深度学习TF—5.tf.kears高层API

最新推荐文章于 2022-03-21 10:45:09 发布
最新推荐文章于 2022-03-21 10:45:09 发布
阅读量285 收藏 2
点赞数 2
分类专栏： # 深度学习-Tensorflow 文章标签： tf.keras
本BLOG上原创文章未经本人许可，不得用于商业用途，转载请注明出处。
本文链接：https://blog.csdn.net/weixin_46649052/article/details/114004234
版权
本文详细介绍了tf.keras的使用，包括metrics的实战、模型的编译、训练、评估与预测，以及如何自定义层和网络，如Sequential、Model、Layer。此外，还探讨了模型的加载与保存，包括weights、整个模型的保存和save_model的使用。
摘要由CSDN通过智能技术生成

文章目录

一、metrics

  • 新建一个评价指标
acc_meter = metrics.Accuracy()
loss_meter = metrics.Mean()
  • update data- 添加数据
loss_meter.update_state(loss)
acc_meter.update_state(y,pred)
  • result().numpy()-显示结果
print(step,'loss:', loss_meter.result().numpy())
...
print(step,'Evaluate Acc:',total_correct/total, acc_meter.result().numpy())
  • reset_states()-清零
if step % 100 == 0:
	print(step,'loss:', loss_meter.result().numpy())
	# 清除上一个时间戳的数据
	loss_meter.reset_states()

if step % 500 == 0:
	print(step,'Evaluate Acc:',total_correct/total, acc_meter.result().numpy())
	acc_meter.reset_states()
1.实战
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics

# 预处理函数
def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)

    return x, y


batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())

db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz).repeat(10)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz)

# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(10)])
network.build(input_shape=(None, 28 * 28))
network.summary()

# 优化器
optimizer = optimizers.Adam(lr=0.01)

# 评价指标-acc,loss
acc_meter = metrics.Accuracy()
loss_meter = metrics.Mean()


for step, (x, y) in enumerate(db):

    with tf.GradientTape() as tape:
        # [b, 28, 28] => [b, 784]
        x = tf.reshape(x, (-1, 28 * 28))
        # [b, 784] => [b, 10]
        out = network(x)
        # [b] => [b, 10]
        y_onehot = tf.one_hot(y, depth=10)
        # [b]
        loss = tf.reduce_mean(tf.losses.categorical_crossentropy(y_onehot, out, from_logits=True))

        loss_meter.update_state(loss)

    grads = tape.gradient(loss, network.trainable_variables)
    optimizer.apply_gradients(zip(grads, network.trainable_variables))

    if step % 100 == 0:
        print(step, 'loss:', loss_meter.result().numpy())
        loss_meter.reset_states()

    # evaluate
    if step % 500 == 0:
        total, total_correct = 0., 0
        acc_meter.reset_states()

        for step, (x, y) in enumerate(ds_val):
            # [b, 28, 28] => [b, 784]
            x = tf.reshape(x, (-1, 28 * 28))
            # [b, 784] => [b, 10]
            out = network(x)

            # [b, 10] => [b]
            pred = tf.argmax(out, axis=1)
            pred = tf.cast(pred, dtype=tf.int32)
            # bool type
            correct = tf.equal(pred, y)
            # bool tensor => int tensor => numpy
            total_correct += tf.reduce_sum(tf.cast(correct, dtype=tf.int32)).numpy()
            total += x.shape[0]

            acc_meter.update_state(y, pred)

        print(step, 'Evaluate Acc:', total_correct / total, acc_meter.result().numpy())

datasets: (60000, 28, 28) (60000,) 0 255
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                multiple                  200960    
_________________________________________________________________
dense_1 (Dense)              multiple                  32896     
_________________________________________________________________
dense_2 (Dense)              multiple                  8256      
_________________________________________________________________
dense_3 (Dense)              multiple                  2080      
_________________________________________________________________
dense_4 (Dense)              multiple                  330       
=================================================================
Total params: 244,522
Trainable params: 244,522
Non-trainable params: 0
_________________________________________________________________
0 loss: 2.3095727
78 Evaluate Acc: 0.1032 0.1032
100 loss: 0.49836162
200 loss: 0.24281283
300 loss: 0.20814449
400 loss: 0.19040857
500 loss: 0.1471103
78 Evaluate Acc: 0.956 0.956
600 loss: 0.15806517
700 loss: 0.13501912
800 loss: 0.13778095
900 loss: 0.13771541
1000 loss: 0.11204889
78 Evaluate Acc: 0.9666 0.9666
1100 loss: 0.10818114
1200 loss: 0.10698662
1300 loss: 0.10993517
1400 loss: 0.10309881
1500 loss: 0.092004016
78 Evaluate Acc: 0.9658 0.9658
1600 loss: 0.09988546
1700 loss: 0.09517718
1800 loss: 0.102653
1900 loss: 0.10128655
2000 loss: 0.084593534
78 Evaluate Acc: 0.9696 0.9696
2100 loss: 0.089395694
2200 loss: 0.084114745
2300 loss: 0.08294669
2400 loss: 0.0765419
2500 loss: 0.07786285
78 Evaluate Acc: 0.9716 0.9716
2600 loss: 0.08739958
2700 loss: 0.08950595
2800 loss: 0.08106578
2900 loss: 0.06466477
3000 loss: 0.077431396
78 Evaluate Acc: 0.9707 0.9707
3100 loss: 0.08382876
3200 loss: 0.076059125
3300 loss: 0.07230227
3400 loss: 0.05853687
3500 loss: 0.07312769
78 Evaluate Acc: 0.9703 0.9703
3600 loss: 0.07384481
3700 loss: 0.08926408
3800 loss: 0.066682965
3900 loss: 0.05534654
4000 loss: 0.073996484
78 Evaluate Acc: 0.9741 0.9741
4100 loss: 0.066883035
4200 loss: 0.070191
4300 loss: 0.08581101
4400 loss: 0.07324687
4500 loss: 0.056211904
78 Evaluate Acc: 0.9751 0.9751
4600 loss: 0.05384313

二、compile&fit&Evaluate&Predict

1.compile—编译模型

指定训练时loss（损失）、optimizer（优化器）和metrics（评价指标）的选择

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets,optimizers,losses,metrics,layers,Sequential

# 数据预处理
def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)

    return x, y

batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())
x = x.reshape((-1,28*28))
x_val = x_val.reshape((-1,28*28))
# 数据集加载并预处理
db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz).repeat(10)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz, drop_remainder=True)

# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(1)])
network.build(input_shape=(None, 28 * 28))

