这篇文章主要讲解了“TensorFlow2的CNN图像分类方法是什么”，文中的讲解内容简单清晰，易于学习与理解，下面请大家跟着小编的思路慢慢深入，一起来研究和学习“TensorFlow2的CNN图像分类方法是什么”吧！
1. 导包
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import tensorflow as tf
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
2. 图像分类 fashion_mnist
数据处理
# 原始数据
(X_train_all, y_train_all),(X_test, y_test) = tf.keras.datasets.fashion_mnist.load_data()
# 训练集、验证集拆分
X_train, X_valid, y_train, y_valid = train_test_split(X_train_all, y_train_all, test_size=0.25)
# 数据标准化，你也可以用除以255的方式实现归一化
# 注意最后reshape中的1，代表图像只有一个channel，即当前图像是灰度图
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train.reshape(-1, 28 * 28)).reshape(-1, 28, 28, 1)
X_valid_scaled = scaler.transform(X_valid.reshape(-1, 28 * 28)).reshape(-1, 28, 28, 1)
X_test_scaled = scaler.transform(X_test.reshape(-1, 28 * 28)).reshape(-1, 28, 28, 1)
构建CNN模型
model = tf.keras.models.Sequential()
# 多个卷积层
model.add(tf.keras.layers.Conv2D(filters=32, kernel_size=[5, 5], padding="same", activation="relu", input_shape=(28, 28, 1)))
model.add(tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2))
model.add(tf.keras.layers.Conv2D(filters=64, kernel_size=[5, 5], padding="same", activation="relu"))
model.add(tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2))
# 将前面卷积层得出的多维数据转为一维
# 7和前面的kernel_size、padding、MaxPool2D有关
# Conv2D: 28*28 -> 28*28 (因为padding="same")
# MaxPool2D: 28*28 -> 14*14
# Conv2D: 14*14 -> 14*14 (因为padding="same")
# MaxPool2D: 14*14 -> 7*7
model.add(tf.keras.layers.Reshape(target_shape=(7 * 7 * 64,)))
# 传入全连接层
model.add(tf.keras.layers.Dense(1024, activation="relu"))
model.add(tf.keras.layers.Dense(10, activation="softmax"))
# compile
model.compile(loss = "sparse_categorical_crossentropy",
optimizer = "sgd",
metrics = ["accuracy"])
模型训练
callbacks = [
tf.keras.callbacks.EarlyStopping(min_delta=1e-3, patience=5)
]
history = model.fit(X_train_scaled, y_train, epochs=15,
validation_data=(X_valid_scaled, y_valid),
callbacks = callbacks)
Train on 50000 samples, validate on 10000 samples
Epoch 1/15
50000/50000 [==============================] - 17s 343us/sample - loss: 0.5707 - accuracy: 0.7965 - val_loss: 0.4631 - val_accuracy: 0.8323
Epoch 2/15
50000/50000 [==============================] - 13s 259us/sample - loss: 0.3728 - accuracy: 0.8669 - val_loss: 0.3573 - val_accuracy: 0.8738
...
