【tensorflow】T3.天气图片分类

>- **🍨 本文为[🔗365天深度学习训练营](https://mp.weixin.qq.com/s/0dvHCaOoFnW8SCp3JpzKxg) 中的学习记录博客**
>- **🍖 原作者：[K同学啊](https://mtyjkh.blog.csdn.net/)**

1. 导入数据

import os, PIL, pathlib
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import models, layers, preprocessing

data_dir = "/Volumes/T7 Shield/code/weather_photos/"
data_dir = pathlib.Path(data_dir)

#导入了相关lib和本次需要的dataset

2. 查看数据

image_count = len(list(data_dir.glob("*/*.jpg")))

print(image_count)

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

#out[1]: 1125 （图片总数）

#out[2]: 第一张图片

3. 数据预处理

加载数据：

image_height = 180
image_width = 180
batch_size = 32

train_ds = preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split = 0.2,
    subset = "training",
    seed = 42,
    image_size = (image_height, image_width),
    batch_size = batch_size
    
    )

val_ds = preprocessing.image_dataset_from_directory(
    data_dir,
    validation_split = 0.2,
    subset = "validation",
    seed = 42,
    image_size = (image_height, image_width),
    batch_size = batch_size
    
    )

#out[3]: 

Found 1125 files belonging to 4 classes.
Using 900 files for training.
Found 1125 files belonging to 4 classes.
Using 225 files for validation.

数据可视化：

class_names = train_ds.class_names
print(class_names)

plt.figure(figsize = (20,10))

for images, labels in train_ds.take(1):
    for i in range(20):
        plt.subplot(5, 10, i+1)
        
        plt.imshow(images[i].numpy().astype("uint8"))
        plt.title(class_names[labels[i]])
        plt.axis("off")
        

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

#out[4]: (32, 180, 180, 3)
(32,)

#和一张图

数据集prefetch：

AUTOTUNE = tf.data.AUTOTUNE

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

4. 搭建模型然后summary一下

classes_num = 4

model = models.Sequential([
    layers.experimental.preprocessing.Rescaling(1./255, input_shape = (image_height, image_width, 3)),
    
    layers.Conv2D(16, (3, 3), activation = "relu", input_shape = (image_height, image_width, 3)),
    layers.AveragePooling2D((2,2)),
    layers.Conv2D(32, (3, 3), activation = "relu"),
    layers.AveragePooling2D((2,2)),
    layers.Conv2D(64, (3, 3), activation = "relu"),
    layers.Dropout(0.3),
    
    layers.Flatten(),
    layers.Dense(128, activation = "relu"),
    layers.Dense(classes_num)
    
    ])

model.summary()

#out[5]: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
________________________________________________

 5. 编译和训练模型：

opt = keras.optimizers.legacy.Adam(learning_rate = 0.001)

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

epoch = 10

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

#out[6]:
Epoch 1/10
29/29 [==============================] - 7s 184ms/step - loss: 1.1090 - accuracy: 0.5800 - val_loss: 0.6656 - val_accuracy: 0.7200
Epoch 2/10
29/29 [==============================] - 5s 169ms/step - loss: 0.5341 - accuracy: 0.8044 - val_loss: 0.5162 - val_accuracy: 0.8400
Epoch 3/10
29/29 [==============================] - 5s 171ms/step - loss: 0.4334 - accuracy: 0.8256 - val_loss: 0.4499 - val_accuracy: 0.8356
Epoch 4/10
29/29 [==============================] - 5s 168ms/step - loss: 0.3435 - accuracy: 0.8767 - val_loss: 0.4158 - val_accuracy: 0.8356
Epoch 5/10
29/29 [==============================] - 5s 168ms/step - loss: 0.3020 - accuracy: 0.8722 - val_loss: 0.3915 - val_accuracy: 0.8622
Epoch 6/10
29/29 [==============================] - 5s 164ms/step - loss: 0.2420 - accuracy: 0.9022 - val_loss: 0.4223 - val_accuracy: 0.8622
Epoch 7/10
29/29 [==============================] - 5s 165ms/step - loss: 0.1917 - accuracy: 0.9311 - val_loss: 0.3256 - val_accuracy: 0.8978
Epoch 8/10
29/29 [==============================] - 5s 165ms/step - loss: 0.1495 - accuracy: 0.9500 - val_loss: 0.3683 - val_accuracy: 0.8844
Epoch 9/10
29/29 [==============================] - 5s 165ms/step - loss: 0.1163 - accuracy: 0.9678 - val_loss: 0.4332 - val_accuracy: 0.8578
Epoch 10/10
29/29 [==============================] - 5s 165ms/step - loss: 0.0860 - accuracy: 0.9678 - val_loss: 0.3672 - val_accuracy: 0.8933

6. 模型评估：

acc = history.history["accuracy"]
valacc = history.history["val_accuracy"]

loss = history.history["loss"]
valloss = history.history["val_loss"]

epoch_range = range(epoch)

plt.figure(figsize = (12, 4))

plt.subplot(1, 2, 1)
plt.plot(epoch_range, acc, label = "Training Accuracy")
plt.plot(epoch_range, valacc, label = "Validation Accuracy")
plt.legend(loc = "lower right")
plt.title("Accuracy")

plt.subplot(1, 2, 2)
plt.plot(epoch_range, loss, label = "Training Loss")
plt.plot(epoch_range, valloss, label = "Validation loss")
plt.legend(loc = "lower right")
plt.title("Loss")

plt.show()

#out[7]:

7. 总结

这次加入了Dropout和数据预处理的部分内容，并且把归一化加在了模型中。而且与之前不同的是把learning rate指定了数值，但是调了两下发现都不如0.001好hhhhh。这次包括refetch和image dataset from directory了解了很多数据预处理的内容，而且一边做一边回顾了之前学过的学习率相关的知识，收获不少。