深度学习MLP/LeNet/AlexNet/GoogLeNet/ResNet在三个不同数据集上的分类效果实践腾讯云开发者社区|hwdb数据集_彩票

实验结果如下表所示模型在不同数据集上的准确度

MNIST

FashionMNIST

HWDB1

MLP

97.76%

87.73%

84.17%

LeNet

98.68%

85.82%

91.33%

AlexNet

98.91%

90.57%

89.67%

GoogLeNet

99.27%

90.27%

91.50%

ResNet

99.21%

91.35%

93.67%

1.激活函数使用relu

2.卷积之后引入标准化层(BN层)

3.使用了Dropout防止过拟合

inception模块结构图：

GoogLeNet创新点：

v1:引入了Inception结构,并使用1x1卷积和来压缩通道数（减少参数量。Inception作用：代替人工确定卷积层中的过滤器类型或者确定是否需要创建卷积层和池化层，即：不需要人为的决定使用哪个过滤器，是否需要池化层等，由网络自行决定这些参数，可以给网络添加所有可能值，将输出连接起来，网络自己学习它需要什么样的参数。

v2：引入BN层，BN作用：加速网络训练/防止梯度消失。

v3：(1)将Inception内部的BN层推广到外部。(2)优化了网络结构，将较大的二维卷积拆成两个较小的一维卷积，比如将3x3拆成1x3和3x1。这样节省了大量参数，加速运算，并增加了一层非线性表达能力。

v4：引入残差结构。

num_epochs=10batch_size=50learning_rate=0.001model=MLP()data_loader=MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))测试——MLP-Mnist准确率：97.76%

sparse_categorical_accuracy=tf.keras.metrics.SparseCategoricalAccuracy()num_batches=int(data_loader.num_test_data//batch_size)forbatch_indexinrange(num_batches):start_index,end_index=batch_index*\batch_size,(batch_index+1)*batch_sizey_pred=model.predict(data_loader.test_data[start_index:end_index])sparse_categorical_accuracy.update_state(y_true=data_loader.test_label[start_index:end_index],y_pred=y_pred)print("testaccuracy:%f"%sparse_categorical_accuracy.result())代码语言：javascript复制testaccuracy:0.977600训练——MLP-Fashion_MNIST进行10轮训练

num_epochs=10batch_size=50learning_rate=0.001model=MLP()data_loader=Fashion_MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):print("epoch%d"%(e+1))forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)print("batch%d:loss%f"%(batch_index,loss.numpy()))grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))测试——MLP-Fashion_MNIST准确率：87.73%

num_epochs=100batch_size=50learning_rate=0.001model=MLP()data_loader=HWDB1Loader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))测试——MLP-HWDB1准确率：84.17%

num_epochs=10batch_size=50learning_rate=0.001model=LeNet()data_loader=MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))测试——LeNet-Mnist准确率：98.68%

sparse_categorical_accuracy=tf.keras.metrics.SparseCategoricalAccuracy()num_batches=int(data_loader.num_test_data//batch_size)forbatch_indexinrange(num_batches):start_index,end_index=batch_index*\batch_size,(batch_index+1)*batch_sizey_pred=model.predict(data_loader.test_data[start_index:end_index])sparse_categorical_accuracy.update_state(y_true=data_loader.test_label[start_index:end_index],y_pred=y_pred)print("testaccuracy:%f"%sparse_categorical_accuracy.result())代码语言：javascript复制testaccuracy:0.986800训练——LeNet-Fashion_MNIST进行10轮训练

num_epochs=10batch_size=50learning_rate=0.001model=LeNet()data_loader=Fashion_MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))测试——LeNet-Fashion_MNIST准确率：85.82%

num_epochs=100batch_size=50learning_rate=0.001model=LeNet()data_loader=HWDB1Loader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))测试——LeNet-HWDB1准确率：91.33%

num_epochs=10batch_size=50learning_rate=0.001model=AlexNet()data_loader=MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss0.067972epoch2:loss0.008088epoch3:loss0.002195epoch4:loss0.000097epoch5:loss0.008648epoch6:loss0.002917epoch7:loss0.024030epoch8:loss0.000092epoch9:loss0.003540epoch10:loss0.003601测试——AlexNet-Mnist准确率：98.91%

sparse_categorical_accuracy=tf.keras.metrics.SparseCategoricalAccuracy()num_batches=int(data_loader.num_test_data//batch_size)forbatch_indexinrange(num_batches):start_index,end_index=batch_index*\batch_size,(batch_index+1)*batch_sizey_pred=model.predict(data_loader.test_data[start_index:end_index])sparse_categorical_accuracy.update_state(y_true=data_loader.test_label[start_index:end_index],y_pred=y_pred)print("testaccuracy:%f"%sparse_categorical_accuracy.result())代码语言：javascript复制testaccuracy:0.989100训练——AlexNet-Fashion_MNIST进行10轮训练

num_epochs=10batch_size=50learning_rate=0.001model=AlexNet()data_loader=Fashion_MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss0.512883epoch2:loss0.372691epoch3:loss0.317794epoch4:loss0.209880epoch5:loss0.259238epoch6:loss0.246677epoch7:loss0.325049epoch8:loss0.085606epoch9:loss0.231111epoch10:loss0.108106测试——AlexNet-Fashion_MNIST准确率:90.57%

