前言

上一节学习了以TensorFlow为底端的keras接口最简单的使用，这里就继续学习怎么写卷积分类模型和各种保存方法(仅保存权重、权重和网络结构同时保存)

国际惯例，参考博客：
官方教程
【注】其实不用看博客，直接翻到文末看我的colab就行，里面涵盖了学习方法，包括自己的出错内容和一些简单笔记，下面为了展示方便，每次都重新定义了网络结构，对Python熟悉的大佬可以直接def create_model():函数，把模型结构保存起来，后面直接调用就行

构建卷积模型分类

回顾一下上篇博客介绍的构建模型方法，有两种写法：

model = keras.models.Sequential([keras.layers.Flatten(...),keras.layers.Dense(...),...
])

model = keras.models.Sequential()
model.add(keras.layers.Flatten(...))
model.add(keras.layers.Dense(...))
])

第一种简单，第二种舒服，本博文采用第二种写法构建一个简单的卷积网络

引入相关包

保存模型需要路径(引入os)，数据归一化处理(引入numpy)，此外注意，虽然我们学习keras,但是不仅要引入keras，还得引入tensorflow，具体原因后续再说

import tensorflow as tf
from tensorflow import keras
import numpy as np
import matplotlib.pyplot as plt
import os

构建数据集

还是用mnist吧，后续根据需要出一个训练本地图片数据的教程，看看是不是还得数据流操作
注意，要把标签改为单热度编码格式，数据也得归一化

mnist_dataset = keras.datasets.mnist
(train_x,train_y),(test_x,test_y)= mnist_dataset.load_data()
train_y = keras.utils.to_categorical(train_y,10)
test_y = keras.utils.to_categorical(test_y,10)
train_x = train_x / 255.0
test_x = test_x / 255.0

还得注意就是keras的卷积操作接受的数据是一个思维矩阵，需要指定是channels_first即<样本，通道，行，列>, 还是channels_last即<样本，行，列，通道>，默认最后的维度是通道(channels_last)

train_x = train_x[ ..., np.newaxis ]
test_x = test_x[..., np.newaxis ]
print(train_x.shape)#(60000, 28, 28, 1)

构建模型

构建简单的AlexNet,但是直接用这个结构可能有问题，因为输入图片总共28\*28，经过多次卷积池化会越变越小，最后可能都不够做卷积池化了，为稍微改了改

model = keras.models.Sequential()
model.add( keras.layers.Conv2D( filters = 64, kernel_size=(11,11),strides = (1,1), padding='valid', activation= tf.keras.activations.relu) )
model.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model.add( keras.layers.Conv2D( filters = 192, kernel_size=(5,5),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model.add( keras.layers.Conv2D( filters = 256, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))model.add( keras.layers.Flatten() )
model.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model.add( keras.layers.Dropout(rate=0.5) )
model.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model.add( keras.layers.Dropout(rate=0.5) )model.add( keras.layers.Dense(units=10 , activation= keras.activations.softmax ) )

编译和训练模型

在keras中关于交叉熵分类有两个函数sparse_categorical_crossentropy和categorical_crossentropy，这里就出现了第一个坑，如果将标签[batch_size,10]输入到编译器，使用sparse_...的时候回报错

logits and labels must have the same first dimension,got logits shape [200,10] and labels shape [2000]

好像是默认把他拉长拼起来了，所以我们要使用后者

model.compile( optimizer= keras.optimizers.Adam(),loss= keras.losses.categorical_crossentropy, metrics=['accuracy'] )

然后就可以训练模型了

model.fit(train_x,train_y,epochs=2, batch_size=200)
'''
Epoch 1/2
60000/60000 [==============================] - 26s 435us/step - loss: 0.2646 - acc: 0.9110
Epoch 2/2
60000/60000 [==============================] - 24s 407us/step - loss: 0.0510 - acc: 0.9855
<tensorflow.python.keras.callbacks.History at 0x7f65fe2de940>
'''

还能看网络结构和参数，使用summary()函数

model.summary()
'''
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              multiple                  7808      
_________________________________________________________________
max_pooling2d (MaxPooling2D) multiple                  0         
_________________________________________________________________
conv2d_1 (Conv2D)            multiple                  307392    
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 multiple                  0         
_________________________________________________________________
conv2d_2 (Conv2D)            multiple                  663936    
_________________________________________________________________
conv2d_3 (Conv2D)            multiple                  1327488   
_________________________________________________________________
conv2d_4 (Conv2D)            multiple                  884992    
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 multiple                  0         
_________________________________________________________________
flatten (Flatten)            multiple                  0         
_________________________________________________________________
dense (Dense)                multiple                  4198400   
_________________________________________________________________
dropout (Dropout)            multiple                  0         
_________________________________________________________________
dense_1 (Dense)              multiple                  16781312  
_________________________________________________________________
dropout_1 (Dropout)          multiple                  0         
_________________________________________________________________
dense_2 (Dense)              multiple                  40970     
=================================================================
Total params: 24,212,298
Trainable params: 24,212,298
Non-trainable params: 0
_________________________________________________________________
'''

