【人工智能】实验二基于CNN的图像分类
这个算是个记录吧，这个是黑龙江大学-马吉权老师的人工智能课，这个老师听说科研很好，但是他交的就是有点云里雾里，主要是概念，没有实际代码，令人感叹，老师主要是让大伙自学，不过没有期末考试，还挺好过的。
---------------------------------------------------------------------------------------------------------------------------------
实验的话就是把用cifar100把图片分割为32*32 60k，然后的话主要是分为4个步骤，即
- load datasets
- build network
- train
- test
然后整个分割内容是这样的
这个是老师给的待完成代码
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)
conv_layers = [ # 5 units of conv + max pooling
#Please build vgg13 network according to the demonstration in readme file, bellow is a instance for unit 1
#Please complete other parts. (unit 2 to unit 5)
# unit 1
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'),
# unit 2
#____________________________________________________________________________
#____________________________________________________________________________
#____________________________________________________________________________
# unit 3
#____________________________________________________________________________
#____________________________________________________________________________
#____________________________________________________________________________
# unit 4
#____________________________________________________________________________
#____________________________________________________________________________
#____________________________________________________________________________
# unit 5
#____________________________________________________________________________
#____________________________________________________________________________
#____________________________________________________________________________
]
def preprocess(x, y):
# [0~1]
x = tf.cast(x, dtype=tf.float32) / 255.
y = tf.cast(y, dtype=tf.int32)
return x,y
(x,y), (x_test, y_test) = datasets.cifar100.load_data()
y = tf.squeeze(y, axis=1)
y_test = tf.squeeze(y_test, axis=1)
print(x.shape, y.shape, x_test.shape, y_test.shape)
train_db = tf.data.Dataset.from_tensor_slices((x,y))
train_db = train_db.shuffle(1000).map(preprocess).batch(128)
test_db = tf.data.Dataset.from_tensor_slices((x_test,y_test))
test_db = test_db.map(preprocess).batch(64)
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():
# [b, 32, 32, 3] => [b, 1, 1, 512]
conv_net = Sequential(conv_layers)
# Please add your code in blank
fc_net = Sequential([
layers.Dense(____________________________),
layers.Dense(____________________________),
layers.Dense(100, activation=None), #you can try other activation function to evaluate and compare
])
# Please add your code in blank
conv_net.build(input_shape=[None, ____, _____, ___])
fc_net.build(input_shape=[None, ____])
optimizer = optimizers.Adam(lr=1e-4)
# [1, 2] + [3, 4] => [1, 2, 3, 4]
variables = conv_net.trainable_variables + fc_net.trainable_variables
for epoch in range(50):
for step, (x,y) in enumerate(train_db):
with tf.GradientTape() as tape:
# [b, 32, 32, 3] => [b, 1, 1, 512]
out = conv_net(x)
# flatten, => [b, 512]
out = tf.reshape(out, [-1, 512])
# [b, 512] => [b, 100]
logits = fc_net(out)
# [b] => [b, 100]
y_onehot = tf.one_hot(y, depth=100)
# compute loss
loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
loss = tf.reduce_mean(loss)
# Please add your code in blank
grads = tape.gradient(___________, ______________)
optimizer.apply_gradients(zip(__________, ___________))
if step %100 == 0:
print(epoch, step, 'loss:', float(loss))
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)
prob = tf.nn.softmax(logits, axis=1)
pred = tf.argmax(prob, axis=1)
pred = tf.cast(pred, dtype=tf.int32)
correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
correct = tf.reduce_sum(correct)
total_num += x.shape[0]
total_correct += int(correct)
acc = total_correct / total_num
print(epoch, 'acc:', acc)
if __name__ == '__main__':
main()这个是补完后的代码
import tensorflow as tf
from tensorflow.keras import layers, optimizers, datasets, Sequential
import os
import time
print("tf.version", tf.__version__)
# print("tf is using GPU", tf.test.is_gpu_available())
print("tf is using GPU", tf.config.list_physical_devices(device_type='GPU'))
print(tf.config.experimental.list_physical_devices(device_type='CPU'))
start_time = time.perf_counter()
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
tf.random.set_seed(2345)
#
# 二维卷积层layers.Conv2D：参数(
# 卷积核个数的变化filters
# 表示卷积核的大小kernel_size
# 是否对周围进行填充padding(==“same”四周会填充0，==“valid”不会填充)
# 激活函数activation
# )
# 池化层layers.MaxPool2D：参数(
# 池化核的尺寸pool_size，默认是2×2
# 移动步长strides，默认是池化核尺寸，即2
# 是否对周围进行填充padding(==“same”四周会填充0，==“valid”不会填充)
# )
#
# 首先我们知道Sequential这个容器，接受一个13层的list.我们先组成list；网络的第一部分。
conv_layers = [ # 5 units of conv + max pooling
# unit 1
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'),
# unit 2
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'),
# unit 3
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'),
# unit 4
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'),
# unit 5
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')
]
# 进行数据预处理，仅仅是类型的转换。 [0~1]
# cast数据类型转换函数 tf.cast(x, dtype, name=None) x：待转换的数据(张量)，dtype：目标数据类型，name：可选参数，定义操作的名称
def preprocess(x, y):
x = tf.cast(x, dtype=tf.float32) / 255.