# 编译模型
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss = losses.CategoricalCrossentropy(from_logits=True),
                metrics = ['accuracy'])
2.fit—训练模型
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets,optimizers,losses,metrics,layers,Sequential

# 数据预处理
def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)

    return x, y

batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())
x = x.reshape((-1,28*28))
x_val = x_val.reshape((-1,28*28))
# 数据集加载并预处理
db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz).repeat(10)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz, drop_remainder=True)

# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(1)])
network.build(input_shape=(None, 28 * 28))

# 编译模型
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss = losses.CategoricalCrossentropy(from_logits=True),
                metrics = ['accuracy'])

# 训练模型
network.fit(db,epochs=100)
3.Evaluate—评估模型
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets,optimizers,losses,metrics,layers,Sequential
print(tf.__version__)
# 数据预处理
def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)

    return x, y

batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())
x = x.reshape((-1,28*28))
x_val = x_val.reshape((-1,28*28))
# 数据集加载并预处理
db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz).repeat(10)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz, drop_remainder=True)

# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(1)])
network.build(input_shape=(None, 28 * 28))

# 编译模型
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss = losses.CategoricalCrossentropy(from_logits=True),
                metrics = ['accuracy'])

# 训练模型
network.fit(db,epochs=10,validation_data = ds_val)
4.predict—预测
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets,optimizers,losses,metrics,layers,Sequential
from sklearn.metrics import accuracy_score
import numpy as np
print(tf.__version__)
# 数据预处理
def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)

    return x, y

batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())

x = x.reshape((-1,28*28))
y = tf.one_hot(y,depth=10)
x_val = x_val.reshape((-1,28*28))
y_val = tf.one_hot(y_val,depth=10)

# 数据集加载并预处理
db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz).repeat(10)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz, drop_remainder=True)

# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(10)])
network.build(input_shape=(None, 28 * 28))

# 编译模型
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss = losses.CategoricalCrossentropy(from_logits=True),
                metrics = ['accuracy'])

# 训练模型
network.fit(db,epochs=5,validation_data = ds_val,validation_freq=2)
network.summary()

# val
network.evaluate(ds_val)

# predict
pred = network.predict(x_val)

y_true = tf.argmax(y_val,axis=1)
y_pred = tf.argmax(pred,axis=1)
correct = tf.equal(y_true,y_pred)
total_correct = tf.reduce_sum(tf.cast(correct,dtype=np.int32)).numpy()
print(total_correct/x_val.shape[0])

Epoch 1/5
4690/4690 [==============================] - 16s 3ms/step - loss: 0.1098 - accuracy: 0.9695
Epoch 2/5
4690/4690 [==============================] - 18s 4ms/step - loss: 0.0531 - accuracy: 0.9873 - val_loss: 0.1227 - val_accuracy: 0.9776
Epoch 3/5
4690/4690 [==============================] - 19s 4ms/step - loss: 0.0448 - accuracy: 0.9902
Epoch 4/5
4690/4690 [==============================] - 18s 4ms/step - loss: 0.0376 - accuracy: 0.9923 - val_loss: 0.1778 - val_accuracy: 0.9763
Epoch 5/5
4690/4690 [==============================] - 19s 4ms/step - loss: 0.0368 - accuracy: 0.9921
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 256)               200960    
_________________________________________________________________
dense_1 (Dense)              (None, 128)               32896     
_________________________________________________________________
dense_2 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_3 (Dense)              (None, 32)                2080      
_________________________________________________________________
dense_4 (Dense)              (None, 10)                330       
=================================================================
Total params: 244,522
Trainable params: 244,522
Non-trainable params: 0
_________________________________________________________________
78/78 [==============================] - 0s 3ms/step - loss: 0.1899 - accuracy: 0.9758
tf.Tensor([7 2 1 ... 4 5 6], shape=(10000,), dtype=int64)
tf.Tensor([7 2 1 ... 4 5 6], shape=(10000,), dtype=int64)
0.9737

三、自定义层或网络

1.keras.Sequential
# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(10)])
# 建立网络参数
network.build(input_shape=(None, 28 * 28))
2.keras.Model / keras.layers.Layer

继承keras.layers.Layer来实现自定义层，自己的逻辑在call()方法中

  • __init__

  • call()

继承keras.Model来实现自定义网络，中间有一个小类继承keras.layers.Layer

  • __init__

  • call()

  • Model:compile / fit / evaluate
3.自定义层
# 自定义Dense层
class MyDense(layers.Layer):
	# 初始化方法
	def __init__(self,inp_dim,outp_dim):
		# 调用母类的初始化
		super(MyDense,self).__init__()
		# self.add_variable作用是在创建这两个Variable时，同时告诉类这两个variable是需要创建的
		# 当两个容器拼接时，会把这两个variable交给上面的容器来管理，统一管理，不需要人为管理参数
		# 这个函数在母类中实现，所以可以直接调用
		self.kernel = self.add_variable('w',[inp_dim,outp_dim])
		self.bias = self.add_variable('b',[outp_dim])
	
	def call(self,inputs,training = None):
		out = inputs @	self.kernel + self.bias
		return out
4.自定义网络
# 自定义Dense层
class MyDense(layers.Layer):
	# 初始化方法
	def __init__(self,inp_dim,outp_dim):
		# 调用母类的初始化
		super(MyDense,self).__init__()
		# self.add_variable作用是在创建这两个Variable时，同时告诉类这两个variable是需要创建的
		# 当两个容器拼接时，会把这两个variable交给上面的容器来管理，统一管理，不需要人为管理参数
		# 这个函数在母类中实现，所以可以直接调用
		self.kernel = self.add_variable('w',[inp_dim,outp_dim])
		self.bias = self.add_variable('b',[outp_dim])
	
	def call(self,inputs,training = None):
		out = inputs @	self.kernel + self.bias
		return out

# 利用自定义层，创建自定义网络(5层)
class MyModel(keras.Model):
	def __init__(self):
		super(MyModel,self).__init__()
		self.fc1 = MyDense(28*28,256)
		self.fc2 = MyDense(256,128)
		self.fc3 = MyDense(128,64)
		self.fc4 = MyDense(64,32)
		self.fc5 = MyDense(32,10)

	# 定义前向传播
	def call(self,inputs,training = None):
		x = self.fc1(inputs)
		x = tf.nn.relu(x)
		
		x = self.fc2(x)
		x = tf.nn.relu(x)	

		x = self.fc3(x)
		x = tf.nn.relu(x)
		x = self.fc4(x)
		x = tf.nn.relu(x)
		x = self.fc5(x)
		return x
5.自定义网络实战—手写数字识别
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics
from tensorflow import keras


