CNN识别4种天气状态（TensorFlow，网络优化）

陌上花开 ~

已于 2022-08-29 23:04:12 修改

阅读量1.1k

点赞数 1

分类专栏： # tensorflow # cv 文章标签： tensorflow cnn 深度学习卷积神经网络

于 2022-08-13 22:36:59 首次发布

本文链接：https://blog.csdn.net/qq_45550375/article/details/126325300

版权

cv 同时被 2 个专栏收录

11 篇文章 5 订阅

订阅专栏

tensorflow

5 篇文章 1 订阅

订阅专栏

项目数据及源码

可在github下载：

https://github.com/chenshunpeng/Weather-recognition-based-on-CNN

1.数据处理

设置GPU环境

import tensorflow as tf

gpus = tf.config.list_physical_devices("GPU")

if gpus:
    gpu0 = gpus[0]                                        #如果有多个GPU，仅使用第0个GPU
    tf.config.experimental.set_memory_growth(gpu0, True)  #设置GPU显存用量按需使用
    tf.config.set_visible_devices([gpu0],"GPU")

导入数据

import matplotlib.pyplot as plt
import os,PIL

# 设置随机种子尽可能使结果可以重现
import numpy as np
np.random.seed(1)

# 设置随机种子尽可能使结果可以重现
import tensorflow as tf
tf.random.set_seed(1)

from tensorflow import keras
from tensorflow.keras import layers,models

import pathlib

设置数据地址

data_dir = "E:\demo_study\jupyter\Jupyter_notebook\Weather-recognition-based-on-CNN\weather_photos"
data_dir = pathlib.Path(data_dir)

查看数据

数据集一共分为cloudy、rain、shine、sunrise四类，分别存放于weather_photos文件夹中以各自名字命名的子文件夹中

查看图片总数：

image_count = len(list(data_dir.glob('*/*.jpg')))
print("图片总数为：",image_count)

输出：

图片总数为： 1125

查看第一张图片：

roses = list(data_dir.glob('sunrise/*.jpg'))
PIL.Image.open(str(roses[0]))

输出：

请添加图片描述

1.1.数据预处理

加载数据:

使用image_dataset_from_directory方法将磁盘中的数据加载到tf.data.Dataset中

设置数据参数：

batch_size = 64
img_height = 180
img_width = 180

加载训练集数据：

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.2,
    subset="training",
    seed=123,
    image_size=(img_height, img_width),
    batch_size=batch_size)

输出：

Found 1125 files belonging to 4 classes.
Using 900 files for training.

加载验证集数据：

val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split=0.2,
    subset="validation",
    seed=123,
    image_size=(img_height, img_width),
    batch_size=batch_size)

输出：

Found 1125 files belonging to 4 classes.
Using 225 files for validation.

通过class_names输出数据集的标签，标签按字母顺序对应于目录名称：

class_names = train_ds.class_names
print(class_names)

输出：

['cloudy', 'rain', 'shine', 'sunrise']

查看train_ds的数据类型：

train_ds

输出：

<PrefetchDataset element_spec=(TensorSpec(shape=(None, 180, 180, 3), dtype=tf.float32, name=None), TensorSpec(shape=(None,), dtype=tf.int32, name=None))>

1.2.数据可视化

# 每次画的图不一样

plt.figure(figsize=(12, 10))

for images, labels in train_ds.take(1):
    for i in range(30):
        ax = plt.subplot(5, 6, i + 1)

        plt.imshow(images[i].numpy().astype("uint8"))
        plt.title(class_names[labels[i]])
        plt.savefig('pic1.jpg', dpi=600) #指定分辨率保存
        plt.axis("off")

输出：
请添加图片描述

查看图片形状：

for image_batch, labels_batch in train_ds:
    print(image_batch.shape)
    print(labels_batch.shape)
    break

Image_batch是形状的张量（32,180,180,3）。这是一批形状180x180x3的32张图片（最后一维指的是彩色通道RGB）。
Label_batch是形状（32，）的张量，这些标签对应32张图片

输出：

(64, 180, 180, 3)
(64,)

