【深度学习 走进tensorflow2.0】训练模型以及保存最佳模型

无意中发现了一个巨牛的人工智能教程，忍不住分享一下给大家。教程不仅是零基础，通俗易懂，而且非常风趣幽默，像看小说一样！觉得太牛了，所以分享给大家。点这里可以跳转到教程。人工智能教程

项目目录： 数据集： 下载 二分类数据集：cats_and_dogs_filtered文件夹

wget https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip

目录结构如下：

. ├── [drwxr-x---] cats_and_dogs_filtered │ ├── [drwxr-x---] train │ │ ├── [drwxr-x---] cats │ │ └── [drwxr-x---] dogs │ └── [drwxr-x---] validation │ ├── [drwxr-x---] cats │ └── [drwxr-x---] dogs

1、数据预处理 2、数据增强 2、创建模型 4、编译模型 5、模型自动保存定义 6、模型训练 7、模型保存回调以及自动调节学习率回调

模型保存：

├── [-rw-rw-r--] model ├── [-rw-rw-r--] model_class.json ├── [-rw-rw-r--] model_ex-001_acc-0.500000.h5 ├── [-rw-rw-r--] model_ex-002_acc-0.501116.h5 ├── [-rw-rw-r--] model_ex-004_acc-0.527902.h5 ├── [-rw-rw-r--] model_ex-005_acc-0.540179.h5 ├── [-rw-rw-r--] model_ex-006_acc-0.549107.h5 ├── [-rw-rw-r--] model_ex-007_acc-0.600446.h5 ├── [-rw-rw-r--] model_ex-008_acc-0.646205.h5 ├── [-rw-rw-r--] model_ex-010_acc-0.648438.h5 ├── [-rw-rw-r--] model_ex-013_acc-0.650670.h5 └── [-rw-rw-r--] model_ex-014_acc-0.672991.h5

关键点： 1、训练量很少时，通常会发生过度拟合。解决此问题的一种方法是扩充数据集，使其具有足够数量的训练示例。数据增强采用通过使用产生真实感图像的随机变换增强样本来从现有训练样本生成更多训练数据的方法。目标是模型在训练期间永远不会看到两次完全相同的图片。这有助于使模型暴露于数据的更多方面，并且可以更好地进行概括。 tf.keras使用ImageDataGenerator类来实现这一点。将不同的转换传递给数据集，它将在训练过程中加以应用。

2、tf.keras.callbacks.ModelCheckpoint函数 自动保存模型

tf.keras.callbacks.ModelCheckpoint 参数： filepath：string，保存模型文件的路径。 monitor：要监测的数量。 verbose：详细信息模式，0或1。 save_best_only：如果save_best_only=True，被监测数量的最佳型号不会被覆盖。 mode：{auto，min，max}之一。如果save_best_only=True，那么是否覆盖保存文件的决定就取决于被监测数据的最大或者最小值。对于val_acc，这应该是max，对于val_loss这应该是min，等等。在auto模式中，方向是从监测数量的名称自动推断出来的。 save_weights_only：如果为True，则仅保存模型的权重（model.save_weights(filepath)），否则保存完整模型（model.save(filepath)）。 period：检查点之间的间隔（epoch数）。

3、TensorFlow回调函数：tf.keras.callbacks.ReduceLROnPlateau 当指标停止提升时，降低学习速率。一旦学习停止，模型通常会将学习率降低2-10倍。该回调监测数量，如果没有看到epoch的 ‘patience’ 数量的改善，那么学习率就会降低。

reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=5, min_lr=0.001) model.fit(X_train, Y_train, callbacks=[reduce_lr]) 参数： monitor：要监测的数量。 factor：学习速率降低的因素。new_lr = lr * factor patience：没有提升的epoch数，之后学习率将降低。 verbose：int。0：安静，1：更新消息。 mode：{auto，min，max}之一。在min模式下，当监测量停止下降时，lr将减少；在max模式下，当监测数量停止增加时，它将减少；在auto模式下，从监测数量的名称自动推断方向。 min_delta：对于测量新的最优化的阀值，仅关注重大变化。 cooldown：在学习速率被降低之后，重新恢复正常操作之前等待的epoch数量。 min_lr：学习率的下限。

