TensorFlow 是一个端到端开源机器学习平台，由谷歌大脑开发和维护。TensorFlow对深度学习常用的数学方法和数据操作进行了封装，使用者利用TensorFlow可以更加方便快速地搭建和训练神经网络。
可以将许多层次加入一个普通的Python列表然后作为参数传入模型来构造顺序结构的神经网络，如果想搭建如GoogLeNet一样存在分支的网络这可以使用tensorflow.keras.layers.concatenate实现，将多个层次构成的列表作为参数inputs传入方法可以做到。
本文用LeNet进行数字识别，LeNet结构简单易于理解和实现但涉及到的知识一点都不少，是入手深度学习框架的不错的选择。LeNet-5是最早的卷积神经网络之一。论文提出的卷积层、池化层的概念，也提到本文所使用的MNIST数据集。MNIST数据集已经集成进了Tensorflow框架，可以通过调用一个方法直接使用。
环境
Cuda compilation tools, release 11.2, V11.2.152
Build cuda_11.2.r11.2/compiler.29618528_0
Python 3.8.10
tensorflow-gpu 2.5.0
代码
#!~/.conda/envs python
# -*- coding: utf-8 -*-
import functools # 来源Python标准库的包，用于处理函数和可调用对象的工具
import numpy as np # Numpy是常用数学包，提供了很多数学计算方法
import tensorflow.keras as keras # https://keras.io/zh/ 可看作Tensorflow的一个子包，但也可单独使用
import tensorflow.keras.optimizers as opts # 引入优化器
from tensorflow import ( # 提倡在导入大量模块和方法时加括号
GradientTape, # 自动求导类，求导之前得先监听变量，用其监听的变量所做的各种操作来确定这个导该怎么求
Tensor, # 描述矩阵的类，调试中可以看到其不同实例的名字互相关联
clip_by_value, # 用来限制变量取值范围的方法，给这个变量设置上下限
convert_to_tensor, # 将列表或者numpy.ndarray转换为Tensor
float32, # Tensorflow提供32位浮点数
reshape # 对Tensor形状进行变更
)
from tensorflow import math as tfmath # 框架内进行一些数学运算，用的是自己封装的类和方法
from tensorflow import reduce_max # 求Tensor矩阵中值最大的一个元素
from tensorflow.keras.layers import (AvgPool2D, Conv2D, Dense, Flatten, InputLayer) # 导入所需层
from tensorflow.keras.losses import BinaryCrossentropy # 框架自带交叉熵损失函数，这次我们不用这个，我们自己来写
from tensorflow.python.keras.callbacks import Callback # 每轮拟合都调用一次的回调函数类的基类
from tensorflow.python.keras.engine import data_adapter # 将不同的输入数据转换为tf.dataset
np.set_printoptions(suppress=True, threshold=np.inf) # 设置不用科学计数法
SHAPE = (-1, 28, 28, 1) #神经网络输入形状，-1代表自动设置长度
(x, y), (x_t, y_t) = keras.datasets.mnist.load_data() # 读数据，读取出来的数据形状为((n,28,28,1),(n,10)),((m,28,28,1),(m,10))
x = reshape(x, SHAPE) # 改成(-1, 28, 28, 1)是要描述图片的通道数，1就是灰度图片，3应该是RGB彩色图片
x_t = reshape(x_t, SHAPE) # 同上
y = [[1 if j == y[i] else 0 for j in range(10)] for i in range(len(y))] # 这里y的值就是0～9，这不利于计算，提前将1个值处理成10个值
y_t = [[1 if j == y_t[i] else 0 for j in range(10)] for i in range(len(y_t))] # 同上
# 取一定量的数据减少内存占用，否则就溢出
rands = np.random.randint(0, len(x), (2000,)) # 随机一些序号
x = [x[i] for i in range(len(x)) if i in rands] # 按照序号获取列表中的一些数值
y = [y[i] for i in range(len(y)) if i in rands] # 按照序号获取列表中的一些数值
x = convert_to_tensor(x) # 将列表转换为Tensor对象
y = convert_to_tensor(y, dtype=float32) # 将列表转换为Tensor对象并设置Tensor对象的dtype
# dtype描述了对象内部存储的数据类型，这个Tensor中的每一个值都转化为这个类型
print('x.shape = %s, y.shape = %s' % (x.shape, y.shape)) # 打印数据形状
# 如果y的第0维度与输出层神经元数目不同就会影响损失函数工作，除非在损失函数中将y转换成正确的形状
EPOCHS = 500 # 最大拟合500轮
NETWORK = [
InputLayer(input_shape=SHAPE[1:], name="input"), # 输入层，确定第一个隐层的权重矩阵尺寸
Conv2D(
filters=32, # 卷积核数目，有许多神经元对输入做一致的操作，不妨让它们的权重矩阵就也可以同步变化，换种说法：共享权重
kernel_size=(3, 3), # 卷积核大小，就是权重矩阵的shape
activation='relu', # 激励函数的名字一定要写对
padding='same', # 这个参数描述卷积层是否对输入进行填充，默认是'valid'（不填充），具体怎么填充是框架里已经做好的事情
name="Conv_1" # 给这一层起个名字
),
AvgPool2D( # 平均池化层，对输入的几个相邻值合并成一个值（求平均）来降维
pool_size=(2, 2), # 描述“窗口”大小，对多少值求平均，这里取了2行2列4个值求平均
strides=None, # “窗口”的步长，为None时就自动让“窗口”永不重合
padding='valid', # 无填充
name="Pool_1" # 名字
),
Conv2D(filters=64, kernel_size=(3, 3), activation='relu', padding='same', name="Conv_2"), # 同一网络中层的名字要互不相同
AvgPool2D(name="Pool_2"), # 输出的shape应该是(None,7,7,64)
Flatten(), # 平展层用来改变矩阵形状，其没有可训练参数，仅仅将输入的矩阵处理成全连接层所接受的形状
Dense(units=64, activation='relu', name="Dense_1"), # 经过平展层后输入应该是(None,3136)