1.3.配置数据集

shuffle() ：打乱数据，详细可参考：数据集shuffle方法中buffer_size的理解

prefetch() ：预取数据，加速运行，详细可参考：Better performance with the tf.data API

cache() ：将数据集缓存到内存当中，加速运行

推荐一篇博客：【学习笔记】使用tf.data对预处理过程优化

prefetch()功能详细介绍：它使得训练步骤的预处理和模型执行部分重叠起来，原来是：

在这里插入图片描述

prefetch()之后是：

在这里插入图片描述
~~当然，不这么处理也可以的~~

AUTOTUNE = tf.data.AUTOTUNE

train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

2.构建CNN网络

卷积神经网络（CNN）的输入是张量 (Tensor) 形式的： (image_height, image_width, color_channels)，包含了图像高度、宽度及颜色信息（color_channels 为(R,G,B)，分别对应 RGB 的三个颜色通道）

首先通过layers.experimental.preprocessing.Rescaling对图像进行处理（官方解释：rescales and offsets the values of a batch of image (e.g. go from inputs in the [0, 255] range to inputs in the [0, 1] range）

函数详细介绍可看：tf.keras.layers.Rescaling

设置scale=1./255，可将[0, 255]范围内的输入重新缩放为[0, 1]范围内
参数input_shape指明图像的大小格式

2.1.池化层

池化层介绍：

在CNN中通常会在相邻的卷积层之间加入一个池化层，池化层可以有效的缩小参数矩阵的尺寸，从而减少最后连接层的中的参数数量

常见的2类池化层可看下图（借鉴自知乎：传送门）：
在这里插入图片描述

池化层的作用：

缩小特征图一个显而易见的好处是减少参数量，降维、去除冗余信息、对特征进行压缩、简化网络复杂度、减小计算量、减小内存消耗
池化最重要的一个作用是扩大神经元，也就是特征图中一个像素的感受野大小。 在浅层卷积中，特征图还很大，一个像素能接收到的实际图像面积很小。虽然逐层的卷积一定程度上能够沟通相邻的神经元，但作用有限。而池化，通过简单粗暴地合并相邻的若干个神经元，使缩小后的特征图上的神经元能够获得来自原图像更大范围的信息，从而提取出更高阶的特征
平均池化和最大池化分别代表了合并相邻神经元信息的两种策略。平均池化能够更好地保留所有神经元的信息，能够体现出对应区域的平均响应情况；而最大池化则能够保留细节，保留对应区域的最大响应，防止强烈的响应被周围的神经元弱化。

这里用的是平均池化层（官方可看：tf.keras.layers.AveragePooling2D），其模板为：

tf.keras.layers.AveragePooling2D(
    pool_size=(2, 2),
    strides=None,
    padding='valid',
    data_format=None,
    **kwargs
)

在这里我们默认填充形状为valid（有效），此时不添加填充，结果输出的形状是： $\lfloor \dfrac{input\_shape-pool\_size}{strides}\rfloor +1( input\_shape\geq pool\_size)$

如果填充形状为same（有效），此时将添加填充，结果输出的形状是：（特别地，如果跨度为1，则输出形状与输入形状相同） $\lfloor \dfrac{input\_shape-1}{strides}\rfloor +1$

2.2.卷积层

输入图片矩阵 $I$ 大小： $w\times w$
卷积核 $K$ 大小： $k\times k$
步长 $S$ 大小： $s$
填充 $P$ 大小： $p$

卷积输出大小的计算公式为： $o=\dfrac{\left( w-k+2p\right) }{s}+1$

可以用tf.keras.layers.Conv2D()（官方可看：tf.keras.layers.Conv2D）构造卷积核，代码结构为：

tf.keras.layers.Conv2D(
    filters,
    kernel_size,
    strides=(1, 1),
    padding='valid',
    data_format=None,
    dilation_rate=(1, 1),
    groups=1,
    activation=None,
    use_bias=True,
    kernel_initializer='glorot_uniform',
    bias_initializer='zeros',
    kernel_regularizer=None,
    bias_regularizer=None,
    activity_regularizer=None,
    kernel_constraint=None,
    bias_constraint=None,
    **kwargs
)