# 数据预处理
def preprocess(x, y):
    """
    x is a simple image, not a batch
    """
    x = tf.cast(x, dtype=tf.float32) / 255.
    x = tf.reshape(x, [28 * 28])
    y = tf.cast(y, dtype=tf.int32)
    y = tf.one_hot(y, depth=10)
    return x, y


batchsz = 128
# 数据集加载
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())

db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz)

sample = next(iter(db))
print(sample[0].shape, sample[1].shape)

# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(10)])
network.build(input_shape=(None, 28 * 28))
network.summary()

# 自定义构建多层网络
# 自定义层
class MyDense(layers.Layer):

    def __init__(self, inp_dim, outp_dim):
        super(MyDense, self).__init__()

        self.kernel = self.add_variable('w', [inp_dim, outp_dim])
        self.bias = self.add_variable('b', [outp_dim])

    def call(self, inputs, training=None):
        out = inputs @ self.kernel + self.bias

        return out

# 自定义网络
class MyModel(keras.Model):

    def __init__(self):
        super(MyModel, self).__init__()

        self.fc1 = MyDense(28 * 28, 256)
        self.fc2 = MyDense(256, 128)
        self.fc3 = MyDense(128, 64)
        self.fc4 = MyDense(64, 32)
        self.fc5 = MyDense(32, 10)

    def call(self, inputs, training=None):
        x = self.fc1(inputs)
        x = tf.nn.relu(x)
        x = self.fc2(x)
        x = tf.nn.relu(x)
        x = self.fc3(x)
        x = tf.nn.relu(x)
        x = self.fc4(x)
        x = tf.nn.relu(x)
        x = self.fc5(x)

        return x


network = MyModel()

network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )

network.fit(db, epochs=5, validation_data=ds_val,
            validation_freq=2)

network.evaluate(ds_val)

sample = next(iter(ds_val))
x = sample[0]
y = sample[1]  # one-hot
pred = network.predict(x)  # [b, 10]
# convert back to number 
y = tf.argmax(y, axis=1)
pred = tf.argmax(pred, axis=1)

print(pred)
print(y)
6.自定义网络实战—CIFAR10

在这里插入图片描述

import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics

# 数据预处理
def preprocess(x,y):
    # [-1,1]
    x = 2 * tf.cast(x,dtype=tf.float32) / 255. - 1
    y = tf.cast(y,dtype=tf.int32)
    return x,y

batchsz = 128
# 数据集的加载
# x[b,32,32,3]  y[b,1]
(x,y),(x_val,y_val) = datasets.cifar10.load_data()

# 消去[b,1]的1这个维度
y = tf.squeeze(y)
y_val = tf.squeeze(y_val)

y = tf.one_hot(y,depth=10)
y_val = tf.one_hot(y_val,depth=10)
print('datasets:',x.shape,y.shape,x.min(),x.max())
# datasets: (50000, 32, 32, 3) (50000, 10) 0 255

# 构建两个数据集
train_db = tf.data.Dataset.from_tensor_slices((x,y))
train_db = train_db.map(preprocess).shuffle(10000).batch(batchsz)
test_db = tf.data.Dataset.from_tensor_slices((x_val,y_val))
test_db = test_db.map(preprocess).batch(batchsz)

sample = next(iter(train_db))
print('batch:',sample[0].shape,sample[1].shape)

# 创建自己的层
# replace standard layers.Dense
class MyDense(layers.Layer):
    def __init__(self,inp_dim,outp_dim):
        super(MyDense,self).__init__()

        self.kernel = self.add_variable('w',[inp_dim,outp_dim])
        # self.bias = self.add_variable('b',[outp_dim])

    # 构建前向传播
    def call(self,input,training = None):
        x = input @ self.kernel
        return x

# 构建自定义网络(5层)
class MyNetwork(keras.Model):
    def __init__(self):
        super(MyNetwork,self).__init__()

        # 优化-使参数变大-但容易造成过拟合
        self.fc1 = MyDense(32*32*3,256)
        self.fc2 = MyDense(256,128)
        self.fc3 = MyDense(128,64)
        self.fc4 = MyDense(64,32)
        self.fc5 = MyDense(32,10)

    def call(self,inputs,training=None):
        """

        :param inputs: [b,32,32,3]
        :param training:
        :return:
        """
        # 打平操作
        x = tf.reshape(inputs,[-1,32*32*3])
        x = self.fc1(x)
        x = tf.nn.relu(x)
        x = self.fc2(x)
        x = tf.nn.relu(x)
        x = self.fc3(x)
        x = tf.nn.relu(x)
        x = self.fc4(x)
        x = tf.nn.relu(x)
        # x[b,32]->[b,10]
        x = self.fc5(x)
        return x

network = MyNetwork()
network.compile(optimizer = optimizers.Adam(lr = 1e-3),
                loss = tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics = ['accuracy'])

network.fit(train_db,epochs=15,validation_data = test_db,validation_freq=1)

# 保存模型权值
network.evaluate(test_db)
network.save_weights('ckpt/weights.ckpt')
del network
print('saved to ckpt/weights.ckpt')

network = MyNetwork()
network.compile(optimizer = optimizers.Adam(lr = 1e-3),
                loss = tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics = ['accuracy'])