y = tf.cast(y, dtype=tf.int32)
return x, y
# load datasets 根据要求使用的是cifar100
(x, y), (x_test, y_test) = datasets.cifar100.load_data()
y = tf.squeeze(y) # 或者tf.squeeze(y, axis=1)把1维度的squeeze掉。
y_test = tf.squeeze(y_test) # 或者tf.squeeze(y, axis=1)把1维度的squeeze掉。
print(x.shape, y.shape, x_test.shape, y_test.shape)
# from_tensor_slices 将输入的张量的第一个维度看做样本的个数，沿其第一个维度将tensor切片，得到的每个切片是一个样本数据。实现了输入张量的自动切片。
# batch size指的就是更新梯度中使用的样本数。如果把batch_size设置为数据集的长度，就成了批量梯度下降算法，batch_size设置为1就是随机梯度下降算法
# shuffle的功能为打乱dataset中的元素，它有一个参数buffersize，表示打乱时使用的buffer的大小，不设置会报错，buffer_size越大，打乱程度越大
# map接收一个函数，Dataset中的每个元素都会被当作这个函数的输入，并将函数返回值作为新的Dataset
# 训练集
train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.shuffle(1000).map(preprocess).batch(128)
# 测试集
test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.map(preprocess).batch(128)
# 测试一下sample的形状。
sample = next(iter(train_db))
print('sample:', sample[0].shape, sample[1].shape,
tf.reduce_min(sample[0]), tf.reduce_max(sample[0])) # 值范围为[0,1]
def main():
# 把参数放进Sequential容器
# 输入：[b, 32, 32, 3] => 输出[b, 1, 1, 512]
conv_net = Sequential(conv_layers)
# conv_net.build(input_shape=[None, 32, 32, 3])
# x = tf.random.normal([4, 32, 32, 3])
# out = conv_net(x)
# print(out.shape)
# 创建全连接层网络；网络的第二部分；第二部分的输入为第一部分的输出。
fc_net = Sequential([
layers.Dense(256, activation=tf.nn.relu),
layers.Dense(128, activation=tf.nn.relu),
layers.Dense(100, activation=None),
])
# keras.optimizers.Adam 梯度学习率优化器，根据次数的增多，动态的减小学习率，以实现效率和效果的兼得
# （如果学习率过大，很可能会越过最优值，反而如果学习率过小，优化的效率可能很低，导致过长的运算时间）
# 这里其实把一个网络分成2个来写，
conv_net.build(input_shape=[None, 32, 32, 3])
fc_net.build(input_shape=[None, 512])
# 创建一个优化器
#1e-1 99999999
optimizer = optimizers.Adam(learning_rate=0.001)
conv_net.summary()
fc_net.summary()
# 下面的+表示拼接。python中的list列表拼接，2个列表变为一个。
# 例如：[1, 2] + [3, 4] => [1, 2, 3, 4]
# tf.get_variable和tf.Variable的默认选项是True，表示该变量是否可训练
variables = conv_net.trainable_variables + fc_net.trainable_variables
for epoch in range(50):
for step, (x, y) in enumerate(train_db):
with tf.GradientTape() as tape:
# [b, 32, 32, 3] => [b, 1, 1, 512]
out = conv_net(x)
# 之后squeeze或者reshape为平坦的flatten；flatten, => [b, 512]
out = tf.reshape(out, [-1, 512])
# 送入全连接层输入，得到输出logits
# [b, 512] => [b, 100]
logits = fc_net(out)
# [b] => [b, 100]转换为热编码。
y_onehot = tf.one_hot(y, depth=100)
# compute loss 结果维度[b]
loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
loss = tf.reduce_mean(loss)
# 梯度求解
grads = tape.gradient(loss, variables)
# 梯度更新
optimizer.apply_gradients(zip(grads, variables))
if step % 100 == 0:
print(epoch, step, 'loss:', float(loss))
# 做测试
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)
# 预测可能性。
prob = tf.nn.softmax(logits, axis=1)
pred = tf.argmax(prob, axis=1)
pred = tf.cast(pred, dtype=tf.int32)
# 拿到预测值pred和真实值比较。
correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
correct = tf.reduce_sum(correct)
total_num += x.shape[0]
total_correct += int(correct) # 转换为numpy数据
acc = total_correct / total_num
print(epoch, 'acc:', acc)
if __name__ == '__main__':
start=time.perf_counter()
main()
end = time.perf_counter()
print('Running time: %s min %s s' % int((end - start)/60) %((end-start)%60) )keras看自己环境所在位置决定是否要加TensorFlow为前缀
让后末尾记录一个折磨了自己挺久的东西
这个报错是因为TensorFlow和keras版本不匹配 conda install keras就可以了让万能的conda帮你匹配一下。就是说改进的话看看怎么优化比较好。
实际上所有数据跑30次的话正确率和时间都很慢，属于是折磨了。