可以用测试集评估模型

print(test_x.shape, test_y.shape)
model.evaluate( test_x ,test_y)
'''
(10000, 28, 28, 1) (10000, 10)
10000/10000 [==============================] - 3s 283us/step
[0.03784323987539392, 0.9897]
'''

还能预测单张图片,但是要注意输入的第一个维度是样本数, 记得增加一个维度

test_img_idx = 1000
test_img = test_x[test_img_idx,...]
test_img= test_img[np.newaxis,...]
img_prob = model.predict( test_img )plt.figure()
plt.imshow( np.squeeze(test_img) )
plt.title(img_prob.argmax())

在这里插入图片描述

保存模型

训练过程中保存检查点

函数为keras.callbacks.ModelCheckpoint

checkpoint_path='./train_save/mnist.ckpt'
checkpoint_dir= os.path.dirname(checkpoint_path)
# 创建检查回调点
cp_callback= keras.callbacks.ModelCheckpoint( checkpoint_path, save_weights_only= True, verbose=1 )model.fit(train_x, train_y,epochs=2, validation_data=(test_x,test_y), callbacks=[cp_callback] )
'''
Train on 60000 samples, validate on 10000 samples
Epoch 1/2
59968/60000 [============================>.] - ETA: 0s - loss: 0.1442 - acc: 0.9681
Epoch 00001: saving model to ./train_save/mnist.ckpt
WARNING:tensorflow:This model was compiled with a Keras optimizer (<tensorflow.python.keras.optimizers.Adam object at 0x7faff48fae80>) but is being saved in TensorFlow format with `save_weights`. The model's weights will be saved, but unlike with TensorFlow optimizers in the TensorFlow format the optimizer's state will not be saved.Consider using a TensorFlow optimizer from `tf.train`.
60000/60000 [==============================] - 93s 2ms/step - loss: 0.1442 - acc: 0.9681 - val_loss: 0.0693 - val_acc: 0.9811
Epoch 2/2
59968/60000 [============================>.] - ETA: 0s - loss: 0.0757 - acc: 0.9840
Epoch 00002: saving model to ./train_save/mnist.ckpt
WARNING:tensorflow:This model was compiled with a Keras optimizer (<tensorflow.python.keras.optimizers.Adam object at 0x7faff48fae80>) but is being saved in TensorFlow format with `save_weights`. The model's weights will be saved, but unlike with TensorFlow optimizers in the TensorFlow format the optimizer's state will not be saved.Consider using a TensorFlow optimizer from `tf.train`.
60000/60000 [==============================] - 92s 2ms/step - loss: 0.0757 - acc: 0.9839 - val_loss: 0.0489 - val_acc: 0.9876
<tensorflow.python.keras.callbacks.History at 0x7fafead25f98>
'''

发现有个warnning，意思是说模型使用的是keras的优化器，保存以后不是tensorflow能直接使用的模型格式，好像少了个状态，需要使用tensorflow自带的优化器，好吧，调整代码

import os
model.compile(optimizer = tf.train.AdamOptimizer(),loss = keras.losses.categorical_crossentropy, metrics=['accuracy'] )checkpoint_path='./train_save2/mnist.ckpt'
checkpoint_dir= os.path.dirname(checkpoint_path)
# 创建检查回调点
cp_callback= keras.callbacks.ModelCheckpoint( checkpoint_path, save_weights_only= True, verbose=1 )model.fit(train_x, train_y,epochs=2, validation_data=(test_x,test_y), callbacks=[cp_callback] )
'''
Train on 60000 samples, validate on 10000 samples
Epoch 1/2
59936/60000 [============================>.] - ETA: 0s - loss: 0.2241 - acc: 0.9325
Epoch 00001: saving model to ./train_save2/mnist.ckpt
60000/60000 [==============================] - 60s 1ms/step - loss: 0.2239 - acc: 0.9326 - val_loss: 0.1009 - val_acc: 0.9765
Epoch 2/2
59936/60000 [============================>.] - ETA: 0s - loss: 0.0866 - acc: 0.9801
Epoch 00002: saving model to ./train_save2/mnist.ckpt
60000/60000 [==============================] - 56s 930us/step - loss: 0.0867 - acc: 0.9800 - val_loss: 0.0591 - val_acc: 0.9855
<tensorflow.python.keras.callbacks.History at 0x7fad1a90d7b8>
'''