# 加载模型权值
network.load_weights('ckpt/weights.ckpt')
print('load weights from file')
network.evaluate(test_db)

Epoch 14/15
  1/391 [..............................] - ETA: 2:59 - loss: 0.6248 - accuracy: 0.8047
  8/391 [..............................] - ETA: 24s - loss: 0.6025 - accuracy: 0.7744 
 14/391 [>.............................] - ETA: 15s - loss: 0.5613 - accuracy: 0.7952
 20/391 [>.............................] - ETA: 11s - loss: 0.5669 - accuracy: 0.7969
 26/391 [>.............................] - ETA: 9s - loss: 0.5580 - accuracy: 0.8029 
 32/391 [=>............................] - ETA: 8s - loss: 0.5757 - accuracy: 0.7932
 38/391 [=>............................] - ETA: 7s - loss: 0.5719 - accuracy: 0.7926
 44/391 [==>...........................] - ETA: 6s - loss: 0.5721 - accuracy: 0.7933
 50/391 [==>...........................] - ETA: 5s - loss: 0.5669 - accuracy: 0.7962
 56/391 [===>..........................] - ETA: 5s - loss: 0.5710 - accuracy: 0.7939
 62/391 [===>..........................] - ETA: 5s - loss: 0.5740 - accuracy: 0.7941
 68/391 [====>.........................] - ETA: 4s - loss: 0.5731 - accuracy: 0.7945
 75/391 [====>.........................] - ETA: 4s - loss: 0.5753 - accuracy: 0.7922
 81/391 [=====>........................] - ETA: 4s - loss: 0.5745 - accuracy: 0.7936
 88/391 [=====>........................] - ETA: 4s - loss: 0.5727 - accuracy: 0.7936
 94/391 [======>.......................] - ETA: 3s - loss: 0.5742 - accuracy: 0.7927
101/391 [======>.......................] - ETA: 3s - loss: 0.5736 - accuracy: 0.7932
107/391 [=======>......................] - ETA: 3s - loss: 0.5724 - accuracy: 0.7934
114/391 [=======>......................] - ETA: 3s - loss: 0.5749 - accuracy: 0.7926
120/391 [========>.....................] - ETA: 3s - loss: 0.5757 - accuracy: 0.7934
126/391 [========>.....................] - ETA: 3s - loss: 0.5722 - accuracy: 0.7951
133/391 [=========>....................] - ETA: 3s - loss: 0.5721 - accuracy: 0.7955
139/391 [=========>....................] - ETA: 2s - loss: 0.5717 - accuracy: 0.7955
146/391 [==========>...................] - ETA: 2s - loss: 0.5715 - accuracy: 0.7954
152/391 [==========>...................] - ETA: 2s - loss: 0.5694 - accuracy: 0.7959
159/391 [===========>..................] - ETA: 2s - loss: 0.5688 - accuracy: 0.7957
166/391 [===========>..................] - ETA: 2s - loss: 0.5699 - accuracy: 0.7948
173/391 [============>.................] - ETA: 2s - loss: 0.5699 - accuracy: 0.7953
180/391 [============>.................] - ETA: 2s - loss: 0.5691 - accuracy: 0.7954
187/391 [=============>................] - ETA: 2s - loss: 0.5686 - accuracy: 0.7957
193/391 [=============>................] - ETA: 2s - loss: 0.5687 - accuracy: 0.7956
200/391 [==============>...............] - ETA: 2s - loss: 0.5694 - accuracy: 0.7952
207/391 [==============>...............] - ETA: 1s - loss: 0.5688 - accuracy: 0.7954
214/391 [===============>..............] - ETA: 1s - loss: 0.5673 - accuracy: 0.7951
221/391 [===============>..............] - ETA: 1s - loss: 0.5672 - accuracy: 0.7953
228/391 [================>.............] - ETA: 1s - loss: 0.5661 - accuracy: 0.7958
234/391 [================>.............] - ETA: 1s - loss: 0.5651 - accuracy: 0.7959
240/391 [=================>............] - ETA: 1s - loss: 0.5638 - accuracy: 0.7964
247/391 [=================>............] - ETA: 1s - loss: 0.5638 - accuracy: 0.7962
254/391 [==================>...........] - ETA: 1s - loss: 0.5627 - accuracy: 0.7971
261/391 [===================>..........] - ETA: 1s - loss: 0.5635 - accuracy: 0.7968
268/391 [===================>..........] - ETA: 1s - loss: 0.5642 - accuracy: 0.7966
275/391 [====================>.........] - ETA: 1s - loss: 0.5638 - accuracy: 0.7969
282/391 [====================>.........] - ETA: 1s - loss: 0.5633 - accuracy: 0.7972
289/391 [=====================>........] - ETA: 1s - loss: 0.5626 - accuracy: 0.7973
296/391 [=====================>........] - ETA: 0s - loss: 0.5625 - accuracy: 0.7973
302/391 [======================>.......] - ETA: 0s - loss: 0.5629 - accuracy: 0.7968
309/391 [======================>.......] - ETA: 0s - loss: 0.5641 - accuracy: 0.7967
318/391 [=======================>......] - ETA: 0s - loss: 0.5652 - accuracy: 0.7964
332/391 [========================>.....] - ETA: 0s - loss: 0.5661 - accuracy: 0.7960
347/391 [=========================>....] - ETA: 0s - loss: 0.5674 - accuracy: 0.7959
362/391 [==========================>...] - ETA: 0s - loss: 0.5676 - accuracy: 0.7957
376/391 [===========================>..] - ETA: 0s - loss: 0.5684 - accuracy: 0.7957
389/391 [============================>.] - ETA: 0s - loss: 0.5698 - accuracy: 0.7956
391/391 [==============================] - 4s 10ms/step - loss: 0.5697 - accuracy: 0.7956 - val_loss: 1.9200 - val_accuracy: 0.5195
Epoch 15/15
  1/391 [..............................] - ETA: 2:55 - loss: 0.6455 - accuracy: 0.7812
  8/391 [..............................] - ETA: 24s - loss: 0.5190 - accuracy: 0.8135 
 15/391 [>.............................] - ETA: 14s - loss: 0.5051 - accuracy: 0.8161
 22/391 [>.............................] - ETA: 10s - loss: 0.4930 - accuracy: 0.8224
 29/391 [=>............................] - ETA: 8s - loss: 0.4935 - accuracy: 0.8217 
 36/391 [=>............................] - ETA: 7s - loss: 0.4941 - accuracy: 0.8238
 43/391 [==>...........................] - ETA: 6s - loss: 0.4999 - accuracy: 0.8212
 50/391 [==>...........................] - ETA: 5s - loss: 0.5044 - accuracy: 0.8181
 57/391 [===>..........................] - ETA: 5s - loss: 0.5097 - accuracy: 0.8177
