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！
成都创新互联公司是网站建设专家,致力于互联网品牌建设与网络营销，专业领域包括成都网站设计、网站制作、电商网站制作开发、小程序开发、微信营销、系统平台开发，与其他网站设计及系统开发公司不同，我们的整合解决方案结合了恒基网络品牌建设经验和互联网整合营销的理念，并将策略和执行紧密结合，且不断评估并优化我们的方案，为客户提供全方位的互联网品牌整合方案！
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: https://www.kaggle.com/c/dogs-vs-cats/
百度网盘: https://pan.baidu.com/s/13hw4LK8ihR6-6-8mpjLKDA 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！
