调节使用Keras创建模型的超参数对于结果的影响

最新推荐文章于 2022-01-09 19:00:08 发布

顾弦之

最新推荐文章于 2022-01-09 19:00:08 发布

阅读量1.1k

点赞数

CC 4.0 BY-SA版权

分类专栏：卷积网络深度学习文章标签： CNN卷积网络中的超参数超参数如何调节 Keras 深度学习

本文链接：https://blog.youkuaiyun.com/weixin_42354073/article/details/89366388

本文探讨了超参数如何影响基于Keras的CNN模型的性能。主要关注了学习率和batch_size两个关键超参数。通过实验发现，学习率过大会导致模型训练效果差，而适当降低学习率能获得更好的损失函数和准确率。batch_size的增减对训练集的影响较小，但在测试集上，减小batch_size可能导致泛化能力下降。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

超参数对于CNN模型效果的影响

上一篇文章已经详细介绍使用Keras创建的基础模型，这一篇文章详细探究超参数对于神经网络结果的影响。

模型当前的训练结果

训练样本以及测试样本
X_train shape: (600, 64, 64, 3)
Y_train shape: (600, 1)
X_test shape: (150, 64, 64, 3)
Y_test shape: (150, 1)
输入时人脸图片，分类器需要识别出人是否处于“笑”和“不笑”的状态，笑，输出1，不笑，输出0
优化器的设置

happyModel.compile(optimizer=keras.optimizers.Adam(lr=0.001,beta_1=0.9,beta_2=0.999,epsilon=1e-08, decay=0.0), loss=‘binary_crossentropy’, metrics=[‘accuracy’])

学习率=0.001

训练包的大小以及优化次数设置

happyModel.fit(x=X_train, y=Y_train, batch_size=16, epochs=20)

批量包大小=16，优化次数设置为20
对于训练样本得到的结果如下：

Epoch 1/20
600/600 [==============================] - 6s 10ms/step - loss: 0.5106 - acc: 0.7683
Epoch 2/20
600/600 [==============================] - 4s 7ms/step - loss: 0.2408 - acc: 0.9233
Epoch 3/20
600/600 [==============================] - 4s 6ms/step - loss: 0.1509 - acc: 0.9583
Epoch 4/20
600/600 [==============================] - 4s 6ms/step - loss: 0.1059 - acc: 0.9750
Epoch 5/20
600/600 [==============================] - 4s 6ms/step - loss: 0.0827 - acc: 0.9850
Epoch 6/20
600/600 [==============================] - 4s 6ms/step - loss: 0.0665 - acc: 0.9900
Epoch 7/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0478 - acc: 0.9917
Epoch 8/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0422 - acc: 0.9917
Epoch 9/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0313 - acc: 0.9983
Epoch 10/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0302 - acc: 0.9917
Epoch 11/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0228 - acc: 0.9967
Epoch 12/20
600/600 [==============================] - 4s 6ms/step - loss: 0.0227 - acc: 0.9983
Epoch 13/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0254 - acc: 0.9933A: 0s - loss: 0.0274 - acc:
Epoch 14/20
600/600 [==============================] - 4s 6ms/step - loss: 0.0127 - acc: 0.9967
Epoch 15/20
600/600 [==============================] - 4s 6ms/step - loss: 0.0140 - acc: 0.9967A
Epoch 16/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0090 - acc: 1.0000
Epoch 17/20
600/600 [==============================] - 4s 6ms/step - loss: 0.0153 - acc: 1.0000
Epoch 18/20
600/600 [==============================] - 4s 6ms/step - loss: 0.0114 - acc: 0.9983
Epoch 19/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0084 - acc: 1.0000
Epoch 20/20
600/600 [==============================] - 4s 7ms/step - loss: 0.0123 - acc: 0.9967