这里网络构造了3个3x3的卷积核，使用relu激活函数

num_classes = 4

model = models.Sequential([
    layers.experimental.preprocessing.Rescaling(1./255, input_shape=(img_height, img_width, 3)),
    
    layers.Conv2D(16, (3, 3), activation='relu', input_shape=(img_height, img_width, 3)), # 卷积层1，卷积核3*3  
    layers.AveragePooling2D((2, 2)),               # 池化层1，2*2采样
    layers.Conv2D(32, (3, 3), activation='relu'),  # 卷积层2，卷积核3*3
    layers.AveragePooling2D((2, 2)),               # 池化层2，2*2采样
    layers.Conv2D(64, (3, 3), activation='relu'),  # 卷积层3，卷积核3*3
    layers.Dropout(0.3),                    # 防止过拟合，提高模型泛化能力
    
    layers.Flatten(),                       # Flatten层，连接卷积层与全连接层
    layers.Dense(128, activation='relu'),   # 全连接层，特征进一步提取
    layers.Dense(num_classes)               # 输出层，输出预期结果
])

model.summary()  # 打印网络结构

输出如下：

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 rescaling (Rescaling)       (None, 180, 180, 3)       0         
                                                                 
 conv2d (Conv2D)             (None, 178, 178, 16)      448       
                                                                 
 average_pooling2d (AverageP  (None, 89, 89, 16)       0         
 ooling2D)                                                       
                                                                 
 conv2d_1 (Conv2D)           (None, 87, 87, 32)        4640      
                                                                 
 average_pooling2d_1 (Averag  (None, 43, 43, 32)       0         
 ePooling2D)                                                     
                                                                 
 conv2d_2 (Conv2D)           (None, 41, 41, 64)        18496     
                                                                 
 dropout (Dropout)           (None, 41, 41, 64)        0         
                                                                 
 flatten (Flatten)           (None, 107584)            0         
                                                                 
 dense (Dense)               (None, 128)               13770880  
                                                                 
 dense_1 (Dense)             (None, 4)                 516       
                                                                 
=================================================================
Total params: 13,794,980
Trainable params: 13,794,980
Non-trainable params: 0
_________________________________________________________________

2.3.编译设置

损失函数（loss）：用于衡量模型在训练期间的准确率，这里用sparse_categorical_crossentropy，原理与categorical_crossentropy（多类交叉熵损失）一样，不过真实值采用的整数编码（例如第0个类用数字0表示，第3个类用数字3表示，官方可看：tf.keras.losses.SparseCategoricalCrossentropy）
优化器（optimizer）：决定模型如何根据其看到的数据和自身的损失函数进行更新，这里是Adam（官方可看：tf.keras.optimizers.Adam）
评价函数（metrics）：用于监控训练和测试步骤，本次使用accuracy，即被正确分类的图像的比率（官方可看：tf.keras.metrics.Accuracy）

# 设置优化器
opt = tf.keras.optimizers.Adam(learning_rate=0.001)

model.compile(optimizer=opt,
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

2.4.模型训练

epochs = 30

history = model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=epochs
)

这里 $s t e p$ （ $i t er a t i o n$ ）的个数为： $step=\lceil \dfrac{exampleNums∗epoch }{batch size}\rceil=\lceil \dfrac{(1125-225)∗1}{64}\rceil=\lceil 14.0625\rceil=15$

