环境
tensorflow 2.1
最好用GPU
Cifar10数据集
CIFAR-10 数据集的分类是机器学习中一个公开的基准测试问题。任务的目标对一组32x32 RGB的图像进行分类,这个数据集涵盖了10个类别:飞机, 汽车, 鸟, 猫, 鹿, 狗, 青蛙, 马, 船以及卡车。
下面代码仅仅只是做显示Cifar10数据集用
import numpy as npimport matplotlib.pyplot as pltimport tensorflow as tfdef showPic(X_train, y_train):# 看看数据集中的一些样本:每个类别展示一些classes = ['plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']num_classes = len(classes)samples_per_class = 7for y, cls in enumerate(classes):idxs = np.flatnonzero(y_train == y)# 一个类别中挑出一些idxs = np.random.choice(idxs, samples_per_class, replace=False)for i, idx in enumerate(idxs):plt_idx = i * num_classes + y + 1plt.subplot(samples_per_class, num_classes, plt_idx)plt.imshow(X_train[idx].astype('uint8'))plt.axis('off')if i == 0:plt.title(cls)plt.show()if __name__ == '__main__':(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()showPic(x_train, y_train)
模型
Resnet:把前一层的数据直接加到下一层里。减少数据在传播过程中过多的丢失。
SENet: 学习每一层的通道之间的关系
Inception: 每一层都用不同的核(1x1,3x3,5x5)来学习.防止因为过小的核或者过大的核而学不到图片的特征。
用Resnet ,SENet, Inceptiont网络训练Cifar10 或者Cifar 100.
训练数据:Cifar10 或者 Cifar 100
训练集上准确率:97.11%左右
验证集上准确率:90.22%左右
测试集上准确率:88.6%
训练时间在GPU上:一小时多
权重大小:21.8 MB
训练的历程
普通网络(65%左右)-> 数据增强(70%左右)->模型增强(进入Resnet 和SEnet) 80%左右 ->模型的结构做了调整(86% )- > 加inception 模型后(88.6%)
训练集与验证集上的测试结果:
347/351 [============================>.] - ETA: 0s - loss: 0.0827 - sparse_categorical_accuracy: 0.9711348/351 [============================>.] - ETA: 0s - loss: 0.0828 - sparse_categorical_accuracy: 0.9710349/351 [============================>.] - ETA: 0s - loss: 0.0827 - sparse_categorical_accuracy: 0.9710350/351 [============================>.] - ETA: 0s - loss: 0.0826 - sparse_categorical_accuracy: 0.9711351/351 [==============================] - 20s 57ms/step - loss: 0.0826 - sparse_categorical_accuracy: 0.9711 - val_loss: 0.3952 - val_sparse_categorical_accuracy: 0.9022
测试集上的结果
79/79 - 4s - loss: 0.4005 - sparse_categorical_accuracy: 0.8842[0.40052215495630156, 0.8842]time 4.130274534225464
下面是完整的代码,运行前建一下这个目录weights4_5,不想写代码自动化建了。
如果要训练Cifar100,直接把cifar10 改成cifar100就可以了。不需要改其它地方
import tensorflow as tfimport tensorflow.keras as kerasimport tensorflow.keras.layers as layersimport time as timeimport tensorflow.keras.preprocessing.image as imageimport matplotlib.pyplot as pltimport osdef senet_block(inputs, ratio):shape = inputs.shapechannel_out = shape[-1]# print(shape)# (2, 28, 28, 32) , [1,28,28,1], [1,28,28,1]squeeze = layers.GlobalAveragePooling2D()(inputs)# [2, 1, 1, 32]# print(squeeze.shape)# 第二层,全连接层# [2,32]# print(squeeze.shape)shape_result = layers.Flatten()(squeeze)# print(shape_result.shape)# [32,2]shape_result = layers.Dense(int(channel_out / ratio), activation='relu')(shape_result)# shape_result = layers.BatchNormalization()(shape_result)# [2,32]shape_result = layers.Dense(channel_out, activation='sigmoid')(shape_result)# shape_result = layers.BatchNormalization()(shape_result)# 第四层,点乘# print('heres2')excitation_output = tf.reshape(shape_result, [-1, 1, 1, channel_out])# print(excitation_output.shape)h_output = excitation_output * inputsreturn h_outputdef inception_block(input, input_filter, output_filter):reception_filter = output_filterprint('reception_filter',reception_filter)r1 = layers.Conv2D(filters=reception_filter, kernel_size=(1, 