 64/391 [===>..........................] - ETA: 4s - loss: 0.5112 - accuracy: 0.8174
 71/391 [====>.........................] - ETA: 4s - loss: 0.5097 - accuracy: 0.8168
 78/391 [====>.........................] - ETA: 4s - loss: 0.5115 - accuracy: 0.8172
 85/391 [=====>........................] - ETA: 4s - loss: 0.5161 - accuracy: 0.8148
 92/391 [======>.......................] - ETA: 3s - loss: 0.5176 - accuracy: 0.8145
 99/391 [======>.......................] - ETA: 3s - loss: 0.5187 - accuracy: 0.8149
106/391 [=======>......................] - ETA: 3s - loss: 0.5168 - accuracy: 0.8155
113/391 [=======>......................] - ETA: 3s - loss: 0.5177 - accuracy: 0.8148
119/391 [========>.....................] - ETA: 3s - loss: 0.5190 - accuracy: 0.8147
125/391 [========>.....................] - ETA: 3s - loss: 0.5164 - accuracy: 0.8159
132/391 [=========>....................] - ETA: 2s - loss: 0.5162 - accuracy: 0.8159
139/391 [=========>....................] - ETA: 2s - loss: 0.5149 - accuracy: 0.8156
146/391 [==========>...................] - ETA: 2s - loss: 0.5149 - accuracy: 0.8157
153/391 [==========>...................] - ETA: 2s - loss: 0.5139 - accuracy: 0.8161
159/391 [===========>..................] - ETA: 2s - loss: 0.5161 - accuracy: 0.8150
165/391 [===========>..................] - ETA: 2s - loss: 0.5156 - accuracy: 0.8154
171/391 [============>.................] - ETA: 2s - loss: 0.5135 - accuracy: 0.8162
177/391 [============>.................] - ETA: 2s - loss: 0.5148 - accuracy: 0.8158
183/391 [=============>................] - ETA: 2s - loss: 0.5155 - accuracy: 0.8155
189/391 [=============>................] - ETA: 2s - loss: 0.5171 - accuracy: 0.8147
195/391 [=============>................] - ETA: 2s - loss: 0.5189 - accuracy: 0.8140
201/391 [==============>...............] - ETA: 1s - loss: 0.5175 - accuracy: 0.8144
208/391 [==============>...............] - ETA: 1s - loss: 0.5165 - accuracy: 0.8144
214/391 [===============>..............] - ETA: 1s - loss: 0.5185 - accuracy: 0.8137
221/391 [===============>..............] - ETA: 1s - loss: 0.5182 - accuracy: 0.8140
228/391 [================>.............] - ETA: 1s - loss: 0.5175 - accuracy: 0.8143
234/391 [================>.............] - ETA: 1s - loss: 0.5170 - accuracy: 0.8144
240/391 [=================>............] - ETA: 1s - loss: 0.5161 - accuracy: 0.8150
246/391 [=================>............] - ETA: 1s - loss: 0.5168 - accuracy: 0.8142
253/391 [==================>...........] - ETA: 1s - loss: 0.5165 - accuracy: 0.8140
259/391 [==================>...........] - ETA: 1s - loss: 0.5169 - accuracy: 0.8136
265/391 [===================>..........] - ETA: 1s - loss: 0.5164 - accuracy: 0.8138
271/391 [===================>..........] - ETA: 1s - loss: 0.5161 - accuracy: 0.8139
278/391 [====================>.........] - ETA: 1s - loss: 0.5155 - accuracy: 0.8142
284/391 [====================>.........] - ETA: 1s - loss: 0.5156 - accuracy: 0.8140
291/391 [=====================>........] - ETA: 0s - loss: 0.5143 - accuracy: 0.8148
298/391 [=====================>........] - ETA: 0s - loss: 0.5146 - accuracy: 0.8146
305/391 [======================>.......] - ETA: 0s - loss: 0.5148 - accuracy: 0.8146
312/391 [======================>.......] - ETA: 0s - loss: 0.5151 - accuracy: 0.8144
325/391 [=======================>......] - ETA: 0s - loss: 0.5148 - accuracy: 0.8142
339/391 [=========================>....] - ETA: 0s - loss: 0.5161 - accuracy: 0.8137
354/391 [==========================>...] - ETA: 0s - loss: 0.5169 - accuracy: 0.8136
369/391 [===========================>..] - ETA: 0s - loss: 0.5193 - accuracy: 0.8127
383/391 [============================>.] - ETA: 0s - loss: 0.5197 - accuracy: 0.8126
391/391 [==============================] - 4s 10ms/step - loss: 0.5200 - accuracy: 0.8126 - val_loss: 2.0124 - val_accuracy: 0.5189
 1/79 [..............................] - ETA: 0s - loss: 1.6155 - accuracy: 0.5625
10/79 [==>...........................] - ETA: 0s - loss: 1.8749 - accuracy: 0.5273
19/79 [======>.......................] - ETA: 0s - loss: 1.9776 - accuracy: 0.5169
27/79 [=========>....................] - ETA: 0s - loss: 1.9817 - accuracy: 0.5194
36/79 [============>.................] - ETA: 0s - loss: 1.9576 - accuracy: 0.5252
45/79 [================>.............] - ETA: 0s - loss: 1.9520 - accuracy: 0.5274
54/79 [===================>..........] - ETA: 0s - loss: 1.9581 - accuracy: 0.5268
63/79 [======================>.......] - ETA: 0s - loss: 1.9572 - accuracy: 0.5255
72/79 [==========================>...] - ETA: 0s - loss: 1.9786 - accuracy: 0.5215
79/79 [==============================] - 0s 6ms/step - loss: 2.0124 - accuracy: 0.5189
saved to ckpt/weights.ckpt
load weights from file
 1/79 [..............................] - ETA: 5s - loss: 1.6155 - accuracy: 0.5625
10/79 [==>...........................] - ETA: 0s - loss: 1.8749 - accuracy: 0.5273
19/79 [======>.......................] - ETA: 0s - loss: 1.9776 - accuracy: 0.5169
28/79 [=========>....................] - ETA: 0s - loss: 1.9824 - accuracy: 0.5176
37/79 [=============>................] - ETA: 0s - loss: 1.9426 - accuracy: 0.5283
46/79 [================>.............] - ETA: 0s - loss: 1.9576 - accuracy: 0.5275
55/79 [===================>..........] - ETA: 0s - loss: 1.9660 - accuracy: 0.5259
64/79 [=======================>......] - ETA: 0s - loss: 1.9604 - accuracy: 0.5242
73/79 [==========================>...] - ETA: 0s - loss: 1.9814 - accuracy: 0.5205
79/79 [==============================] - 1s 7ms/step - loss: 2.0124 - accuracy: 0.5189