Dense(units=10, activation='sigmoid', name="Dense_2") # 每一个神经元用上一层的64个值判断输入是否是自己的输出的类别
]
model = keras.Sequential(layers=NETWORK, name='LeNet') # 将层顺序连接形成神经网络模型
def step(self, data): # 定义训练步骤方法，模型每轮拟合都进行
print('data type: %s' % type(data)) # data是tuple类型的
print(data) # data里面有两个Tensor对象
x, y, _ = data_adapter.unpack_x_y_sample_weight(data) # 把元组拆成单个值
print(x.shape) # 这里有一行输出，在训练过程中并不会显示出来
with GradientTape(persistent=True, watch_accessed_variables=False) as tape:
tape.watch(self.trainable_variables) # 监控网络参数进行过的运算
y_hat = self(x, training=True) # 得出模型输出
loss = self.loss(y, y_hat) # 计算模型损失
grads = tape.gradient(loss, self.trainable_variables) # GradientTape 自动求导
self.optimizer.apply_gradients( # 优化网络参数
zip(grads, self.trainable_variables) # 函数用于将可迭代的对象作为参数，将对象中对应的元素打包成一个个元组，然后返回由这些元组组成的列表。
)
accuracy = tfmath.reduce_mean((y_hat*y)+((1-y_hat)*(1-y))) # 计算正确率，不算也行，但是早停需要用所以就进行计算
return {'loss': loss, 'accuracy': accuracy} # 返回值是字典
class Callback_(Callback): # 定义回调类
def on_epoch_end(self, epoch, logs={}): # self就是Model对象的指针
acc = logs.get('accuracy') # logs是step的返回值
if acc > 0.97: # 如果能达到97%的训练误差就停止训练
self.model.stop_training = True # 设置停止训练
def log(v: Tensor): # 定义一个求对数的方法
return tfmath.log(clip_by_value(v, 1e-8, reduce_max(v))) # 限制v的取值防止出现nan
def binary_crossentropy_(y, y_hat): # 自定义的交叉熵损失函数
loss = log(y_hat) * y + (log(1 - y_hat) * (1-y)) # L_i=log_e(hat{y}_i)y_i+log_e(1-hat{y}_i)(1-y_i)
loss = -tfmath.reduce_mean(loss) # (1/N)∑_{i}L_i
return loss # 返回损失值
opt = opts.Nadam(learning_rate=.0001) # 定义优化器
model.compile( # 模型训练相关配置
optimizer=opt, # 指定优化器
loss=binary_crossentropy_, # 使用我们上面定义的损失函数
# loss=BinaryCrossentropy(from_logits=True), # 使用框架自带的损失函数
metrics=['accuracy'] # 定义模型评估指标，名字和上面一致
)
model.train_step = functools.partial(step, model) # partial负责把model赋给step的self参数
model.fit(x, y, epochs=EPOCHS, callbacks=[Callback_()]) # 填入数据和标签，定义回调列表
model.summary() # 打印网络结构
rands = np.random.randint(0, len(y_t), (30,)) # 取30个测试数据进行打印
for index in rands: # 遍历30个序号
y_hat = model.predict(convert_to_tensor([x_t[index]])) # 非训练下获取网络输出
mi = np.argmax(y_t[index]) # 算出第index个测试数据是什么数字
mti = np.argmax(y_hat[0]) # 算出模型计算出第index个测试数据是什么数字
print('[%s]::%s' % (', '.join(['%s%.04f\033[0m' % (('\033[32m'if i == mi else '\033[31m') if i == mti else '', y_hat[0][i]) for i in range(len((y_hat[0])))]), mi)) # 打印输出
控制台
2021-08-04 09:05:56.073437: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
2021-08-04 09:05:58.276934: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcuda.so.1
2021-08-04 09:05:58.312833: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:58.313281: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1733] Found device 0 with properties:
pciBusID: 0000:01:00.0 name: NVIDIA GeForce GTX 1050 Ti computeCapability: 6.1
coreClock: 1.62GHz coreCount: 6 deviceMemorySize: 3.95GiB deviceMemoryBandwidth: 104.43GiB/s
2021-08-04 09:05:58.313349: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
2021-08-04 09:05:58.332447: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublas.so.11
2021-08-04 09:05:58.332597: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublasLt.so.11
2021-08-04 09:05:58.342695: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcufft.so.10
2021-08-04 09:05:58.347238: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcurand.so.10
2021-08-04 09:05:58.352323: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcusolver.so.11