1), activation='relu', padding='same')(input)r1 = layers.BatchNormalization()(r1 )r1 = senet_block(r1, 8)r3 = layers.Conv2D(filters=reception_filter, kernel_size=(3, 3), activation='relu', padding='same')(input)r3 = layers.BatchNormalization()(r3 )r3 = senet_block(r3, 8)r5 = layers.Conv2D(filters=reception_filter, kernel_size=(5, 5), activation='relu', padding='same')(input)r5 = layers.BatchNormalization()(r5 )r5 = senet_block(r5, 8)mx = tf.keras.layers.MaxPool2D(pool_size=(3, 3), strides=1, padding="same")(input)mx = layers.Conv2D(filters=reception_filter, kernel_size=(1, 1), activation='relu', padding='same')(mx)output = tf.keras.layers.concatenate([r1, r3, r5,mx], axis=-1)output = layers.Conv2D(filters=reception_filter, kernel_size=(3, 3), activation='relu', padding='same')(output)print('output',output.shape)return outputdef res_block(input, input_filter, output_filter):# input = 32, output = 96res_x = inception_block(input, input_filter, output_filter)# 96res_x = layers.Conv2D(filters=output_filter, kernel_size=(3, 3), activation=None, padding='same')(res_x )res_x = layers.BatchNormalization()(res_x )res_x = senet_block(res_x, 8)if input_filter == output_filter:identity = inputelse: #需要升维或者降维identity = layers.Conv2D(filters=output_filter, kernel_size=(1,1), padding='same')(input)print(identity.shape)print(res_x.shape)x = layers.Add()([identity, res_x])output = layers.Activation('relu')(x)return outputdef my_model():inputs = keras.Input(shape=(32,32,3), name='img')h1 = layers.Conv2D(filters=16, kernel_size=(3, 3), strides=(1, 1), padding='same', activation='relu')(inputs)h1 = layers.BatchNormalization()(h1)h1 = senet_block(h1, 8)block1_out = res_block(h1, 16, 32)block2_out = layers.MaxPool2D(pool_size=(2, 2))(block1_out)block2_out = res_block(block2_out, 32,64)block3_out = layers.MaxPool2D(pool_size=(2, 2))(block2_out)block3_out = res_block(block3_out, 64,128)block4_out = layers.MaxPool2D(pool_size=(2, 2))(block3_out)block4_out = res_block(block4_out, 128,256)h3 = layers.GlobalAveragePooling2D()(block4_out)h3 = layers.Flatten()(h3)h3 = layers.BatchNormalization()(h3)h3 = layers.Dense(64, activation='relu')(h3)h3 = layers.BatchNormalization()(h3)outputs = layers.Dense(10, activation='softmax')(h3)deep_model = keras.Model(inputs, outputs, name='resnet')pile(optimizer=keras.optimizers.Adam(),loss=keras.losses.SparseCategoricalCrossentropy(),#metrics=['accuracy'])metrics=[keras.metrics.SparseCategoricalAccuracy()])deep_model.summary()#keras.utils.plot_model(deep_model, 'my_resNet.png', show_shapes=True)return deep_modelcurrent_max_loss = 9999weight_file='./weights4_5/model.h5'def train_my_model(deep_model):(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()train_datagen = image.ImageDataGenerator(rescale=1 / 255,rotation_range=40, # 角度值,0-180.