四、模型的加载与保存

模型的加载与保存一共有三个模式，分别为：

  • save / load weights
    最干净、最轻量级的模型，只保存网络参数，适用于有源代码的情况下
  • save / load entire model
    最简单粗暴，将模型的所有状态都保存下来，可以用来恢复
  • save_model
    模型的一种保存模式，与pyt中的ONNX模式相对应，适用于工厂环境部署
    python代码可以用C++来解析
1.save / load weights
import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics


def preprocess(x, y):
    """
    x is a simple image, not a batch
    """
    x = tf.cast(x, dtype=tf.float32) / 255.
    x = tf.reshape(x, [28 * 28])
    y = tf.cast(y, dtype=tf.int32)
    y = tf.one_hot(y, depth=10)
    return x, y


batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())

db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz)

sample = next(iter(db))
print(sample[0].shape, sample[1].shape)

network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(10)])
network.build(input_shape=(None, 28 * 28))
network.summary()

network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )

network.fit(db, epochs=3, validation_data=ds_val, validation_freq=2)

network.evaluate(ds_val)

# 保存模型的参数
network.save_weights('weights.ckpt')
print('saved weights.')
del network

# 构建多层网络
network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(10)])
network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )
# 加载模型的参数
network.load_weights('weights.ckpt')
print('loaded weights!')
network.evaluate(ds_val)

Epoch 2/3
  1/469 [..............................] - ETA: 13:50 - loss: 0.1488 - accuracy: 0.9688
 19/469 [>.............................] - ETA: 43s - loss: 0.1262 - accuracy: 0.9634  
 39/469 [=>............................] - ETA: 20s - loss: 0.1342 - accuracy: 0.9619
 57/469 [==>...........................] - ETA: 13s - loss: 0.1324 - accuracy: 0.9635
 76/469 [===>..........................] - ETA: 10s - loss: 0.1379 - accuracy: 0.9623
 93/469 [====>.........................] - ETA: 8s - loss: 0.1348 - accuracy: 0.9632 
110/469 [======>.......................] - ETA: 6s - loss: 0.1370 - accuracy: 0.9618
130/469 [=======>......................] - ETA: 5s - loss: 0.1375 - accuracy: 0.9615
150/469 [========>.....................] - ETA: 4s - loss: 0.1384 - accuracy: 0.9618
169/469 [=========>....................] - ETA: 3s - loss: 0.1384 - accuracy: 0.9614
187/469 [==========>...................] - ETA: 3s - loss: 0.1369 - accuracy: 0.9619
207/469 [============>.................] - ETA: 2s - loss: 0.1385 - accuracy: 0.9612
229/469 [=============>................] - ETA: 2s - loss: 0.1387 - accuracy: 0.9612
251/469 [===============>..............] - ETA: 2s - loss: 0.1393 - accuracy: 0.9610
274/469 [================>.............] - ETA: 1s - loss: 0.1388 - accuracy: 0.9615
297/469 [=================>............] - ETA: 1s - loss: 0.1378 - accuracy: 0.9616
319/469 [===================>..........] - ETA: 1s - loss: 0.1373 - accuracy: 0.9618
342/469 [====================>.........] - ETA: 0s - loss: 0.1366 - accuracy: 0.9621
363/469 [======================>.......] - ETA: 0s - loss: 0.1356 - accuracy: 0.9624
385/469 [=======================>......] - ETA: 0s - loss: 0.1362 - accuracy: 0.9623
407/469 [=========================>....] - ETA: 0s - loss: 0.1358 - accuracy: 0.9624
429/469 [==========================>...] - ETA: 0s - loss: 0.1350 - accuracy: 0.9627
450/469 [===========================>..] - ETA: 0s - loss: 0.1342 - accuracy: 0.9629
466/469 [============================>.] - ETA: 0s - loss: 0.1343 - accuracy: 0.9629
469/469 [==============================] - 3s 7ms/step - loss: 0.1344 - accuracy: 0.9629 - val_loss: 0.1209 - val_accuracy: 0.9648
Epoch 3/3
  1/469 [..............................] - ETA: 14:16 - loss: 0.1254 - accuracy: 0.9609
 20/469 [>.............................] - ETA: 42s - loss: 0.1014 - accuracy: 0.9695  
 39/469 [=>............................] - ETA: 21s - loss: 0.1063 - accuracy: 0.9700
 60/469 [==>...........................] - ETA: 13s - loss: 0.1006 - accuracy: 0.9703
 82/469 [====>.........................] - ETA: 9s - loss: 0.1041 - accuracy: 0.9690 
105/469 [=====>........................] - ETA: 7s - loss: 0.1089 - accuracy: 0.9676
128/469 [=======>......................] - ETA: 5s - loss: 0.1072 - accuracy: 0.9684
151/469 [========>.....................] - ETA: 4s - loss: 0.1056 - accuracy: 0.9692
171/469 [=========>....................] - ETA: 3s - loss: 0.1089 - accuracy: 0.9688
189/469 [===========>..................] - ETA: 3s - loss: 0.1094 - accuracy: 0.9688
208/469 [============>.................] - ETA: 2s - loss: 0.1122 - accuracy: 0.9681
228/469 [=============>................] - ETA: 2s - loss: 0.1099 - accuracy: 0.9687
250/469 [==============>...............] - ETA: 2s - loss: 0.1093 - accuracy: 0.9691
270/469 [================>.............] - ETA: 1s - loss: 0.1088 - accuracy: 0.9692
291/469 [=================>............] - ETA: 1s - loss: 0.1081 - accuracy: 0.9696
312/469 [==================>...........] - ETA: 1s - loss: 0.1079 - accuracy: 0.9700
334/469 [====================>.........] - ETA: 1s - loss: 0.1082 - accuracy: 0.9700
356/469 [=====================>........] - ETA: 0s - loss: 0.1086 - accuracy: 0.9699
378/469 [=======================>......] - ETA: 0s - loss: 0.1083 - accuracy: 0.9699
401/469 [========================>.....] - ETA: 0s - loss: 0.1071 - accuracy: 0.9700
422/469 [=========================>....] - ETA: 0s - loss: 0.1081 - accuracy: 0.9698
441/469 [===========================>..] - ETA: 0s - loss: 0.1089 - accuracy: 0.9697
459/469 [============================>.] - ETA: 0s - loss: 0.1083 - accuracy: 0.9700
469/469 [==============================] - 3s 6ms/step - loss: 0.1082 - accuracy: 0.9701
 1/79 [..............................] - ETA: 0s - loss: 0.0620 - accuracy: 0.9844
11/79 [===>..........................] - ETA: 0s - loss: 0.1625 - accuracy: 0.9616
21/79 [======>.......................] - ETA: 0s - loss: 0.1902 - accuracy: 0.9576
32/79 [===========>..................] - ETA: 0s - loss: 0.1910 - accuracy: 0.9570
41/79 [==============>...............] - ETA: 0s - loss: 0.1845 - accuracy: 0.9573
50/79 [=================>............] - ETA: 0s - loss: 0.1695 - accuracy: 0.9605
60/79 [=====================>........] - ETA: 0s - loss: 0.1499 - accuracy: 0.9645
70/79 [=========================>....] - ETA: 0s - loss: 0.1389 - accuracy: 0.9667
79/79 [==============================] - 0s 5ms/step - loss: 0.1372 - accuracy: 0.9664
saved weights.
loaded weights!
 1/79 [..............................] - ETA: 6s - loss: 0.0620 - accuracy: 0.9844
11/79 [===>..........................] - ETA: 0s - loss: 0.1625 - accuracy: 0.9616
21/79 [======>.......................] - ETA: 0s - loss: 0.1902 - accuracy: 0.9576
30/79 [==========>...................] - ETA: 0s - loss: 0.1884 - accuracy: 0.9581
39/79 [=============>................] - ETA: 0s - loss: 0.1914 - accuracy: 0.9559
49/79 [=================>............] - ETA: 0s - loss: 0.1724 - accuracy: 0.9600
59/79 [=====================>........] - ETA: 0s - loss: 0.1522 - accuracy: 0.9640
69/79 [=========================>....] - ETA: 0s - loss: 0.1404 - accuracy: 0.9665
78/79 [============================>.] - ETA: 0s - loss: 0.1388 - accuracy: 0.9663
79/79 [==============================] - 1s 6ms/step - loss: 0.1372 - accuracy: 0.9664
2.save / load entire model
import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics

