2d_1 (MaxPooling2 (None, 5, 5, 64)          0         
  14. _________________________________________________________________
  15. dropout (Dropout)            (None, 5, 5, 64)          0         
  16. _________________________________________________________________
  17. flatten (Flatten)            (None, 1600)              0         
  18. _________________________________________________________________
  19. dense (Dense)                (None, 32)                51232     
  20. _________________________________________________________________
  21. dense_1 (Dense)              (None, 1)                 33        
  22. =================================================================
  23. Total params: 103,425
  24. Trainable params: 103,425
  25. Non-trainable params: 0
  26. _________________________________________________________________

三，训练模型

训练模型通常有3种方法，内置fit方法，内置train_on_batch方法，以及自定义训练循环。此处我们选择最常用也最简单的内置fit方法。

  1. import datetime
  2. import os
  4. stamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
  5. logdir = os.path.join('data', 'autograph', stamp)
  7. ## 在 Python3 下建议使用 pathlib 修正各操作系统的路径
  8. # from pathlib import Path
  9. # stamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
  10. # logdir = str(Path('./data/autograph/' + stamp))
  12. tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
  14. model.compile(
  15.         optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
  16.         loss=tf.keras.losses.binary_crossentropy,
  17.         metrics=["accuracy"]
  18.     )
  20. history = model.fit(ds_train,epochs= 10,validation_data=ds_test,
  21.                     callbacks = [tensorboard_callback],workers = 4)
  1. Train for 100 steps, validate for 20 steps
  2. Epoch 1/10
  3. 100/100 [==============================] - 16s 156ms/step - loss: 0.4830 - accuracy: 0.7697 - val_loss: 0.3396 - val_accuracy: 0.8475
  4. Epoch 2/10
  5. 100/100 [==============================] - 14s 142ms/step - loss: 0.3437 - accuracy: 0.8469 - val_loss: 0.2997 - val_accuracy: 0.8680
  6. Epoch 3/10
  7. 100/100 [==============================] - 13s 131ms/step - loss: 0.2871 - accuracy: 0.8777 - val_loss: 0.2390 - val_accuracy: 0.9015
  8. Epoch 4/10
  9. 100/100 [==============================] - 12s 117ms/step - loss: 0.2410 - accuracy: 0.9040 - val_loss: 0.2005 - val_accuracy: 0.9195
  10. Epoch 5/10
  11. 100/100 [==============================] - 13s 130ms/step - loss: 0.1992 - accuracy: 0.9213 - val_loss: 0.1949 - val_accuracy: 0.9180
  12. Epoch 6/10
  13. 100/100 [==============================] - 14s 136ms/step - loss: 0.1737 - accuracy: 0.9323 - val_loss: 0.1723 - val_accuracy: 0.9275
  14. Epoch 7/10
  15. 100/100 [==============================] - 14s 139ms/step - loss: 0.1531 - accuracy: 0.9412 - val_loss: 0.1670 - val_accuracy: 0.9310
  16. Epoch 8/10
  17. 100/100 [==============================] - 13s 134ms/step - loss: 0.1299 - accuracy: 0.9525 - val_loss: 0.1553 - val_accuracy: 0.9340
  18. Epoch 9/10
  19. 100/100 [==============================] - 14s 137ms/step - loss: 0.1158 - accuracy: 0.9556 - val_loss: 0.1581 - val_accuracy: 0.9340
  20. Epoch 10/10
  21. 100/100 [==============================] - 14s 142ms/step - loss: 0.1006 - accuracy: 0.9617 - val_loss: 0.1614 - val_accuracy: 0.9345

四，评估模型

  1. %load_ext tensorboard
  2. #%tensorboard --logdir ./data/keras_model
  1. from tensorboard import notebook
  2. notebook.list()
  1. #在tensorboard中查看模型
  2. notebook.start("--logdir ./data/keras_model")

image.png

  1. import pandas as pd 
  2. dfhistory = pd.DataFrame(history.history)
  3. dfhistory.index = range(1,len(dfhistory) + 1)
  4. dfhistory.index.name = 'epoch'
  6. dfhistory

image.png

  1. %matplotlib inline
  2. %config InlineBackend.figure_format = 'svg'
  4. import matplotlib.pyplot as plt
  6. def plot_metric(history, metric):
  7.     train_metrics = history.history[metric]
  8.     val_metrics = history.history['val_'+metric]
  9.     epochs = range(1, len(train_metrics) + 1)
  10.     plt.plot(epochs, train_metrics, 'bo--')
  11.     plt.plot(epochs, val_metrics, 'ro-')
  12.     plt.title('Training and validation '+ metric)
  13.     plt.xlabel("Epochs")
  14.     plt.ylabel(metric)
  15.     plt.legend(["train_"+metric, 'val_'+metric])
  16.     plt.show()
  1. plot_metric(history,"loss")

image.png

  1. plot_metric(history,"accuracy")

image.png

  1. #可以使用evaluate对数据进行评估
  2. val_loss,val_accuracy = model.evaluate(ds_test,workers=4)
  3. print(val_loss,val_accuracy)
  1. 0.16139143370091916 0.9345

五，使用模型

可以使用model.predict(ds_test)进行预测。

也可以使用model.predict_on_batch(x_test)对一个批量进行预测。

  1. model.predict(ds_test)
  1. array([[9.9996173e-01],
  2.        [9.5104784e-01],
  3.        [2.8648047e-04],
  4.        ...,
  5.        [1.1484033e-03],
  6.        [3.5589080e-02],
  7.        [9.8537153e-01]], dtype=float32)
  1. for x,y in ds_test.take(1):
  2.     print(model.predict_on_batch(x[0:20]))
  1. tf.Tensor(
  2. [[3.8065155e-05]
  3.  [8.8236779e-01]
  4.  [9.1433197e-01]
  5.  [9.9921846e-01]
  6.  [6.4052093e-01]
  7.  [4.9970779e-03]
  8.  [2.6735585e-04]
  9.  [9.9842811e-01]
  10.  [7.9198682e-01]
  11.  [7.4823302e-01]
  12.  [8.7208226e-03]
  13.  [9.3951421e-03]
  14.  [9.9790359e-01]
  15.  [9.9998581e-01]
  16.  [2.1642199e-05]
  17.  [1.7915063e-02]
  18.  [2.5839690e-02]
  19.  [9.7538447e-01]
  20.  [9.7393811e-01]
  21.  [9.7333014e-01]], shape=(20, 1), dtype=float32)

六，保存模型

推荐使用TensorFlow原生方式保存模型。

  1. # 保存权重，该方式仅仅保存权重张量
  2. model.save_weights('./data/tf_model_weights.ckpt',save_format = "tf")
  1. # 保存模型结构与模型参数到文件,该方式保存的模型具有跨平台性便于部署
  3. model.save('./data/tf_model_savedmodel', save_format="tf")
  4. print('export saved model.')
  6. model_loaded = tf.keras.models.load_model('./data/tf_model_savedmodel')
  7. model_loaded.evaluate(ds_test)
  1. [0.16139124035835267, 0.9345]