Epoch 13/15
50000/50000 [==============================] - 12s 244us/sample - loss: 0.1625 - accuracy: 0.9407 - val_loss: 0.2489 - val_accuracy: 0.9112
Epoch 14/15
50000/50000 [==============================] - 12s 240us/sample - loss: 0.1522 - accuracy: 0.9451 - val_loss: 0.2584 - val_accuracy: 0.9104
Epoch 15/15
50000/50000 [==============================] - 12s 237us/sample - loss: 0.1424 - accuracy: 0.9500 - val_loss: 0.2521 - val_accuracy: 0.9114
作图
def plot_learning_curves(history):
pd.DataFrame(history.history).plot(figsize=(8, 5))
plt.grid(True)
#plt.gca().set_ylim(0, 1)
plt.show()
plot_learning_curves(history)
测试集评估准确率
model.evaluate(X_test_scaled, y_test)
[0.269884311157465, 0.9071]
可以看到使用CNN后，图像分类的准确率明显提升了。之前的模型是0.8747，现在是0.9071。
3. 图像分类 Dogs vs. Cats
3.1 原始数据
原始数据下载
Kaggle: /tupian/20230522/pp 百度网盘: /tupian/20230522/init 提取码 dmp4
读取一张图片，并展示
image_string = tf.io.read_file("C:/Users/Skey/Downloads/datasets/cat_vs_dog/train/cat.28.jpg")
image_decoded = tf.image.decode_jpeg(image_string)
plt.imshow(image_decoded)
3.2 利用Dataset加载图片
由于原始图片过多，我们不能将所有图片一次加载入内存。Tensorflow为我们提供了便利的Dataset API，可以从硬盘中一批一批的加载数据，以用于训练。
处理本地图片路径与标签
# 训练数据的路径
train_dir = "C:/Users/Skey/Downloads/datasets/cat_vs_dog/train/"
train_filenames = [] # 所有图片的文件名
train_labels = [] # 所有图片的标签
for filename in os.listdir(train_dir):
train_filenames.append(train_dir + filename)
if (filename.startswith("cat")):
train_labels.append(0) # 将cat标记为0
else:
train_labels.append(1) # 将dog标记为1
# 数据随机拆分郑州人流哪家医院做的好 http://www.csyhjlyy.com/
X_train, X_valid, y_train, y_valid = train_test_split(train_filenames, train_labels, test_size=0.2)
定义一个解码图片的方法
def _decode_and_resize(filename, label):
image_string = tf.io.read_file(filename) # 读取图片
image_decoded = tf.image.decode_jpeg(image_string) # 解码
image_resized = tf.image.resize(image_decoded, [256, 256]) / 255.0 # 重置size，并归一化
return image_resized, label
定义 Dataset，用于加载图片数据
# 训练集
train_dataset = tf.data.Dataset.from_tensor_slices((train_filenames, train_labels))
train_dataset = train_dataset.map(
map_func=_decode_and_resize, # 调用前面定义的方法，解析filename，转为特征和标签
num_parallel_calls=tf.data.experimental.AUTOTUNE)
train_dataset = train_dataset.shuffle(buffer_size=128) # 设置缓冲区大小
train_dataset = train_dataset.batch(32) # 每批数据的量
train_dataset = train_dataset.prefetch(tf.data.experimental.AUTOTUNE) # 启动预加载图片，也就是说CPU会提前从磁盘加载数据，不用等上一次训练完后再加载
# 验证集
valid_dataset = tf.data.Dataset.from_tensor_slices((valid_filenames, valid_labels))
valid_dataset = valid_dataset.map(
map_func=_decode_and_resize,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
valid_dataset = valid_dataset.batch(32)
3.3 构建CNN模型，并训练
构建模型与编译
model = tf.keras.Sequential([
# 卷积，32个filter(卷积核)，每个大小为3*3，步长为1
tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(256, 256, 3)),
# 池化，默认大小2*2，步长为2
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Conv2D(32, 5, activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(64, activation='relu'),
tf.keras.layers.Dense(2, activation='softmax')
])
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
loss=tf.keras.losses.sparse_categorical_crossentropy,
metrics=[tf.keras.metrics.sparse_categorical_accuracy]
)
模型总览
model.summary()
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_2 (Conv2D) (None, 254, 254, 32) 896
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 127, 127, 32) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 123, 123, 32) 25632
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 61, 61, 32) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 119072) 0
_________________________________________________________________
dense_2 (Dense) (None, 64) 7620672
_________________________________________________________________
dense_3 (Dense) (None, 2) 130
=================================================================
Total params: 7,647,330
Trainable params: 7,647,330
Non-trainable params: 0
开始训练
model.fit(train_dataset, epochs=10, validation_data=valid_dataset)
由于数据量大，此处训练时间较久
需要注意的是此处打印的step，每个step指的是一个batch(例如32个样本一个batch)
模型评估
test_dataset = tf.data.Dataset.from_tensor_slices((valid_filenames, valid_labels))
test_dataset = test_dataset.map(_decode_and_resize)
test_dataset = test_dataset.batch(32)
print(model.metrics_names)
print(model.evaluate(test_dataset))
感谢各位的阅读，以上就是“TensorFlow2的CNN图像分类方法是什么”的内容了，经过本文的学习后，相信大家对TensorFlow2的CNN图像分类方法是什么这一问题有了更深刻的体会，具体使用情况还需要大家实践验证。这里是创新互联，小编将为大家推送更多相关知识点的文章，欢迎关注！
