老规矩–妹妹镇楼:
前面已经介绍过了VGG网络模型,一共13层,这里说的13层指的是10层卷积层和3层全连接层,并没有包括池化层。下面代码详细地将VGG的13层网络模型复现,并用CIFAR100数据集进行训练,测试。
代码中附有详细的注释,从数据的预处理,到训练,再到测试。
import tensorflow as tf
from tensorflow.keras import layers, optimizers, datasets, Sequential
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
tf.random.set_seed(2345)
#VGG13总共13层,指的是10层卷积层和3层全连接层
"""
一共5个单元的卷积层和池化层,每个单元2个卷积层和一个池化层
"""
#卷积层和池化层
conv_layers = [
#第一个单元 2个卷积和一个池化
layers.Conv2D(64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.Conv2D(64, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2, 2], strides=2, padding="same"),
#第二个单元
layers.Conv2D(128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.Conv2D(128, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2, 2], strides=2, padding="same"),
#第三个单元
layers.Conv2D(256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.Conv2D(256, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2, 2], strides=2, padding="same"),
#第四个单元
layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2, 2], strides=2, padding="same"),
#第五个单元
layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.Conv2D(512, kernel_size=[3, 3], padding="same", activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2, 2], strides=2, padding="same"),
]
#数据预处理
def preprocess(x, y):
#数据到0-1之间
x = tf.cast(x, dtype=tf.float32) / 255.
y = tf.cast(y, dtype=tf.int32)
return x,y
#数据自动下载加载
(x, y), (x_test, y_test) = datasets.cifar100.load_data()
print(x.shape, y.shape, x_test.shape, y_test.shape)
#标签 (64, 1),需要squeeze掉 1
y = tf.squeeze(y, axis=1)
y_test = tf.squeeze(y_test, axis=1)
#数据的处理
#训练数据的处理
train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.shuffle(1000).map(preprocess).batch(64)
#测试数据的处理
test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.map(preprocess).batch(64)
#取出一个sample查看 每个样本数据的尺寸,标签的尺寸,最大值最小值
sample = next(iter(train_db))
print("sample:", sample[0].shape, sample[1].shape,
tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))
def main():
#卷积层
conv_net = Sequential(conv_layers)
#test卷积层和池化层
#conv_net.build(input_shape=[None, 32, 32, 3])
# x = tf.random.normal([4, 32, 32, 3])
# out = conv_net(x)
# print(out.shape)
#全连接层
fc_net = Sequential([
layers.Dense(256, activation=tf.nn.relu),
layers.Dense(128, activation=tf.nn.relu),
layers.Dense(100, activation=None),
])
conv_net.build(input_shape=[None, 32, 32, 3])
fc_net.build(input_shape=[None, 512])
#设置优化器,优化学习率
optimizer = optimizers.Adam(lr=1e-4)
#每个层键的参数变量结合起来
variables = conv_net.trainable_variables + fc_net.trainable_variables
#处理完数据后,开始训练
for epoch in range(50):
for step, (x, y) in enumerate(train_db):
#自动求解梯度
with tf.GradientTape() as tape:
#卷积层和池化层
#[b, 32, 32, 3] -> [b, 1, 1, 512]
out = conv_net(x)
#flatten -> [b, 512]
out = tf.reshape(out, [-1, 512])
#全连接层
#[b, 512] -> [b, 100]
logits = fc_net(out)
#走完卷积层,池化层,全连接层得到输出值
#求解Loss值,需要将y的尺寸补齐到logits的尺寸
#[b] -> [b, 100]
y_onehot = tf.one_hot(y, depth=100)
loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
#计算Loss的均值,每个样本有一个loss均值
loss = tf.reduce_mean(loss)
#反向传播,对所有的层间参数进行求导
grads = tape.gradient(loss, variables)
#更新梯度
optimizer.apply_gradients(zip(grads, variables))
#每100个样本打印一次结果
if step%100 == 0:
print(epoch, step, "loss:", float(loss))
#测试模型
total_num = 0
total_correct = 0
for x, y in test_db:
out = conv_net(x)
out = tf.reshape(out, [-1, 512])
logits = fc_net(out)
#softmax转换成概率
prob = tf.nn.softmax(logits, axis=1)
#选择概率最大的作为预测值
pred = tf.argmax(prob, axis=1)
pred = tf.cast(pred, dtype=tf.int32)
#计算正确预测的数量
correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
correct = tf.reduce_sum(correct)
total_num += x.shape[0]
total_correct += int(correct)
#计算正确率
acc = total_correct / total_num
print(epoch, "acc:", acc)
if __name__ == '__main__':
main()