这回没出错了，尝试构建一个没训练的模型，将参数载入进来

model_test = keras.models.Sequential()
model_test.add( keras.layers.Conv2D( filters = 64, kernel_size=(11,11),strides = (1,1), padding='valid', activation= tf.keras.activations.relu) )
model_test.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test.add( keras.layers.Conv2D( filters = 192, kernel_size=(5,5),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test.add( keras.layers.Conv2D( filters = 256, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))model_test.add( keras.layers.Flatten() )
model_test.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test.add( keras.layers.Dropout(rate=0.5) )
model_test.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test.add( keras.layers.Dropout(rate=0.5) )model_test.add( keras.layers.Dense(units=10 , activation= keras.activations.softmax ) )model_test.compile(optimizer = tf.train.RMSPropOptimizer(learning_rate=0.01),loss = keras.losses.categorical_crossentropy, metrics=['accuracy'] )

载入最近的检查点

! ls train_save
latest = tf.train.latest_checkpoint('train_save2')#checkpoint  mnist.ckpt.data-00000-of-00001  mnist.ckpt.index
type(latest)#str
loss,acc = model_test.evaluate(test_x,test_y)
print("未载入权重时:准确率{:5.2f}%".format(100*acc))
model_test.load_weights(latest)
loss,acc = model_test.evaluate(test_x,test_y)
print("载入权重时:准确率{:5.2f}%".format(100*acc))
'''
10000/10000 [==============================] - 3s 308us/step
未载入权重时:准确率 9.60%
10000/10000 [==============================] - 3s 264us/step
载入权重时:准确率98.60%
'''

间隔保存

也可以指定多少次训练保存一次检查点，这样能够有效防止过拟合，以后自己可以挑选比较好的训练参数

checkpoint_path="train_save3/cp-{epoch:04d}.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)
cp_callback = keras.callbacks.ModelCheckpoint(checkpoint_path,verbose=1, save_weights_only=True, period=1)
model.fit(train_x,train_y,epochs=2, callbacks=[cp_callback], validation_data=[test_x,test_y],verbose=1)
'''
Train on 60000 samples, validate on 10000 samples
Epoch 1/2
59936/60000 [============================>.] - ETA: 0s - loss: 0.0447 - acc: 0.9897
Epoch 00001: saving model to train_save3/cp-0001.ckpt
60000/60000 [==============================] - 61s 1ms/step - loss: 0.0446 - acc: 0.9897 - val_loss: 0.0421 - val_acc: 0.9920
Epoch 2/2
59936/60000 [============================>.] - ETA: 0s - loss: 0.0478 - acc: 0.9885
Epoch 00002: saving model to train_save3/cp-0002.ckpt
60000/60000 [==============================] - 61s 1ms/step - loss: 0.0478 - acc: 0.9884 - val_loss: 0.0590 - val_acc: 0.9859
<tensorflow.python.keras.callbacks.History at 0x7fafea497dd8>
'''

重新构建一个未训练的模型，调用第一次的训练结果

model_test1 = keras.models.Sequential()
model_test1.add( keras.layers.Conv2D( filters = 64, kernel_size=(11,11),strides = (1,1), padding='valid', activation= tf.keras.activations.relu) )
model_test1.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test1.add( keras.layers.Conv2D( filters = 192, kernel_size=(5,5),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test1.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test1.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test1.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test1.add( keras.layers.Conv2D( filters = 256, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test1.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))model_test1.add( keras.layers.Flatten() )
model_test1.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test1.add( keras.layers.Dropout(rate=0.5) )
model_test1.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test1.add( keras.layers.Dropout(rate=0.5) )model_test1.add( keras.layers.Dense(units=10 , activation= keras.activations.softmax ) )model_test1.compile(optimizer = tf.train.AdamOptimizer(learning_rate=0.01),loss = keras.losses.categorical_crossentropy, metrics=['accuracy'] )