# 数据预处理
def preprocess(x, y):
    """
    x is a simple image, not a batch
    """
    x = tf.cast(x, dtype=tf.float32) / 255.
    x = tf.reshape(x, [28 * 28])
    y = tf.cast(y, dtype=tf.int32)
    y = tf.one_hot(y, depth=10)
    return x, y


batchsz = 128
# 数据集加载
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())

db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz)

sample = next(iter(db))
print(sample[0].shape, sample[1].shape)

network = Sequential([layers.Dense(256, activation='relu'),
                      layers.Dense(128, activation='relu'),
                      layers.Dense(64, activation='relu'),
                      layers.Dense(32, activation='relu'),
                      layers.Dense(10)])
network.build(input_shape=(None, 28 * 28))
network.summary()

network.compile(optimizer=optimizers.Adam(lr=0.01),
                loss=tf.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy']
                )

network.fit(db, epochs=3, validation_data=ds_val, validation_freq=2)

network.evaluate(ds_val)

# 保存整个模型
network.save('model.h5')
print('saved total model.')
del network

print('loaded model from file.')
# 加载整个模型
network = tf.keras.models.load_model('model.h5', compile=False)

x_val = tf.cast(x_val, dtype=tf.float32) / 255.
x_val = tf.reshape(x_val, [-1, 28 * 28])
y_val = tf.cast(y_val, dtype=tf.int32)
y_val = tf.one_hot(y_val, depth=10)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val)).batch(128)
network.evaluate(ds_val)

datasets: (60000, 28, 28) (60000,) 0 255
(128, 784) (128, 10)
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 256)               200960    
_________________________________________________________________
dense_1 (Dense)              (None, 128)               32896     
_________________________________________________________________
dense_2 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_3 (Dense)              (None, 32)                2080      
_________________________________________________________________
dense_4 (Dense)              (None, 10)                330       
=================================================================
Total params: 244,522
Trainable params: 244,522
Non-trainable params: 0
_________________________________________________________________
Epoch 1/3
469/469 [==============================] - 1s 2ms/step - loss: 0.2723 - accuracy: 0.9182
Epoch 2/3
469/469 [==============================] - 1s 3ms/step - loss: 0.1363 - accuracy: 0.9628 - val_loss: 0.1280 - val_accuracy: 0.9637
Epoch 3/3
469/469 [==============================] - 1s 2ms/step - loss: 0.1101 - accuracy: 0.9692
79/79 [==============================] - 0s 3ms/step - loss: 0.1372 - accuracy: 0.9673
saved total model.
loaded model from file.
79/79 [==============================] - 0s 1ms/step - loss: 0.1372 - accuracy: 0.9673
3.save_model
tf.saved_model.save(m,'/tmp/saved_model/')

imported = tf.saved_model.load(path)
f = imported.signatures['serving_default']
print(f(x = tf.ones([1,28,28,3])))