num_epochs=10batch_size=50learning_rate=0.001model=AlexNet()data_loader=HWDB1Loader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss2.154262epoch2:loss1.550519epoch3:loss1.117614epoch4:loss0.796395epoch5:loss0.482055epoch6:loss0.261059epoch7:loss0.598217epoch8:loss0.408028epoch9:loss0.527755epoch10:loss0.232926测试——AlexNet-HWDB1准确率：89.67%

num_epochs=10batch_size=50learning_rate=0.001model=Inception(num_blocks=2,num_classes=10)data_loader=MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss0.226245epoch2:loss0.046722epoch3:loss0.043534epoch4:loss0.005478epoch5:loss0.005200epoch6:loss0.020319epoch7:loss0.087346epoch8:loss0.008595epoch9:loss0.035517epoch10:loss0.025345测试——GoogLeNet-Mnist准确率：99.27%

sparse_categorical_accuracy=tf.keras.metrics.SparseCategoricalAccuracy()num_batches=int(data_loader.num_test_data//batch_size)forbatch_indexinrange(num_batches):start_index,end_index=batch_index*\batch_size,(batch_index+1)*batch_sizey_pred=model.predict(data_loader.test_data[start_index:end_index])sparse_categorical_accuracy.update_state(y_true=data_loader.test_label[start_index:end_index],y_pred=y_pred)print("testaccuracy:%f"%sparse_categorical_accuracy.result())代码语言：javascript复制testaccuracy:0.992700训练——GoogLeNet-Fashion_MNIST进行10轮训练

num_epochs=10batch_size=50learning_rate=0.001model=Inception(num_blocks=2,num_classes=10)data_loader=Fashion_MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss0.555443epoch2:loss0.369227epoch3:loss0.424934epoch4:loss0.332159epoch5:loss0.254867epoch6:loss0.216396epoch7:loss0.330004epoch8:loss0.130716epoch9:loss0.225775epoch10:loss0.072902测试——GoogLeNet-Fashion_MNIST准确率：90.27%

num_epochs=10batch_size=50learning_rate=0.001model=Inception(num_blocks=2,num_classes=10)data_loader=HWDB1Loader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss2.175095epoch2:loss1.401912epoch3:loss1.073104epoch4:loss0.582406epoch5:loss0.561393epoch6:loss0.342880epoch7:loss0.225490epoch8:loss0.534456epoch9:loss0.142884epoch10:loss0.083469测试——GoogLeNet-HWDB1准确率:91.50%

num_epochs=10batch_size=50learning_rate=0.001model=ResNet18([2,2,2,2])data_loader=MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss0.009949epoch2:loss0.026236epoch3:loss0.016156epoch4:loss0.000139epoch5:loss0.003694epoch6:loss0.004449epoch7:loss0.001600epoch8:loss0.001913epoch9:loss0.001122epoch10:loss0.005298测试——ResNet-Mnist准确率：99.21%

sparse_categorical_accuracy=tf.keras.metrics.SparseCategoricalAccuracy()num_batches=int(data_loader.num_test_data//batch_size)forbatch_indexinrange(num_batches):start_index,end_index=batch_index*\batch_size,(batch_index+1)*batch_sizey_pred=model.predict(data_loader.test_data[start_index:end_index])sparse_categorical_accuracy.update_state(y_true=data_loader.test_label[start_index:end_index],y_pred=y_pred)print("testaccuracy:%f"%sparse_categorical_accuracy.result())代码语言：javascript复制testaccuracy:0.992100训练——ResNet-Fashion_MNIST进行10轮训练

num_epochs=10batch_size=50learning_rate=0.001model=ResNet18([2,2,2,2])data_loader=Fashion_MNISTLoader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):print("epoch%d"%(e+1))forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)print("batch%d:loss%f"%(batch_index,loss.numpy()))grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))测试——ResNet-Fashion_MNIST准确率：91.35%

num_epochs=10batch_size=50learning_rate=0.001model=ResNet18([2,2,2,2])data_loader=HWDB1Loader()optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate)num_batches=int(data_loader.num_train_data//batch_size)foreinrange(num_epochs):forbatch_indexinrange(num_batches):X,y=data_loader.get_batch(batch_size)withtf.GradientTape()astape:y_pred=model(X)loss=tf.keras.losses.sparse_categorical_crossentropy(y_true=y,y_pred=y_pred)loss=tf.reduce_mean(loss)grads=tape.gradient(loss,model.variables)optimizer.apply_gradients(grads_and_vars=zip(grads,model.variables))print("epoch%d:loss%f"%(e+1,loss.numpy()))代码语言：javascript复制epoch1:loss2.286184epoch2:loss1.170036epoch3:loss0.718102epoch4:loss0.766253epoch5:loss0.271660epoch6:loss0.121803epoch7:loss0.105720epoch8:loss0.029223epoch9:loss0.422725epoch10:loss0.005966

THE END

深度学习MLP/LeNet/AlexNet/GoogLeNet/ResNet在三个不同数据集上的分类效果实践腾讯云开发者社区

深度学习MLP/LeNet/AlexNet/GoogLeNet/ResNet在三个不同数据集上的分类效果实践腾讯云开发者社区

hwdb中文手册[金步国]

Master笔记22613@英文论文书写@DETR@ViT@CLIP@汉字手写数据集@SwinT@StarGANNoNoe

中国科学数据

新型的端到端弱监督篇幅级手写中文文本识别方法PageNet

同盾AI研究院深度学习首席专家李宏宇：解构知识联邦，开创数据“可用不可见”新局面雷峰网

经典论文复现LSGAN：最小二乘生成对抗网络