输出：

Epoch 1/30
15/15 [==============================] - 12s 293ms/step - loss: 1.3749 - accuracy: 0.5144 - val_loss: 0.6937 - val_accuracy: 0.6533
Epoch 2/30
15/15 [==============================] - 2s 144ms/step - loss: 0.6076 - accuracy: 0.7800 - val_loss: 0.4826 - val_accuracy: 0.7956
Epoch 3/30
15/15 [==============================] - 2s 144ms/step - loss: 0.3909 - accuracy: 0.8589 - val_loss: 0.4557 - val_accuracy: 0.7911
Epoch 4/30
15/15 [==============================] - 2s 145ms/step - loss: 0.2943 - accuracy: 0.8878 - val_loss: 0.4268 - val_accuracy: 0.8356
Epoch 5/30
15/15 [==============================] - 2s 144ms/step - loss: 0.2307 - accuracy: 0.9056 - val_loss: 0.4260 - val_accuracy: 0.8400
Epoch 6/30
15/15 [==============================] - 2s 144ms/step - loss: 0.2000 - accuracy: 0.9267 - val_loss: 0.3143 - val_accuracy: 0.8711
Epoch 7/30
15/15 [==============================] - 2s 144ms/step - loss: 0.1496 - accuracy: 0.9367 - val_loss: 0.3277 - val_accuracy: 0.8844
Epoch 8/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0894 - accuracy: 0.9678 - val_loss: 0.2851 - val_accuracy: 0.9200
Epoch 9/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0638 - accuracy: 0.9800 - val_loss: 0.4995 - val_accuracy: 0.8578
Epoch 10/30
15/15 [==============================] - 2s 145ms/step - loss: 0.1132 - accuracy: 0.9622 - val_loss: 0.4961 - val_accuracy: 0.8356
Epoch 11/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0576 - accuracy: 0.9800 - val_loss: 0.3318 - val_accuracy: 0.8844
Epoch 12/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0417 - accuracy: 0.9878 - val_loss: 0.5433 - val_accuracy: 0.8756
Epoch 13/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0301 - accuracy: 0.9944 - val_loss: 0.3797 - val_accuracy: 0.8978
Epoch 14/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0401 - accuracy: 0.9833 - val_loss: 0.3982 - val_accuracy: 0.8489
Epoch 15/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0260 - accuracy: 0.9922 - val_loss: 0.4777 - val_accuracy: 0.8844
Epoch 16/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0139 - accuracy: 0.9978 - val_loss: 0.3858 - val_accuracy: 0.8978
Epoch 17/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0067 - accuracy: 0.9989 - val_loss: 0.3942 - val_accuracy: 0.9156
Epoch 18/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0059 - accuracy: 0.9989 - val_loss: 0.4101 - val_accuracy: 0.8844
Epoch 19/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0039 - accuracy: 1.0000 - val_loss: 0.5176 - val_accuracy: 0.8889
Epoch 20/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0068 - accuracy: 0.9989 - val_loss: 0.3836 - val_accuracy: 0.9156
Epoch 21/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0116 - accuracy: 0.9989 - val_loss: 0.4635 - val_accuracy: 0.8800
Epoch 22/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0062 - accuracy: 0.9989 - val_loss: 0.4315 - val_accuracy: 0.9022
Epoch 23/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0047 - accuracy: 1.0000 - val_loss: 0.5728 - val_accuracy: 0.8933
Epoch 24/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0090 - accuracy: 0.9989 - val_loss: 0.5049 - val_accuracy: 0.8889
Epoch 25/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0210 - accuracy: 0.9911 - val_loss: 0.5712 - val_accuracy: 0.8756
Epoch 26/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0354 - accuracy: 0.9878 - val_loss: 0.6332 - val_accuracy: 0.8889
Epoch 27/30
15/15 [==============================] - 2s 146ms/step - loss: 0.0781 - accuracy: 0.9789 - val_loss: 0.5726 - val_accuracy: 0.8578
Epoch 28/30
15/15 [==============================] - 2s 144ms/step - loss: 0.0713 - accuracy: 0.9767 - val_loss: 0.5084 - val_accuracy: 0.8889
Epoch 29/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0439 - accuracy: 0.9822 - val_loss: 0.5302 - val_accuracy: 0.8711
Epoch 30/30
15/15 [==============================] - 2s 145ms/step - loss: 0.0343 - accuracy: 0.9878 - val_loss: 0.4488 - val_accuracy: 0.8933

保存模型：

model.save('model')

输出日志：

WARNING:absl:Found untraced functions such as _jit_compiled_convolution_op, _jit_compiled_convolution_op, _jit_compiled_convolution_op while saving (showing 3 of 3). These functions will not be directly callable after loading.
INFO:tensorflow:Assets written to: model\assets
INFO:tensorflow:Assets written to: model\assets

3.模型评估（acc：92.00%）

model = tf.keras.models.load_model('model')
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']

loss = history.history['loss']
val_loss = history.history['val_loss']

epochs_range = range(epochs)

plt.figure(figsize=(12, 5))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, acc, label='Training Accuracy')
plt.plot(epochs_range, val_acc, label='Validation Accuracy')
plt.legend()
plt.title('Training and Validation Accuracy')

plt.subplot(1, 2, 2)
plt.plot(epochs_range, loss, label='Training Loss')
plt.plot(epochs_range, val_loss, label='Validation Loss')
plt.legend()
plt.title('Training and Validation Loss')
plt.savefig('pic2.jpg', dpi=600) #指定分辨率保存
plt.show()