4、模型训练

使用fit_generator方法ImageDataGenerator来训练网络。

history = model.fit_generator( train_data_gen, steps_per_epoch=int(num_train / batch_size), epochs=epochs, validation_data=val_data_gen, validation_steps=int(num_test / batch_size), callbacks=[checkpoint,lr_scheduler])

完整训练代码：

# -*- coding: utf-8 -*- from __future__ import absolute_import, division, print_function, unicode_literals import tensorflow as tf from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense, Conv2D, Flatten, Dropout, MaxPooling2D from tensorflow.keras.preprocessing.image import ImageDataGenerator import json import os import numpy as np batch_size = 128 epochs = 15 IMG_HEIGHT = 150 IMG_WIDTH = 150 PATH = os.path.join('/home/dongli/tensorflow2.0/corpus/', 'cats_and_dogs_filtered') train_dir = os.path.join(PATH, 'train') validation_dir = os.path.join(PATH, 'validation') train_cats_dir = os.path.join(train_dir, 'cats') # directory with our training cat pictures train_dogs_dir = os.path.join(train_dir, 'dogs') # directory with our training dog pictures validation_cats_dir = os.path.join(validation_dir, 'cats') # directory with our validation cat pictures validation_dogs_dir = os.path.join(validation_dir, 'dogs') # directory with our validation dog pictures num_cats_tr = len(os.listdir(train_cats_dir)) num_dogs_tr = len(os.listdir(train_dogs_dir)) num_cats_val = len(os.listdir(validation_cats_dir)) num_dogs_val = len(os.listdir(validation_dogs_dir)) total_train = num_cats_tr + num_dogs_tr total_val = num_cats_val + num_dogs_val print('total training cat images:', num_cats_tr) print('total training dog images:', num_dogs_tr) print('total validation cat images:', num_cats_val) print('total validation dog images:', num_dogs_val) print("--") print("Total training images:", total_train) print("Total validation images:", total_val) # 训练集 # 对训练图像应用了重新缩放，45度旋转，宽度偏移，高度偏移，水平翻转和缩放增强。 image_gen_train = ImageDataGenerator( rescale=1./255, rotation_range=45, width_shift_range=.15, height_shift_range=.15, horizontal_flip=True, zoom_range=0.5 ) train_data_gen = image_gen_train.flow_from_directory(batch_size=batch_size, directory=train_dir, shuffle=True, target_size=(IMG_HEIGHT, IMG_WIDTH), class_mode='binary') # 验证集 image_gen_val = ImageDataGenerator(rescale=1./255) val_data_gen = image_gen_val.flow_from_directory(batch_size=batch_size, directory=validation_dir, target_size=(IMG_HEIGHT, IMG_WIDTH), class_mode='binary') # 创建模型 model = Sequential([ Conv2D(16, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH ,3)), MaxPooling2D(), Dropout(0.2), Conv2D(32, 3, padding='same', activation='relu'), MaxPooling2D(), Conv2D(64, 3, padding='same', activation='relu'), MaxPooling2D(), Dropout(0.2), Flatten(), Dense(512, activation='relu'), Dense(1, activation='sigmoid') ]) # 编译模型 model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy']) # 模型总结 model.summary() # 模型保存格式定义 model_class_dir='./model/' class_indices = train_data_gen.class_indices class_json = {} for eachClass in class_indices: class_json[str(class_indices[eachClass])] = eachClass with open(os.path.join(model_class_dir, "model_class.json"), "w+") as json_file: json.dump(class_json, json_file, indent=4, separators=(",", " : "),ensure_ascii=True) json_file.close() print("JSON Mapping for the model classes saved to ", os.path.join(model_class_dir, "model_class.json")) model_name = 'model_ex-{epoch:03d}_acc-{val_accuracy:03f}.h5' trained_model_dir='./model/' model_path = os.path.join(trained_model_dir, model_name) checkpoint = tf.keras.callbacks.ModelCheckpoint( filepath=model_path, monitor='val_accuracy', verbose=1, save_weights_only=True, save_best_only=True, mode='max', period=1) def lr_schedule(epoch): # Learning Rate Schedule lr =1e-3 total_epochs =epochs check_1 = int(total_epochs * 0.9) check_2 = int(total_epochs * 0.8) check_3 = int(total_epochs * 0.6) check_4 = int(total_epochs * 0.4) if epoch > check_1: lr *= 1e-4 elif epoch > check_2: lr *= 1e-3 elif epoch > check_3: lr *= 1e-2 elif epoch > check_4: lr *= 1e-1 return lr #lr_scheduler =tf.keras.callbacks.LearningRateScheduler(lr_schedule) lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2,patience=5, min_lr=0.001) num_train = len(train_data_gen.filenames) num_test = len(val_data_gen.filenames) print(num_train,num_test) # 模型训练 # 使用fit_generator方法ImageDataGenerator来训练网络。 history = model.fit_generator( train_data_gen, steps_per_epoch=int(num_train / batch_size), epochs=epochs, validation_data=val_data_gen, validation_steps=int(num_test / batch_size), callbacks=[checkpoint,lr_scheduler] )