选择第一个检查点载入

loss,acc=model_test1.evaluate(test_x,test_y)
print("未载入权重时:准确率{:5.2f}%".format(100*acc))
model_test1.load_weights("train_save3/cp-0001.ckpt")
loss,acc=model_test1.evaluate(test_x,test_y)
print("载入权重时:准确率{:5.2f}%".format(100*acc))
'''
10000/10000 [==============================] - 3s 260us/step
未载入权重时:准确率10.28%
10000/10000 [==============================] - 2s 244us/step
载入权重时:准确率98.30%
'''

手动保存模型

在训练完毕以后，也可以自行调用save_weights函数保存权重

model.save_weights('./train_save3/mnist_checkpoint')

构建未训练模型

model_test2 = keras.models.Sequential()
model_test2.add( keras.layers.Conv2D( filters = 64, kernel_size=(11,11),strides = (1,1), padding='valid', activation= tf.keras.activations.relu) )
model_test2.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test2.add( keras.layers.Conv2D( filters = 192, kernel_size=(5,5),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test2.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test2.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test2.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test2.add( keras.layers.Conv2D( filters = 256, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test2.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))model_test2.add( keras.layers.Flatten() )
model_test2.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test2.add( keras.layers.Dropout(rate=0.5) )
model_test2.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test2.add( keras.layers.Dropout(rate=0.5) )model_test2.add( keras.layers.Dense(units=10 , activation= keras.activations.softmax ) )model_test2.compile(optimizer = tf.train.AdamOptimizer(learning_rate=0.01),loss = keras.losses.categorical_crossentropy, metrics=['accuracy'] )

读取权重以及评估模型

loss,acc = model_test2.evaluate(test_x,test_y)
print("未载入权重时:准确率{:5.2f}%".format(100*acc))
model_test2.load_weights('./train_save3/mnist_checkpoint')
loss,acc = model_test2.evaluate(test_x,test_y)
print("载入权重时:准确率{:5.2f}%".format(100*acc))
'''
10000/10000 [==============================] - 3s 303us/step
未载入权重时:准确率12.15%
10000/10000 [==============================] - 3s 260us/step
载入权重时:准确率98.59%
'''

全部保存

同时保存模型与参数
构建未训练模型

model_test3 = keras.models.Sequential()
model_test3.add( keras.layers.Conv2D( filters = 64, kernel_size=(11,11),strides = (1,1), padding='valid', activation= tf.keras.activations.relu) )
model_test3.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test3.add( keras.layers.Conv2D( filters = 192, kernel_size=(5,5),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test3.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test3.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test3.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test3.add( keras.layers.Conv2D( filters = 256, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test3.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))model_test3.add( keras.layers.Flatten() )
model_test3.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test3.add( keras.layers.Dropout(rate=0.5) )
model_test3.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test3.add( keras.layers.Dropout(rate=0.5) )model_test3.add( keras.layers.Dense(units=10 , activation= keras.activations.softmax ) )model_test3.compile(optimizer = tf.train.AdamOptimizer(),loss = keras.losses.categorical_crossentropy, metrics=['accuracy'] )
model_test3.fit(train_x,train_y,batch_size=200,epochs=2)

保存

model_test3.save('my_model.h5')
'''
Currently `save` requires model to be a graph network. Consider using `save_weights`, in order to save the weights of the model.
'''

出现错误，意思是需要定义的模型是一个图网络结构，只能保存权重。其实错误原因在于我们的第一层没有定义输入的大小，尝试定义一波

model_test4 = keras.models.Sequential()
model_test4.add( keras.layers.Conv2D( filters = 64, kernel_size=(11,11),strides = (1,1), padding='valid', activation= tf.keras.activations.relu,input_shape=(28,28,1)) )
model_test4.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test4.add( keras.layers.Conv2D( filters = 192, kernel_size=(5,5),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test4.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test4.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test4.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test4.add( keras.layers.Conv2D( filters = 256, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test4.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))model_test4.add( keras.layers.Flatten() )
model_test4.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test4.add( keras.layers.Dropout(rate=0.5) )
model_test4.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test4.add( keras.layers.Dropout(rate=0.5) )model_test4.add( keras.layers.Dense(units=10 , activation= keras.activations.softmax ) )model_test4.compile(optimizer = tf.train.AdamOptimizer(),loss = keras.losses.categorical_crossentropy, metrics=['accuracy'] )
model_test4.fit(train_x,train_y,batch_size=200,epochs=2)
'''
Epoch 1/2
60000/60000 [==============================] - 22s 364us/step - loss: 0.4112 - acc: 0.8556
Epoch 2/2
60000/60000 [==============================] - 21s 342us/step - loss: 0.0584 - acc: 0.9838
<tensorflow.python.keras.callbacks.History at 0x7f137972f4a8>
'''