表示图像随机旋转的角度范围width_shift_range=0.2, # 平移比例,下同height_shift_range=0.2,shear_range=0.2, # 随机错切变换角度zoom_range=0.2, # 随即缩放比例horizontal_flip=True, # 随机将一半图像水平翻转fill_mode='nearest' # 填充新创建像素的方法)test_datagen = image.ImageDataGenerator(rescale=1 / 255)validation_datagen = image.ImageDataGenerator(rescale=1 / 255)train_generator = train_datagen.flow(x_train[:45000], y_train[:45000], batch_size=128)# train_generator = train_datagen.flow(x_train, y_train, batch_size=128)validation_generator = validation_datagen.flow(x_train[45000:], y_train[45000:], batch_size=128)test_generator = test_datagen.flow(x_test, y_test, batch_size=128)begin_time = time.time()if os.path.isfile(weight_file):print('load weight')deep_model.load_weights(weight_file)def save_weight(epoch, logs):global current_max_lossif(logs['val_loss'] is not None and logs['val_loss']< current_max_loss):current_max_loss = logs['val_loss']print('save_weight', epoch, current_max_loss)deep_model.save_weights(weight_file)batch_print_callback = keras.callbacks.LambdaCallback(on_epoch_end=save_weight)callbacks = [tf.keras.callbacks.EarlyStopping(patience=4, monitor='loss'),batch_print_callback,# keras.callbacks.ModelCheckpoint('./weights/model.h5', save_best_only=True),tf.keras.callbacks.TensorBoard(log_dir='logs4_4')]print(train_generator[0][0].shape)history = deep_model.fit_generator(train_generator, steps_per_epoch=351, epochs=200, callbacks=callbacks,validation_data=validation_generator, validation_steps=39, initial_epoch = 0)if (history.history['val_loss'] is not None and history.history['val_loss'] < current_max_loss):current_max_loss = history['val_loss']print('save_weight', current_max_loss)deep_model.save_weights(weight_file)result = deep_model.evaluate_generator(test_generator, verbose=2)print(result)print('time', time.time() - begin_time)def show_result(history):plt.plot(history.history['loss'])plt.plot(history.history['val_loss'])plt.plot(history.history['sparse_categorical_accuracy'])plt.plot(history.history['val_sparse_categorical_accuracy'])plt.legend(['loss', 'val_loss', 'sparse_categorical_accuracy', 'val_sparse_categorical_accuracy'],loc='upper left')plt.show()print(history)show_result(history)def test_module(deep_model):(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()test_datagen = image.ImageDataGenerator(rescale=1 / 255)test_generator = test_datagen.flow(x_test, y_test, batch_size=128)begin_time = time.time()if os.path.isfile(weight_file):print('load weight')deep_model.load_weights(weight_file)result = deep_model.evaluate_generator(test_generator, verbose=2)print(result)print('time', time.time() - begin_time)def predict_module(deep_model):(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()import numpy as npif os.path.isfile(weight_file):print('load weight')deep_model.load_weights(weight_file)print(y_test[0:20])for i in range(20):img = x_test[i][np.newaxis, :]/255y_ = deep_model.predict(img)v = np.argmax(y_)print(v, y_test[i])if __name__ == '__main__':deep_model = my_model()train_my_model(deep_model)#test_module(deep_model)#predict_module(deep_model)
参考训练结果
执行下面命令,访问http://localhost:6006/查看训练过程中,准确率和损失函数变化过程,
tensorboard --logdir=logs4_4
注意点
训练完后,如果你要测试你的训练好的模型就把test_module(deep_model)代码注释打开
if __name__ == '__main__':deep_model = my_model()#train_my_model(deep_model)test_module(deep_model)#predict_module(deep_model)
如果你要进一步预测你的模型就把predict_module(deep_model)代码注释打开
if __name__ == '__main__':deep_model = my_model()#train_my_model(deep_model)#test_module(deep_model)predict_module(deep_model)
如果觉得《[深度学习-TF2实践]应用Tensorflow2.x训练ResNet SeNet和Inception模型在cifar10 测试集上准确率88.6%》对你有帮助,请点赞、收藏,并留下你的观点哦!