tensorflow2 交叉验证

article/2025/10/6 14:39:41

交叉验证在fit()函数的参数里边，完整参数传送

https://blog.csdn.net/Forrest97/article/details/106635664

fit()里边相关交叉验证的参数

validation_data=test，就是自己划分好的测试集validation_steps， 验证样本总数 Total validation Samples/验证样本大小Validation Batch Size，多少组验证样本的数据后代入网络验证，可能每次输出都能看到验证结果，val_loss 和 val_accuracy 会输出验证精度validation_batch_size：一般是32validation_freq：仅当validation_data设置时有效，表示训练完几组epoch后，进行验证

先给一段代码，以下代码的数据集是60000张图片，50000张来训练集，10000张来测试集。其中每张照片为32*32的彩色照片，每个像素点包括RGB三个数值，一共有10中分类结果：飞机、汽车、鸟、猫咪、鹿子、狗子、小青蛙、马儿、船、大卡车。

先用DNN建模后，存储模型，再重新调用模型。

import keras
import tensorflow as tf
from tensorflow.keras import datasets,layers,optimizers,Sequential,metricsdef preprocess(x,y):x = 2*tf.cast(x,dtype=tf.float32)/225.-1.y = tf.cast(y,dtype=tf.int32)return x,ybatch_size = 128
tf.random.set_seed(1)
(x,y),(test_x,test_y) = datasets.cifar10.load_data()
y = tf.squeeze(y)
test_y = tf.squeeze(test_y)
y = tf.one_hot(y,depth=10)
test_y = tf.one_hot(test_y,depth=10)train = tf.data.Dataset.from_tensor_slices((x,y))
train = train.map(preprocess).shuffle(60000).batch(batch_size=batch_size)
test = tf.data.Dataset.from_tensor_slices((test_x,test_y))
test = test.map(preprocess).batch(batch_size=batch_size)sample = next(iter(train))class mydense(layers.Layer):def __init__(self,input_dim,output_dim):super(mydense, self).__init__()self.kernel = self.add_weight('w',[input_dim,output_dim])self.bias = self.add_weight('b',[1,output_dim])def call(self,inputs,training = None):x = inputs @ self.kernel+self.biasreturn xclass my_network(keras.Model):def __init__(self):super(my_network,self).__init__()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):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 = tf.nn.relu(self.fc4(x))x=self.fc5(x)return x
network = my_network()
network.compile(optimizer=optimizers.Adam(lr=1e-3),loss=tf.losses.CategoricalCrossentropy(from_logits=True),metrics=['accuracy'])
#交叉验证validation_steps ： 验证样本总数 Total validation Samples/验证批量大小Validation Batch Size，多少组验证样本的数据后代入网络验证）
#validation_freq：仅当validation_data设置时有效，表示训练完几组epoch后，进行验证。
network.fit(train,epochs=5,validation_data=test,validation_freq=2)
#print("validation_batch_size",network.fit.validation_freq)
network.evaluate(test)
network.save_weights('weights/mynetwork')
del network
print('saved weights')
network = my_network()
network.compile(optimizer=optimizers.Adam(lr=1e-4),loss=tf.losses.CategoricalCrossentropy(from_logits=True),metrics=['accuracy'])
network.load_weights('weights/mynetwork')
network.fit(train,epochs=5,validation_data=test,validation_batch_size=80)
network.evaluate(test)

输出结果，当设置

validation_freq=2时，第一条fit函数输出是隔一次输出测试集的验证结果，就是在epoch/2的整数倍后，验证数据

Epoch 1/5
391/391 [==============================] - 3s 6ms/step - loss: 1.7262 - accuracy: 0.3867
Epoch 2/5
391/391 [==============================] - 4s 9ms/step - loss: 1.4883 - accuracy: 0.4739 - val_loss: 1.4526 - val_accuracy: 0.4891
Epoch 3/5
391/391 [==============================] - 3s 7ms/step - loss: 1.3768 - accuracy: 0.5139
Epoch 4/5
391/391 [==============================] - 4s 10ms/step - loss: 1.2961 - accuracy: 0.5428 - val_loss: 1.4032 - val_accuracy: 0.5021
Epoch 5/5
391/391 [==============================] - 3s 9ms/step - loss: 1.2186 - accuracy: 0.5733
79/79 [==============================] - 0s 5ms/step - loss: 1.3872 - accuracy: 0.5197
saved weights
Epoch 1/5
391/391 [==============================] - 4s 11ms/step - loss: 1.1536 - accuracy: 0.5942 - val_loss: 1.3735 - val_accuracy: 0.5234
Epoch 2/5
391/391 [==============================] - 4s 11ms/step - loss: 1.0945 - accuracy: 0.6120 - val_loss: 1.3713 - val_accuracy: 0.5252
Epoch 3/5
391/391 [==============================] - 4s 11ms/step - loss: 1.0385 - accuracy: 0.6360 - val_loss: 1.3727 - val_accuracy: 0.5346
Epoch 4/5
391/391 [==============================] - 4s 11ms/step - loss: 0.9848 - accuracy: 0.6520 - val_loss: 1.4310 - val_accuracy: 0.5297
Epoch 5/5
391/391 [==============================] - 5s 12ms/step - loss: 0.9356 - accuracy: 0.6687 - val_loss: 1.4377 - val_accuracy: 0.5266
79/79 [==============================] - 0s 5ms/step - loss: 1.4377 - accuracy: 0.5266

当只设置validation_batch_size=88、88、30，第二条fit函数输出是每次都输出val_loss和val_accuracy，且每次结果都一样，如下：

当只设置validation_steps=30、80，第二条fit函数输出是当30时每次都输出val_loss和val_accuracy，如下：

saved weights
Epoch 1/5
391/391 [==============================] - 3s 9ms/step - loss: 1.1536 - accuracy: 0.5942 - val_loss: 1.3933 - val_accuracy: 0.5185
Epoch 2/5
391/391 [==============================] - 4s 10ms/step - loss: 1.0945 - accuracy: 0.6120 - val_loss: 1.3774 - val_accuracy: 0.5286
Epoch 3/5
391/391 [==============================] - 4s 11ms/step - loss: 1.0385 - accuracy: 0.6360 - val_loss: 1.3880 - val_accuracy: 0.5365
Epoch 4/5
391/391 [==============================] - 4s 11ms/step - loss: 0.9848 - accuracy: 0.6520 - val_loss: 1.4451 - val_accuracy: 0.5326
Epoch 5/5
391/391 [==============================] - 4s 10ms/step - loss: 0.9356 - accuracy: 0.6687 - val_loss: 1.4410 - val_accuracy: 0.5318
79/79 [==============================] - 0s 5ms/step - loss: 1.4377 - accuracy: 0.5266

当设置validation_steps=80，超过验证数据集总数，没有进行验证。

结论：

比较有效果的参数设置：

validation_data和validation_batch_size一起用，validation_batch_size不要超过测试集数据的大小

或者validation_data和validation_freq一起用，validation_freq不要超过epoch的大小