fit函数 model_【Tensorflow2.0】7、全流程model.fit模型训练方法

第一种方法直接用keras的fit方法，以fashion mnist为例

本文包括一个完整的深度学习模型构建过程。有tf.data读数据集的方法；使用keras callbacks来实现checkpoint,学习率衰减，Earlystop等功能。最后结合scikit-learn，对模型的推理使用做了说明。

import tensorflow as tf

import datetime

import matplotlib

matplotlib.use("Agg")#这个设置可以使matplotlib保存.png图到磁盘

import matplotlib.pyplot as plt

from functools import partial

import numpy as np

import pandas as pd

import os

from sklearn.metrics import confusion_matrix

from sklearn.metrics import classification_report1

2

3

4

5

6

7

8

9

10

11

12

#jupyter 中开启该选项，否则不执行

%matplotlib inline1

2

配置超参数

batch_size=64

epochs=5

regularizer=1e-3

total_train_samples=60000

total_test_samples=10000

lr_decay_epochs=1

output_folder="./model_output"

#用来保存模型以及我们需要的所有东西

if not os.path.exists(output_folder):

os.makedirs(output_folder)

save_format="hdf5" #或saved_model

if save_format=="hdf5":

save_path_models=os.path.join(output_folder,"hdf5_models")

if not os.path.exists(save_path_models):

os.makedirs(save_path_models)

save_path=os.path.join(save_path_models,"ckpt_epoch{epoch:02d}_val_acc{val_accuracy:.2f}.hdf5")

elif save_format=="saved_model":

save_path_models=os.path.join(output_folder,"saved_models")

if not os.path.exists(save_path_models):

os.makedirs(save_path_models)

save_path=os.path.join(save_path_models,"ckpt_epoch{epoch:02d}_val_acc{val_accuracy:.2f}.ckpt")

#用来保存日志

log_dir= os.path.join(output_folder,'logs_{}'.format(datetime.datetime.now().strftime("%Y%m%d-%H%M%S")))

if not os.path.exists(log_dir):

os.makedirs(log_dir)1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

选择指定显卡及自动调用显存

physical_devices = tf.config.experimental.list_physical_devices('GPU')#列出所有可见显卡

print("All the available GPUs:\n",physical_devices)

if physical_devices:

gpu=physical_devices[0]#显示第一块显卡

tf.config.experimental.set_memory_growth(gpu, True)#根据需要自动增长显存

tf.config.experimental.set_visible_devices(gpu, 'GPU')#只选择第一块1

2

3

4

5

6

All the available GPUs:

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]1

2

准备数据

fashion_mnist=tf.keras.datasets.fashion_mnist

(train_x,train_y),(test_x,test_y)=fashion_mnist.load_data()

train_x,test_x = train_x[...,np.newaxis]/255.0,test_x[...,np.newaxis]/255.0

total_train_sample = train_x.shape[0]

total_test_sample=test_x.shape[0]

class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',

'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']1

2

3

4

5

6

7

8

使用tf.data来准备训练集和测试集

train_ds = tf.data.Dataset.from_tensor_slices((train_x,train_y))

test_ds = tf.data.Dataset.from_tensor_slices((test_x,test_y))

train_ds=train_ds.shuffle(buffer_size=batch_size*10).batch(batch_size).prefetch(buffer_size = tf.data.experimental.AUTOTUNE).repeat()

test_ds = test_ds.batch(batch_size).prefetch(buffer_size = tf.data.experimental.AUTOTUNE)#不加repeat，执行一次就行1

2

3

4

5

准备模型定义

l2 = tf.keras.regularizers.l2(regularizer)#定义模型正则化方法

ini = tf.keras.initializers.he_normal()#定义参数初始化方法

conv2d = partial(tf.keras.layers.Conv2D,activation='relu',padding='same',kernel_regularizer=l2,bias_regularizer=l2)

fc = partial(tf.keras.layers.Dense,activation='relu',kernel_regularizer=l2,bias_regularizer=l2)

maxpool=tf.keras.layers.MaxPooling2D

dropout=tf.keras.layers.Dropout1

2

3

4

5

6

7

开始定义模型,用functional方法

x_input = tf.keras.layers.Input(shape=(28,28,1),name='input_node')

x = conv2d(128,(5,5))(x_input)

x = maxpool((2,2))(x)

x = conv2d(256,(5,5))(x)

x = maxpool((2,2))(x)

x = tf.keras.layers.Flatten()(x)

x = fc(128)(x)

x_output=fc(10,activation=None,name='output_node')(x)

model = tf.keras.models.Model(inputs=x_input,outputs=x_output)1

2

3

4

5

6

7

8

9

10

打印模型结构

print("The model architure:\n")

print(model.summary())1

2

The model architure:

Model: "model"

____