Neural Network学习9 CIFAR10实例以及模型的保存与加载

import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics
import datetime
from matplotlib import pyplot as plt
import io


# datasets用于数据集的管理，layers用于dense层，optimizer优化器，Sequential容器，metrics测试用的度量器


# 对数据集作预处理(对每一个x，y样本)
def preprocess(x, y):
    # [0,255]==>[-1,1]
    x = 2 * tf.cast(x, dtype=tf.float32) / 255. - 1. #有时候0-1不是最好的，这里调到-1~1
    y = tf.cast(y, dtype=tf.int32)
    return x, y


batchsz = 128
# 得到的是x[50k,32,32,3],y[10k,1]所以需要squeeze把那个1去掉
(x, y), (x_val, y_val) = datasets.cifar10.load_data()  # 加载cifar10数据
y = tf.squeeze(y)  # 消去多余的维度 [50k,1]==>[50k]
y_val = tf.squeeze(y_val)
y = tf.one_hot(y, depth=10)
y_val = tf.one_hot(y_val, depth=10)
print(x.shape, y.shape, x_val.shape, y_val.shape)

# 构建数据集
train_db = tf.data.Dataset.from_tensor_slices((x, y))
# map对每一个instance作预处理
train_db = train_db.map(preprocess).shuffle(10000).batch(batchsz)
test_db = tf.data.Dataset.from_tensor_slices((x_val, y_val))
test_db = test_db.map(preprocess).batch(batchsz)

# sample[0]的shape：[128,32,32,3];sample[1]的shape[128,10]
sample = next(iter(train_db))


# 创建自己的一个layer,要继承自layers.Layer,因为要使用标准的keras的接口
class MyDense(layers.Layer):
    # 继承两个函数，代替标准layer.Dense()
    def __init__(self, inp_dim, outp_dim):
        # 调用母类初始化
        super(MyDense, self).__init__()
        self.kernel = self.add_weight('w', [inp_dim, outp_dim])  # add_weight替换原来的add_variable
        # 自定义的层里不用bias
        # self.bias=self.add_variable('b',[outp_dim])

    def call(self, inputs, training=None):
        out = inputs @ self.kernel
        return out


# 自定义网络
class MyModel(keras.Model):
    def __init__(self):
        super(MyModel, self).__init__()
        self.fc1 = MyDense(32 * 32 * 3, 256)
        self.fc2 = MyDense(256, 128)  #都变成256，使得网络变大
        self.fc3 = MyDense(128, 64)
        self.fc4 = MyDense(64, 32)
        self.fc5 = MyDense(32, 10)

    # training 是用来标注现在是training模式还是test模式
    def call(self, inputs, training=None):
        """"
        :param inputs:[b,32,32,3]
        :param training
        :return
        """""
        # 先把x变成Dense可接受的形式
        x = tf.reshape(inputs, [-1, 32 * 32 * 3])
        x = self.fc1(x)  # [b,32*32*3]==>[b,256]
        x = tf.nn.relu(x)
        x = self.fc2(x)  # [b,256]==>[b,128]
        x = tf.nn.relu(x)
        x = self.fc3(x)  # [b,128]==>[b,64]
        x = tf.nn.relu(x)
        x = self.fc4(x)  # [b,64]==>[b,32]
        x = tf.nn.relu(x)
        x = self.fc5(x)  # [b,32]==>[b,10] logits 不加激活函数
        return x


model = MyModel()
# learning rate太大会可能出现梯度消失
model.compile(optimizers=optimizers.Adam(lr=1e-3),
              loss=tf.losses.CategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
# training 5个epoch，每隔一次循环测试一次
model.fit(train_db, epochs=15, validation_data=test_db, validation_freq=1)  # acc 0.4754

# 确定模型是否已经保存了
model.evaluate(test_db)
# 把权值保存到checkpoints
model.save_weights('ckpt/weights.ckpt')
del model
print('saved to ckpt/weights.ckpt')

# 因为只保存了权指，所以还是需要网络创建
model = MyModel()
# learning rate太大会可能出现梯度消失
model.compile(optimizers=optimizers.Adam(lr=1e-3),
              loss=tf.losses.CategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
model.load_weights('ckpt/weights.ckpt')
print('loaded weights from file ')
model.evaluate(test_db) #和前一个一样