2021-08-04 09:05:58.357109: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcusparse.so.11
2021-08-04 09:05:58.358519: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudnn.so.8
2021-08-04 09:05:58.358676: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:58.359252: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:58.360137: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1871] Adding visible gpu devices: 0
2021-08-04 09:05:58.361005: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2021-08-04 09:05:58.361972: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:58.362384: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1733] Found device 0 with properties:
pciBusID: 0000:01:00.0 name: NVIDIA GeForce GTX 1050 Ti computeCapability: 6.1
coreClock: 1.62GHz coreCount: 6 deviceMemorySize: 3.95GiB deviceMemoryBandwidth: 104.43GiB/s
2021-08-04 09:05:58.362536: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:58.362903: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:58.363168: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1871] Adding visible gpu devices: 0
2021-08-04 09:05:58.363670: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
2021-08-04 09:05:59.532060: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1258] Device interconnect StreamExecutor with strength 1 edge matrix:
2021-08-04 09:05:59.532138: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1264] 0
2021-08-04 09:05:59.532165: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1277] 0: N
2021-08-04 09:05:59.532560: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:59.533604: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:59.534597: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 09:05:59.535497: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1418] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 2896 MB memory) -> physical GPU (device: 0, name: NVIDIA GeForce GTX 1050 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1)
x.shape = (1965, 28, 28, 1), y.shape = (1965, 10)
2021-08-04 09:06:01.174218: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:176] None of the MLIR Optimization Passes are enabled (registered 2)
2021-08-04 09:06:01.196600: I tensorflow/core/platform/profile_utils/cpu_utils.cc:114] CPU Frequency: 2299965000 Hz
Epoch 1/500
data type: <class 'tuple'>
(<tf.Tensor 'IteratorGetNext:0' shape=(None, 28, 28, 1) dtype=uint8>, <tf.Tensor 'IteratorGetNext:1' shape=(None, 10) dtype=float32>)
(None, 28, 28, 1)
data type: <class 'tuple'>
(<tf.Tensor 'IteratorGetNext:0' shape=(None, 28, 28, 1) dtype=uint8>, <tf.Tensor 'IteratorGetNext:1' shape=(None, 10) dtype=float32>)
(None, 28, 28, 1)
2021-08-04 09:06:01.823192: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudnn.so.8
2021-08-04 09:06:02.627137: I tensorflow/stream_executor/cuda/cuda_dnn.cc:359] Loaded cuDNN version 8101
2021-08-04 09:06:03.810120: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublas.so.11
2021-08-04 09:06:04.736986: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcublasLt.so.11
62/62 [==============================] - 4s 5ms/step - loss: 0.4304 - accuracy: 0.8350
Epoch 2/500
62/62 [==============================] - 0s 4ms/step - loss: 0.1316 - accuracy: 0.9237
Epoch 3/500
62/62 [==============================] - 0s 4ms/step - loss: 0.0837 - accuracy: 0.9487
Model: "LeNet"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