尝试保存

model_test4.save('my_model.h5')
'''
WARNING:tensorflow:TensorFlow optimizers do not make it possible to access optimizer attributes or optimizer state after instantiation. As a result, we cannot save the optimizer as part of the model save file.You will have to compile your model again after loading it. Prefer using a Keras optimizer instead (see keras.io/optimizers).
'''

又出warning，说是不能使用tensorflow的优化器，要使用keras自带的优化器，好吧，改

model_test5 = keras.models.Sequential()
model_test5.add( keras.layers.Conv2D( filters = 64, kernel_size=(11,11),strides = (1,1), padding='valid', activation= tf.keras.activations.relu,input_shape=(28,28,1)) )
model_test5.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test5.add( keras.layers.Conv2D( filters = 192, kernel_size=(5,5),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test5.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))
model_test5.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test5.add( keras.layers.Conv2D( filters = 384, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test5.add( keras.layers.Conv2D( filters = 256, kernel_size=(3,3),strides = (1,1), padding='same', activation= tf.keras.activations.relu) )
model_test5.add( keras.layers.MaxPool2D( pool_size=(2,2),strides=(2,2) ))model_test5.add( keras.layers.Flatten() )
model_test5.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test5.add( keras.layers.Dropout(rate=0.5) )
model_test5.add( keras.layers.Dense( units=4096, activation= keras.activations.relu ) )
model_test5.add( keras.layers.Dropout(rate=0.5) )model_test5.add( keras.layers.Dense(units=10 , activation= keras.activations.softmax ) )model_test5.compile(optimizer = tf.keras.optimizers.Adam(),loss = keras.losses.categorical_crossentropy, metrics=['accuracy'] )
model_test5.fit(train_x,train_y,batch_size=200,epochs=2)
model_test5.save("my_model2.h5")
'''
Epoch 1/2
60000/60000 [==============================] - 26s 434us/step - loss: 0.2850 - acc: 0.9043
Epoch 2/2
60000/60000 [==============================] - 25s 409us/step - loss: 0.0555 - acc: 0.9847
<tensorflow.python.keras.callbacks.History at 0x7f661033a9e8>
'''

这样就不出错了，尝试调用模型和参数，因为保存了模型结构和参数，所以不需要重新定义网络结构

model_test6= keras.models.load_model("my_model2.h5")
model_test6.summary()
'''
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_10 (Conv2D)           (None, 18, 18, 64)        7808      
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 9, 9, 64)          0         
_________________________________________________________________
conv2d_11 (Conv2D)           (None, 9, 9, 192)         307392    
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 4, 4, 192)         0         
_________________________________________________________________
conv2d_12 (Conv2D)           (None, 4, 4, 384)         663936    
_________________________________________________________________
conv2d_13 (Conv2D)           (None, 4, 4, 384)         1327488   
_________________________________________________________________
conv2d_14 (Conv2D)           (None, 4, 4, 256)         884992    
_________________________________________________________________
max_pooling2d_8 (MaxPooling2 (None, 2, 2, 256)         0         
_________________________________________________________________
flatten_2 (Flatten)          (None, 1024)              0         
_________________________________________________________________
dense_6 (Dense)              (None, 4096)              4198400   
_________________________________________________________________
dropout_4 (Dropout)          (None, 4096)              0         
_________________________________________________________________
dense_7 (Dense)              (None, 4096)              16781312  
_________________________________________________________________
dropout_5 (Dropout)          (None, 4096)              0         
_________________________________________________________________
dense_8 (Dense)              (None, 10)                40970     
=================================================================
Total params: 24,212,298
Trainable params: 24,212,298
Non-trainable params: 0
_________________________________________________________________
'''

稳如狗，做一下测试

测试集上的测试

model_test6.evaluate(test_x,test_y)
'''
10000/10000 [==============================] - 3s 302us/step
[0.05095283883444499, 0.9868]
'''

单张图片的测试

test_img= test_x[5000,...]
test_img=test_img[ np.newaxis,...]
pred_label = model_test6.predict_classes(test_img)
plt.figure()
plt.imshow( np.squeeze(test_img))
plt.title(pred_label)

在这里插入图片描述

后记

这一章主要学习了如何搭建简单的卷积网络，以及集中保存方法：仅权重以及权重和模型结构。
主要记住的就是如果仅保存权重，注意用tensorflow自带的优化器，而保存网络和权重的时候要用keras的优化器
下一章针对深度学习的几个理论做一下理解以及实验，包括BatchNorm、ResNet等。
博客代码链接：戳这里