控制台输出：

Epoch 1/15 2019-10-31 15:07:47.955641: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7 2019-10-31 15:07:49.213954: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0 14/15 [===========================>..] - ETA: 0s - loss: 1.1185 - accuracy: 0.5206 15/15 [==============================] - 14s 924ms/step - loss: 1.0906 - accuracy: 0.5246 - val_loss: 0.6935 - val_accuracy: 0.5000 Epoch 2/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6936 - accuracy: 0.4911 15/15 [==============================] - 13s 884ms/step - loss: 0.6935 - accuracy: 0.4948 - val_loss: 0.6932 - val_accuracy: 0.5011 Epoch 3/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6910 - accuracy: 0.5230 15/15 [==============================] - 12s 809ms/step - loss: 0.6926 - accuracy: 0.5175 - val_loss: 0.6894 - val_accuracy: 0.5000 Epoch 4/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6914 - accuracy: 0.5039 15/15 [==============================] - 12s 830ms/step - loss: 0.6915 - accuracy: 0.5031 - val_loss: 0.6919 - val_accuracy: 0.5279 Epoch 5/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6893 - accuracy: 0.5472 15/15 [==============================] - 11s 764ms/step - loss: 0.6891 - accuracy: 0.5482 - val_loss: 0.6893 - val_accuracy: 0.5402 Epoch 6/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6788 - accuracy: 0.5603 15/15 [==============================] - 12s 802ms/step - loss: 0.6784 - accuracy: 0.5641 - val_loss: 0.6866 - val_accuracy: 0.5491 Epoch 7/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6789 - accuracy: 0.5608 15/15 [==============================] - 12s 801ms/step - loss: 0.6778 - accuracy: 0.5641 - val_loss: 0.6702 - val_accuracy: 0.6004 Epoch 8/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6691 - accuracy: 0.5991 15/15 [==============================] - 12s 781ms/step - loss: 0.6681 - accuracy: 0.5976 - val_loss: 0.6632 - val_accuracy: 0.6462 Epoch 9/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6635 - accuracy: 0.6021 15/15 [==============================] - 12s 812ms/step - loss: 0.6606 - accuracy: 0.6052 - val_loss: 0.6562 - val_accuracy: 0.5982 Epoch 10/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6541 - accuracy: 0.6124 15/15 [==============================] - 12s 811ms/step - loss: 0.6512 - accuracy: 0.6165 - val_loss: 0.6373 - val_accuracy: 0.6484 Epoch 11/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6329 - accuracy: 0.6439 15/15 [==============================] - 12s 799ms/step - loss: 0.6334 - accuracy: 0.6437 - val_loss: 0.6395 - val_accuracy: 0.6406 Epoch 12/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6477 - accuracy: 0.6085 15/15 [==============================] - 11s 758ms/step - loss: 0.6498 - accuracy: 0.6047 - val_loss: 0.6416 - val_accuracy: 0.6183 Epoch 13/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6280 - accuracy: 0.6507 15/15 [==============================] - 12s 825ms/step - loss: 0.6277 - accuracy: 0.6505 - val_loss: 0.6114 - val_accuracy: 0.6507 Epoch 14/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6160 - accuracy: 0.6548 15/15 [==============================] - 12s 795ms/step - loss: 0.6147 - accuracy: 0.6565 - val_loss: 0.6012 - val_accuracy: 0.6730 Epoch 15/15 14/15 [===========================>..] - ETA: 0s - loss: 0.6163 - accuracy: 0.6680 15/15 [==============================] - 12s 793ms/step - loss: 0.6124 - accuracy: 0.6709 - val_loss: 0.6073 - val_accuracy: 0.6629