1.0tensorflow的安装

1.1安装python

python下载 需要python3.x<=3.7
https://www.python.org/ftp/python/3.7.7/python-3.7.7-amd64.exe

安装时勾选Add Python 3.7 to PATH，把python添加到环境变量。

1.2安装tensorflow

打开命令行，执行
pip install tensorflow==2.1.0


pip 会安装tensorflow和一些其他的依赖

1.3安装vc++2015-2019redist…

tensorflow的另一个依赖(很多tensorflow安装失败的原因就是这个没安装)
https://support.microsoft.com/en-us/help/2977003/the-latest-supported-visual-c-downloads

1.4安装CUDA和CUDNN

cuda: https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exelocal
cudnn: https://developer.nvidia.com/rdp/cudnn-download(需要注册nvidia账号)
cudnn下载后是个压缩文件，要把他解压出来放在CUDA里，如下图


高版本CUDA缺失cudart64_101.dll，下载后放在C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\bin里
https://cn.dll-files.com/cudart64_101.dll.html

2.0CASIA实战

2.1CASIA数据集

可以从网上下载casia数据集，
这里以casia数据集为例，现实中可以使用自己需要的数据集。

2.2数据集的处理

建立data和test两个文件夹，把casia复制到里面
目录是这样的./data/000/000_0.bmp
data.py处理数据，其实就是遍历，匹配，删除

import os 
data = './data'
dirs = os.listdir(data) 
for dir in dirs:
    for file in os.listdir(data + '/' + dir):
        if file.endswith("4.bmp"):
            os.remove(data + '/' + dir + '/' + file)
test = './test'
tdirs = os.listdir(test)
for dir in tdirs:
    for file in os.listdir(test + '/' + dir):
        if file.endswith("0.bmp"):
            os.remove(test + '/' + dir + '/' + file)
        if file.endswith("1.bmp"):
            os.remove(test + '/' + dir + '/' + file)
        if file.endswith("2.bmp"):
            os.remove(test + '/' + dir + '/' + file)
        if file.endswith("3.bmp"):
            os.remove(test + '/' + dir + '/' + file)

2.3训练代码

casia.py

import os
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Conv2D, Flatten, Dropout, MaxPooling2D
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import numpy as np
/*我直接建立了个0000，1111，...这样的数组作为标签*/
#data标签
arr = []
for i in range(100):
    for j in range(4):
        arr.append(i)
arr = np.array(arr)
#test标签
tarr = []
for i in range(100):
    tarr.append(i)
tarr = np.array(tarr)

#训练集
pwd='./data'
dirs = os.listdir(pwd)
imgs = []

for dir in dirs:
    for file in os.listdir(pwd + '/' + dir):
        image = tf.io.read_file(pwd + '/' + dir + '/' + file)
        img = tf.image.decode_bmp(image,channels=3)
        imgs.append(img)
print("[*]训练集加载完毕")
print(imgs[0].shape)
#验证集(测试集)
tpwd='./test'
tdirs = os.listdir(tpwd)
timgs = []
for tdir in tdirs:
    for tfile in os.listdir(tpwd + '/' + tdir):
        timage = tf.io.read_file(tpwd + '/' + tdir + '/' + tfile)
        timg = tf.image.decode_bmp(timage,channels=3)
        timgs.append(timg)
print("[*]验证集加载完毕")
print(timgs[0].shape)
#神经网络模型
model = Sequential([
    Conv2D(16, (3,3), padding='same', activation='relu',input_shape=(480,640,3)),
    MaxPooling2D(),
    Conv2D(64, (3,3), padding='same', activation='relu'),
    MaxPooling2D(),
    Conv2D(128, (3,3), padding='same', activation='relu'),
    MaxPooling2D(),
    Flatten(),
    Dense(128, activation='relu'),
    Dense(100, activation='softmax'),
])
model.summary()//打印神经网络模型
#优化器
model.compile(optimizer=tf.keras.optimizers.Adam(),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
#训练
ds = tf.data.Dataset.from_tensor_slices((imgs,arr))
ds = ds.batch(16)
ds = ds.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
model.fit(ds,epochs=20)
tds = tf.data.Dataset.from_tensor_slices((timgs,tarr))
tds = ds.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
model.evaluate(tds, verbose=2)
#保存
tf.saved_model.save(model, "./tmp/")

2.4训练与验证

在命令行运行 python casia.py进行训练
predict.py

import os
import tensorflow as tf
import numpy as np
/*这里显卡内存不够了*/
from tensorflow.compat.v1 import ConfigProto
from tensorflow.compat.v1 import InteractiveSession

config = ConfigProto()
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)
/*显卡内存*/

model_path = './tmp' //加载模型
test_path = "./test/002/002_4.bmp"//这里就是个栗子
model = tf.keras.models.load_model(model_path, custom_objects=None, compile=True)

image = tf.io.read_file(test_path)
img = tf.image.decode_bmp(image,channels=3)
img = img[tf.newaxis, ...]
res = model.predict(
    img, batch_size=None, verbose=0, steps=None, callbacks=None, max_queue_size=10,
    workers=1, use_multiprocessing=False
)
pred = tf.argmax(res, axis=1)
print (pred[0])
print (res[0,pred[0]])