如果对您有帮助，麻烦点赞关注，这真的对我很重要！！！如果需要互关，请评论留言！
在这里插入图片描述

哎呦-_-不错
关注
专栏目录
深度学习2.0-22.Keras高层接口之自定义层或网络
柳杰的博客
10-10 523
文章目录1.keras.Sequential2.Layer/Model3.自定义层4.自定义网络5.自定义网络实战-手写数字识别 1.keras.Sequential 2.Layer/Model 3.自定义层 # 自定义Dense层 class MyDense(layers.Layer): # 初始化方法 def __init__(self,inp_dim,outp_dim): # 调用母类的初始化 super(MyDense,self).__init__() # self.add_
[实训题目EmoProfo]深度学习框架接口深入优化（一）
浅倉家的归档箱
05-08 3970
深度学习框架接口深入优化（一） 深度学习框架接口深入优化（一） 背景 项目整理 letterbox_image函数的深入剖析 图片预处理优化 思路简述 保证正确性基础的实现 实用实现与优化 小结 背景 继上次实现一个基础的优化版本完成后，我们提到了进一步优化的思路，而那之后又有了一些新的想法，在这里重新整理一下 上一个次更新进度中的内容本身是框架性质的东西，还不能用...
深度学习接口——参数设置及训练模型导入
toCVer的博客
12-11 2503
前言 在进行深度学习的研究和项目的训练测试过程中，需要经常对参数进行调整，有的需要通过命令行的方式进行修改，有的则可以作为固定参数放在程序中。 其中对于那种需要在运行程序时选择模式或者命名等只有在运行程序时才确定的参数一般采用命令行输入的方式确定；对于那些训练过程中一定需要的参数，我们可以通过将其写在程序中，来使其发挥作用。 这是我们设置参数部分，当然对于一些我们曾经训练过的模型，我们可以首先将模型保存，然后当再训练的时候，根据不同的指标，选取不同的模型，开始训练。今天我们就来讲解下深度学习过程中会用到的参
深度学习学习日记：Keras接口记录
水野与小太郎的博客
11-11 251
打算用Keras做迁移学习，书上和资料上说用keras做迁移学习比较方便（毕竟大部分迁移学习都是用的那些比较常用的数据集，keras比较适合和方便做） Keras有两种构建模型的方式：sequential（适用于简单模型）和 function API（适用于复杂的模型） sequential Model sequential 有如下常用方法： compile(optimizer, l...
深度学习常用数据集 API（包括 Fashion MNIST）
weixin_30421525的博客
11-13 957
基准数据集 深度学习中经常会使用一些基准数据集进行一些测试。其中 MNIST, Cifar 10, cifar100, Fashion-MNIST 数据集常常被人们拿来当作练手的数据集。为了方便，诸如 Keras、MXNet、Tensorflow 都封装了自己的基础数据集，如 MNIST、cifar 等。如果我们要在不同平台使用这些数据集，还需要了解那些框架是如何组织这些数据集的，需要花费一些不必...
深度学习2.0-20.Keras高层API-metrics
柳杰的博客
10-10 236
文章目录1.metrics1.metrics-新建一个评价指标2.update data- 添加数据3.result().numpy()-显示结果4.reset_states()-清零2.metrics实战 1.metrics 1.metrics-新建一个评价指标 acc_meter = metrucs.Accuracy() loss_meter = metrics.Mean() 2.update data- 添加数据 loss_meter.update_state(loss) acc_meter.u
【Tensorflow2.x学习笔记】- Keras高层接口
qq_42433311的博客
01-03 2416
目录 1.Keras简介 2.常见功能模块 2.1 常见网络层类 2.2 网路容器 3.模型装配、训练、测试 3.1 模型装配 3.2 模型训练 3.3模型测试 4.模型的保存与加载 4.1 张量方式 4.2网络方式 4.3 SaveModel方式 5.自定义网络 5.1自定义网络层 5.2自定义网络 6.测量工具 7.可视化 7.1模型端 7.2浏览器端 1.Keras简介 Keras是一个主要由 Python语言开发的开源神经网络计...
TensorFlow1.x最佳实践：Dataset API+Keras Model+TF Train
等一杯咖啡的博客
06-04 1953
前言 关于深度学习框架，主流的几个TensorFlow、PyTorch以及Keras都有所使用，由于在深度学习领域自己也只算个入门级选手，所以本文只从一个AI新手的角度去尝试分享一些使用框架编程的实践经验。至于标题最佳实践，那也纯粹有些哗众取宠之意，文章对于能够玩转各种框架API的大佬们，也许会贻笑大方。除此以外，本文相当于一个搬运工，并不讲解具体的使用细节，当然会推荐一些已经介绍的很好的文章，看...
Keras+TF实现深度学习中的Attention机制教程
在深度学习领域，Keras是一个高层神经网络API，它能够以TensorFlow作为后端运行。TensorFlow是由谷歌开发的开源机器学习框架，适用于大规模深度学习应用的开发。这两个框架的结合使用为开发者提供了一个强大的开发...
Python_为TensorFlow提供高级API的深度学习库.zip
最新发布
05-24
描述中的“Python_为TensorFlow提供高级API的深度学习库”可能指的是TFLearn，这是一个轻量级的库，它是TensorFlow的一个高层接口，让用户能够快速地搭建和训练深度学习模型。TFLearn提供了许多预定义的层、优化器、...
add_variable
weixin_42187536的博客
03-05 332
def add_variable(self, name, shape, dtype=None, initializer=None, regularizer=None, trainable=True, constraint=None, partitioner=None): 目的：向层添加一个新变量，或获取一个现有变量。 参数：name：变量名称。 shape：可变形状 dtype：变量的类型，默.
tensorflow2的compile & fit函数
进击的Explorer
04-08 9351
1、compile 配置用于训练的模型。 用法： compile ( optimizer='rmsprop', loss=None, metrics=None, loss_weights=None, sample_weight_mode=None, weighted_metrics=None, target_tensors=None, ...
viewer图片查看器JQ版
weixin_48058348的博客
11-19 1295
viewer图片查看器JS版 html <!DOCTYPE html> <html lang="zh-Hans"> <head> <meta charset="utf-8"> <title>JS/jQuery图片查看器viewer.js</title> <link rel="stylesheet" href="./css/viewer.css"> <style>
PyTorch学习笔记（五）：模型定义、修改、保存
m0_67402731的博客
03-21 2826
一、PyTorch模型定义的方式 Module 类是 torch.nn 模块里提供的一个模型构造类 (nn.Module)，是所有神经网络模块的基类，我们可以继承它来定义我们想要的模型； PyTorch模型定义应包括两个主要部分：各个部分的初始化（_init_）；数据流向定义（forward） 基于nn.Module，可以通过Sequential，ModuleList和ModuleDict三种方式定义PyTorch模型。 1.Sequential 对应模块为nn.Sequential()。 当模型的
Variables: 创建、初始化、保存和加载
热门推荐
再难也要坚持
11-30 1万+
引言当你训练一个模型的时候，你使用变量去保存和更新参数。在Tensorflow中变量是内存缓冲区中保存的张量（tensor）。它们必须被显示的初始化，可以在训练完成之后保存到磁盘上。之后，你可以重新加载这些值用于测试和模型分析。 本篇文档引用了如下的Tensorflow类。以下的链接指向它们更加详细的API： tf.Variable 类。 The tf.train.Saver 类。 创建当你创建
timertask run函数未执行_图执行模式下的 TensorFlow 2
weixin_39997443的博客
12-10 109
文 /李锡涵，Google Developers Expert本文节选自《简单粗暴 TensorFlow 2.0》尽管 TensorFlow 2 建议以即时执行模式(Eager Execution)作为主要执行模式，然而，图执行模式(Graph Execution)作为 TensorFlow 2 之前的主要执行模式，依旧对于我们理解 TensorFlow 具有重要意义。尤其是当我们需...
tensorflow之自定义神经网络层
大雄没有叮当猫的博客
08-16 7503
import tensorflow as tf tfe = tf.contrib.eager tf.enable_eager_execution() 大多数情况下，在为机器学习模型编写代码时，您希望在比单个操作和单个变量操作更高的抽象级别上操作。 1.关于图层的一些有用操作 许多机器学习模型可以表达为相对简单的图层的组合和堆叠，TensorFlow提供了一组许多常用图层，以及您从头开始或...
Tensorflow2.0教程-自定义层
知行_那片天
07-07 1万+
Tensorflow2.0教程-自定义层 最全Tensorflow 2.0 入门教程持续更新：https://blog.csdn.net/qq_31456593/article/details/88606284 完整tensorflow2.0教程代码请看https://github.com/czy36mengfei/tensorflow2_tutorials_chinese (欢迎star) 本...
tensorflow之RNNCell及自定义Cell
AI蜗牛之家
01-17 6236
我们在仿真一些论文的时候经常会遇到一些模型，对RNN或者LSTM进行了少许的修改，或者自己定义了一种RNN的结构等情况，比如前面介绍的几篇memory networks、qausi RNN 的论文，往往都需要按照自己定义的方法来构造RNN网络。所以本篇博客就主要总结一下RNNcell的用法以及如何按照自己的需求自定义RNNCell。 tf中RNNCell的用法介绍 我们直接从源码的层面来看一看tf...
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值