输出：
请添加图片描述

4.模型优化（acc：93.78%）

优化思路：

Training loss 不断下降，Validation loss趋于不变，说明网络过拟合，我们尝试将网络结构设置的更复杂一些，学习率调小一些

4.1.优化网络

num_classes = 4


model = models.Sequential([
    layers.experimental.preprocessing.Rescaling(1./255, input_shape=(img_height, img_width, 3)),
    
    layers.Conv2D(16, (3, 3), activation='relu', input_shape=(img_height, img_width, 3)), # 卷积层1，卷积核3*3  
    layers.AveragePooling2D((2, 2)),               # 池化层1，2*2采样
    layers.Conv2D(32, (3, 3), activation='relu'),  # 卷积层2，卷积核3*3
    layers.AveragePooling2D((2, 2)),
    layers.Conv2D(64, (3, 3), activation='relu'),  # 卷积层2，卷积核3*3
    layers.AveragePooling2D((2, 2)),               # 池化层2，2*2采样
    layers.Conv2D(128, (3, 3), activation='relu'),  # 卷积层3，卷积核3*3
    layers.Dropout(0.3),  
    
    layers.Flatten(),                       # Flatten层，连接卷积层与全连接层
    layers.Dense(256, activation='relu'),   # 全连接层，特征进一步提取
    layers.Dense(128, activation='relu'),   # 全连接层，特征进一步提取
    layers.Dense(num_classes)               # 输出层，输出预期结果
])

model.summary()  # 打印网络结构

输出：

Model: "sequential_7"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 rescaling_7 (Rescaling)     (None, 180, 180, 3)       0         
                                                                 
 conv2d_25 (Conv2D)          (None, 178, 178, 16)      448       
                                                                 
 average_pooling2d_18 (Avera  (None, 89, 89, 16)       0         
 gePooling2D)                                                    
                                                                 
 conv2d_26 (Conv2D)          (None, 87, 87, 32)        4640      
                                                                 
 average_pooling2d_19 (Avera  (None, 43, 43, 32)       0         
 gePooling2D)                                                    
                                                                 
 conv2d_27 (Conv2D)          (None, 41, 41, 64)        18496     
                                                                 
 average_pooling2d_20 (Avera  (None, 20, 20, 64)       0         
 gePooling2D)                                                    
                                                                 
 conv2d_28 (Conv2D)          (None, 18, 18, 128)       73856     
                                                                 
 dropout_8 (Dropout)         (None, 18, 18, 128)       0         
                                                                 
 flatten_7 (Flatten)         (None, 41472)             0         
                                                                 
 dense_16 (Dense)            (None, 256)               10617088  
                                                                 
 dense_17 (Dense)            (None, 128)               32896     
                                                                 
 dense_18 (Dense)            (None, 4)                 516       
                                                                 
=================================================================
Total params: 10,747,940
Trainable params: 10,747,940
Non-trainable params: 0
_________________________________________________________________

4.2.优化学习率

# 设置优化器
opt = tf.keras.optimizers.Adam(learning_rate=0.0008)

model.compile(optimizer=opt,
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

训练模型：

epochs = 50

history = model.fit(
  train_ds,
  validation_data=val_ds,
  epochs=epochs
)