Conv_1 (Conv2D) (None, 28, 28, 32) 320
_________________________________________________________________
Pool_1 (AveragePooling2D) (None, 14, 14, 32) 0
_________________________________________________________________
Conv_2 (Conv2D) (None, 14, 14, 64) 18496
_________________________________________________________________
Pool_2 (AveragePooling2D) (None, 7, 7, 64) 0
_________________________________________________________________
flatten (Flatten) (None, 3136) 0
_________________________________________________________________
Dense_1 (Dense) (None, 64) 200768
_________________________________________________________________
Dense_2 (Dense) (None, 10) 650
=================================================================
Total params: 220,234
Trainable params: 220,234
Non-trainable params: 0
_________________________________________________________________
[0.0044, 0.1280, 0.0010, 0.8811, 0.0001, 0.2038, 0.0064, 0.0054, 0.0051, 0.0099]::3
[0.0001, 0.0000, 0.0001, 0.0008, 0.0056, 0.0001, 0.0000, 0.1023, 0.0061, 0.6481]::9
[0.0001, 0.0000, 0.9795, 0.0185, 0.0305, 0.0000, 0.0010, 0.0003, 0.0001, 0.0873]::2
[0.0161, 0.0018, 0.0000, 0.0052, 0.0000, 0.0058, 0.0000, 0.9935, 0.0004, 0.0060]::7
[0.0040, 0.9984, 0.0287, 0.0160, 0.0410, 0.0005, 0.0075, 0.0049, 0.0087, 0.0022]::1
[0.0893, 0.1610, 0.0002, 0.2568, 0.0014, 0.8967, 0.0010, 0.0109, 0.0177, 0.0115]::5
[0.0028, 0.0003, 0.0004, 0.2713, 0.0008, 0.8692, 0.0105, 0.0001, 0.0055, 0.0221]::5
[0.0014, 0.0012, 0.0216, 0.0056, 0.0007, 0.0193, 0.0005, 0.0007, 0.8914, 0.0003]::8
[0.0061, 0.9460, 0.0034, 0.0133, 0.0300, 0.0054, 0.0115, 0.0127, 0.0174, 0.0053]::1
[0.0028, 0.0010, 0.0003, 0.9993, 0.0000, 0.0172, 0.0000, 0.0008, 0.0090, 0.0004]::3
[0.9639, 0.0001, 0.0032, 0.0000, 0.0049, 0.1168, 0.0001, 0.0000, 0.0016, 0.0001]::0
[0.0001, 0.0033, 0.0523, 0.1046, 0.0034, 0.6914, 0.0008, 0.0004, 0.0016, 0.0005]::3
[0.0004, 0.0003, 0.0002, 0.0000, 0.9466, 0.0015, 0.0010, 0.0205, 0.0126, 0.0938]::4
[0.0105, 0.0015, 0.0085, 0.0000, 0.0357, 0.0018, 0.0015, 0.0200, 0.0046, 0.5125]::9
[0.0166, 0.9857, 0.0070, 0.0126, 0.0316, 0.0035, 0.0089, 0.0130, 0.0227, 0.0047]::1
[0.8887, 0.0017, 0.0005, 0.0003, 0.0023, 0.0275, 0.0003, 0.0053, 0.0046, 0.0181]::0
[0.0037, 0.0614, 0.0001, 0.0011, 0.0011, 0.0364, 0.0002, 0.9784, 0.0013, 0.0498]::7
[0.0002, 0.0008, 0.0013, 0.0000, 0.0028, 0.0069, 0.5261, 0.0000, 0.0002, 0.0022]::6
[0.0001, 0.0087, 0.0025, 0.0002, 0.0661, 0.0011, 0.0017, 0.4763, 0.0150, 0.0057]::7
[0.0144, 0.0000, 0.0007, 0.0091, 0.0000, 0.2172, 0.0046, 0.0038, 0.0680, 0.0001]::5
[0.0004, 0.0013, 0.0149, 0.0000, 0.0067, 0.0162, 0.9983, 0.0026, 0.0000, 0.0034]::6
[0.0022, 0.3633, 0.0997, 0.0494, 0.0009, 0.0017, 0.0013, 0.0129, 0.1099, 0.0021]::2
[0.0019, 0.0000, 0.0000, 0.0000, 0.0090, 0.0001, 0.0003, 0.0263, 0.0016, 0.2663]::9
[0.0003, 0.0003, 0.0000, 0.0000, 0.0885, 0.0004, 0.0001, 0.0871, 0.0227, 0.7294]::9
[0.9997, 0.0092, 0.0007, 0.0000, 0.0075, 0.0057, 0.0000, 0.0001, 0.1147, 0.0031]::0
[0.8477, 0.0006, 0.0023, 0.0000, 0.0007, 0.0003, 0.0131, 0.0013, 0.0001, 0.0027]::0
[0.0299, 0.0032, 0.0000, 0.0001, 0.0001, 0.0130, 0.0000, 0.9598, 0.0039, 0.0085]::7
[0.0004, 0.0000, 0.9995, 0.0233, 0.0001, 0.0000, 0.0000, 0.0000, 0.0009, 0.0000]::2
[0.0008, 0.0116, 0.0121, 0.6877, 0.0000, 0.2762, 0.0910, 0.0002, 0.0004, 0.0003]::3
[0.0000, 0.6402, 0.0112, 0.1691, 0.0003, 0.0108, 0.4377, 0.0317, 0.0018, 0.0021]::1