tensorflow基于dataSet的多任务训练代码

from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.applications import resnet
from tensorflow.keras.utils import plot_model
from tensorflow.keras.layers import Input, GlobalAvgPool2D, Dense, Dropout, Lambda, Conv2D
from tensorflow.keras.models import Model
from tensorflow.keras.losses import SparseCategoricalCrossentropy
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.callbacks import ModelCheckpoint, TensorBoard, EarlyStopping
from sklearn.model_selection import train_test_split

def readPictureByPath(path):
    img = cv2.imdecode(np.fromfile(path, np.uint8), -1)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    # return np.expand_dims(img,0)
    return img


import cv2
# 正则匹配使用：
import re
import os
# 此库用于拷贝，删除，移动，复制以及解压缩
import shutil
import numpy as np
import h5py
import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator
# 用于将普通标签转为独热向量
from tensorflow.keras.utils import to_categorical
import matplotlib.pyplot as plt


# 如此读取图像，直接返回numpy.ndarray
# img=cv2.imdecode(np.fromfile("C:/Users/104005162/Desktop/企业微信截图_20220212102256.png",np.uint8),-1)
# print(img.shape)
##转换为bgr图片，注意此时是PNG图片，不能用矩阵直接转换！
##img=img[:,:,::-1]
##bgra
# img=cv2.cvtColor(img,cv2.COLOR_BGRA2RGB)
# 下面是python的无放回抽样：


def readPictureByPath(path):
    img = cv2.imdecode(np.fromfile(path, np.uint8), -1)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    # return np.expand_dims(img,0)
    return img


# 加载数据
xData = []
yLabel = []
with h5py.File("C:/Users/25360/Desktop/model/myTextData", 'r') as d:
    xData = np.array(d['x'])
    yLabel = np.array(d['label'])
#157,4
#print(yLabel.shape)
#双任务y应为(157,2,4)：
yLabelAll=tf.transpose(np.array([yLabel,yLabel]),[1,0,2])
#print(yLabelAll.shape)
#进行切片生成数据集：
dataText=tf.data.Dataset.from_tensor_slices((xData,yLabelAll))
#到多任务输出:
#其中yLabelAll为(157,2,4)：注意，打包时格式应为元组
def toMulitTaskOutPut(xData,yLabelAll):
    yLabelAll=tf.transpose(yLabelAll,[1,0,2])
    return xData,(yLabelAll[0],yLabelAll[1])
#以四个一样的任务进行演示:识别服装类型：
#输入进行额外命名

#batch后出现第一维：bacth前只有图像维
trainData=dataText.take(100).batch(8).map(toMulitTaskOutPut)
TextData=dataText.skip(100).take(20).batch(8).map(toMulitTaskOutPut)
validData=dataText.skip(100).skip(20).batch(8).map(toMulitTaskOutPut)
for  x,y in trainData:
   #(2, 8, 4)
   #print(y.shape)
   break


inputData=Input(shape=(28, 28, 3))
x=Conv2D(10,kernel_size=(5,5),activation='relu')(inputData)
x=GlobalAvgPool2D()(x)
output1=Dense(4,activation="softmax",name="outPut1")(x)
output2=Dense(4,activation="softmax",name="outPut2")(x)
model=Model(inputData,[output1,output2])
model.summary()
#统一一种损失函数：
#model.compile(optimizer=tf.keras.optimizers.Adam(0.001),loss=tf.keras.losses.CategoricalCrossentropy(),metrics=tf.keras.metrics.categorical_accuracy)
#分开计算的损失函数：转换成字典的形式,重点训练输入2
model.compile(optimizer=tf.keras.optimizers.Adam(0.001),loss={"outPut1":tf.keras.losses.CategoricalCrossentropy(),
                                                              "outPut2":tf.keras.losses.CategoricalCrossentropy()},
              metrics={"outPut1":tf.keras.metrics.categorical_accuracy,"outPut2":tf.keras.metrics.categorical_accuracy},
              loss_weights={"outPut1":1,"outPut2":10})
history=model.fit(x=trainData,epochs=100,validation_data=validData)
back=model.evaluate(x=TextData)
#共输出5个值。对应
print(dict(zip(model.metrics_names, back)))
#绘制代际曲线图
plt.plot(history.history['val_outPut1_categorical_accuracy'],c='red')
plt.plot(history.history['val_outPut2_categorical_accuracy'],c='blue')
plt.legend(["val_outPut1","val_outPut2"])
plt.show()