输出：

Epoch 1/50
15/15 [==============================] - 2s 131ms/step - loss: 1.0564 - accuracy: 0.4989 - val_loss: 0.7628 - val_accuracy: 0.5689
Epoch 2/50
15/15 [==============================] - 2s 123ms/step - loss: 0.7389 - accuracy: 0.6589 - val_loss: 0.8257 - val_accuracy: 0.6711
Epoch 3/50
15/15 [==============================] - 2s 126ms/step - loss: 0.5849 - accuracy: 0.7667 - val_loss: 0.5710 - val_accuracy: 0.7422
Epoch 4/50
15/15 [==============================] - 2s 125ms/step - loss: 0.4292 - accuracy: 0.8389 - val_loss: 0.6052 - val_accuracy: 0.7689
Epoch 5/50
15/15 [==============================] - 2s 140ms/step - loss: 0.3802 - accuracy: 0.8511 - val_loss: 0.7504 - val_accuracy: 0.7556
Epoch 6/50
15/15 [==============================] - 2s 123ms/step - loss: 0.3367 - accuracy: 0.8667 - val_loss: 0.4836 - val_accuracy: 0.8311
Epoch 7/50
15/15 [==============================] - 2s 124ms/step - loss: 0.2773 - accuracy: 0.8889 - val_loss: 0.3823 - val_accuracy: 0.8622
Epoch 8/50
15/15 [==============================] - 2s 125ms/step - loss: 0.2457 - accuracy: 0.9067 - val_loss: 0.3668 - val_accuracy: 0.8622
Epoch 9/50
15/15 [==============================] - 2s 126ms/step - loss: 0.2333 - accuracy: 0.9144 - val_loss: 0.4030 - val_accuracy: 0.8489
Epoch 10/50
15/15 [==============================] - 2s 124ms/step - loss: 0.2526 - accuracy: 0.9089 - val_loss: 0.6440 - val_accuracy: 0.8044
Epoch 11/50
15/15 [==============================] - 2s 123ms/step - loss: 0.2331 - accuracy: 0.9122 - val_loss: 0.4930 - val_accuracy: 0.8444
Epoch 12/50
15/15 [==============================] - 2s 126ms/step - loss: 0.1934 - accuracy: 0.9311 - val_loss: 0.3481 - val_accuracy: 0.8844
Epoch 13/50
15/15 [==============================] - 2s 124ms/step - loss: 0.1471 - accuracy: 0.9367 - val_loss: 0.3174 - val_accuracy: 0.9022
Epoch 14/50
15/15 [==============================] - 2s 122ms/step - loss: 0.1141 - accuracy: 0.9656 - val_loss: 0.4393 - val_accuracy: 0.8578
Epoch 15/50
15/15 [==============================] - 2s 125ms/step - loss: 0.0878 - accuracy: 0.9700 - val_loss: 0.4360 - val_accuracy: 0.8978
Epoch 16/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0718 - accuracy: 0.9744 - val_loss: 0.3478 - val_accuracy: 0.8978
Epoch 17/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0561 - accuracy: 0.9844 - val_loss: 0.3561 - val_accuracy: 0.9200
Epoch 18/50
15/15 [==============================] - 2s 124ms/step - loss: 0.1152 - accuracy: 0.9544 - val_loss: 0.3755 - val_accuracy: 0.9111
Epoch 19/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0642 - accuracy: 0.9778 - val_loss: 0.3634 - val_accuracy: 0.8978
Epoch 20/50
15/15 [==============================] - 2s 124ms/step - loss: 0.0347 - accuracy: 0.9922 - val_loss: 0.3544 - val_accuracy: 0.8978
Epoch 21/50
15/15 [==============================] - 2s 132ms/step - loss: 0.0432 - accuracy: 0.9844 - val_loss: 0.7549 - val_accuracy: 0.8222
Epoch 22/50
15/15 [==============================] - 2s 124ms/step - loss: 0.0641 - accuracy: 0.9811 - val_loss: 0.4202 - val_accuracy: 0.8933
Epoch 23/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0295 - accuracy: 0.9900 - val_loss: 0.4618 - val_accuracy: 0.9200
Epoch 24/50
15/15 [==============================] - 2s 124ms/step - loss: 0.0131 - accuracy: 0.9978 - val_loss: 0.4210 - val_accuracy: 0.9067
Epoch 25/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0172 - accuracy: 0.9944 - val_loss: 0.4878 - val_accuracy: 0.8978
Epoch 26/50
15/15 [==============================] - 2s 125ms/step - loss: 0.0086 - accuracy: 0.9978 - val_loss: 0.4908 - val_accuracy: 0.9111
Epoch 27/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0096 - accuracy: 0.9967 - val_loss: 0.5744 - val_accuracy: 0.8978
Epoch 28/50
15/15 [==============================] - 2s 122ms/step - loss: 0.0058 - accuracy: 0.9989 - val_loss: 0.5868 - val_accuracy: 0.8889
Epoch 29/50
15/15 [==============================] - 2s 126ms/step - loss: 0.0196 - accuracy: 0.9911 - val_loss: 0.5722 - val_accuracy: 0.8578
Epoch 30/50
15/15 [==============================] - 2s 126ms/step - loss: 0.0108 - accuracy: 0.9967 - val_loss: 0.5498 - val_accuracy: 0.9022
Epoch 31/50
15/15 [==============================] - 2s 125ms/step - loss: 0.0041 - accuracy: 1.0000 - val_loss: 0.5006 - val_accuracy: 0.9111
Epoch 32/50
15/15 [==============================] - 2s 124ms/step - loss: 0.0020 - accuracy: 1.0000 - val_loss: 0.5010 - val_accuracy: 0.9067
Epoch 33/50
15/15 [==============================] - 2s 130ms/step - loss: 0.0067 - accuracy: 0.9967 - val_loss: 0.5418 - val_accuracy: 0.9156
Epoch 34/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0092 - accuracy: 0.9967 - val_loss: 0.6289 - val_accuracy: 0.9067
Epoch 35/50
15/15 [==============================] - 2s 122ms/step - loss: 0.0152 - accuracy: 0.9933 - val_loss: 0.5908 - val_accuracy: 0.8978
Epoch 36/50
15/15 [==============================] - 2s 123ms/step - loss: 0.0359 - accuracy: 0.9911 - val_loss: 0.5404 - val_accuracy: 0.8756
Epoch 37/50
15/15 [==============================] - 2s 124ms/step - loss: 0.0492 - accuracy: 0.9811 - val_loss: 0.5009 - val_accuracy: 0.9022
Epoch 38/50
15/15 [==============================] - 2s 125ms/step - loss: 0.0395 - accuracy: 0.9844 - val_loss: 0.4718 - val_accuracy: 0.9244
Epoch 39/50
15/15 [==============================] - 2s 125ms/step - loss: 0.0357 - accuracy: 0.9956 - val_loss: 0.6038 - val_accuracy: 0.8844
Epoch 40/50
15/15 [==============================] - 2s 124ms/step - loss: 0.0309 - accuracy: 0.9889 - val_loss: 0.4189 - val_accuracy: 0.9200
Epoch 41/50
15/15 [==============================] - 2s 126ms/step - loss: 0.0093 - accuracy: 0.9989 - val_loss: 0.5180 - val_accuracy: 0.9067
Epoch 42/50
15/15 [==============================] - 2s 125ms/step - loss: 0.0033 - accuracy: 1.0000 - val_loss: 0.4415 - val_accuracy: 0.9244
Epoch 43/50
15/15 [==============================] - 2s 124ms/step - loss: 0.0016 - accuracy: 1.0000 - val_loss: 0.4622 - val_accuracy: 0.9378
Epoch 44/50
15/15 [==============================] - 2s 125ms/step - loss: 5.7711e-04 - accuracy: 1.0000 - val_loss: 0.4805 - val_accuracy: 0.9333
Epoch 45/50
15/15 [==============================] - 2s 125ms/step - loss: 4.1283e-04 - accuracy: 1.0000 - val_loss: 0.4820 - val_accuracy: 0.9244
Epoch 46/50
15/15 [==============================] - 2s 126ms/step - loss: 3.2792e-04 - accuracy: 1.0000 - val_loss: 0.4859 - val_accuracy: 0.9333
Epoch 47/50
15/15 [==============================] - 2s 127ms/step - loss: 2.7573e-04 - accuracy: 1.0000 - val_loss: 0.4932 - val_accuracy: 0.9289
Epoch 48/50
15/15 [==============================] - 2s 123ms/step - loss: 2.7769e-04 - accuracy: 1.0000 - val_loss: 0.4877 - val_accuracy: 0.9333
Epoch 49/50
15/15 [==============================] - 2s 124ms/step - loss: 2.6387e-04 - accuracy: 1.0000 - val_loss: 0.5107 - val_accuracy: 0.9289
Epoch 50/50
15/15 [==============================] - 2s 122ms/step - loss: 2.1140e-04 - accuracy: 1.0000 - val_loss: 0.4979 - val_accuracy: 0.9378

可以发现val_accuracy的最大值在训练的第50次，为93.78%

训练过程可视化图：
请添加图片描述

陌上花开 ~

关注

1
点赞
踩
18

收藏

觉得还不错? 一键收藏
1
评论
CNN识别4种天气状态（TensorFlow，网络优化）

TensorFlow识别4种天气状态（CNN，模型ACC：93.78%）在CNN中通常会在相邻的卷积层之间加入一个池化层，池化层可以有效的缩小参数矩阵的尺寸，从而减少最后连接层的中的参数数量。
复制链接

扫一扫