08_2_模型加载与保存&CIFAR10自定义网络实战
模型保存与加载
Outline
- save/load weights
- 最干净、最轻量级，只保存参数，适用于自己有源代码的情况
- save/load entire model
- 把所有的状态都保存起来
- saved_model
- 模型的保存格式，和pytorch的ONNX对应，适用于交给工厂的生产环境时，直接部署，格式通用，python写的模型可以应用C++工作
1 save/load weights
例子：对于三层网络，只需要掌握[w1,b1,w2,b2,w3,b3]，每次都会更新这6个Variable，保存进文本里去，使用model.save_weights(保存路径)，保存到文件里去
例子：
保存到weights.ckpt
删除network
创建network，使用Sequential都是随机的w和b参数
实战
import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics
def preprocess(x, y):
"""
x is a simple image, not a batch
"""
x = tf.cast(x, dtype=tf.float32) / 255.
x = tf.reshape(x, [28*28])
y = tf.cast(y, dtype=tf.int32)
y = tf.one_hot(y, depth=10)
return x,y
batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())
db = tf.data.Dataset.from_tensor_slices((x,y))
db = db.map(preprocess).shuffle(60000).batch(batchsz)
ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz)
sample = next(iter(db))
print(sample[0].shape, sample[1].shape)
network = Sequential([layers.Dense(256, activation='relu'),
layers.Dense(128, activation='relu'),
layers.Dense(64, activation='relu'),
layers.Dense(32, activation='relu'),
layers.Dense(10)])
network.build(input_shape=(None, 28*28))
network.summary()
#配置训练方法
network.compile(optimizer=optimizers.Adam(lr=0.01),
loss=tf.losses.CategoricalCrossentropy(from_logits=True),
metrics=['accuracy']
)
#训练模型，会先training 3个epoch
network.fit(db, epochs=3, validation_data=ds_val, validation_freq=2)
network.evaluate(ds_val)
####保存模型
network.save_weights('weights.ckpt')
print('saved weights.')
del network
#网络需要与之前的网络一模一样
network = Sequential([layers.Dense(256, activation='relu'),
layers.Dense(128, activation='relu'),
layers.Dense(64, activation='relu'),
layers.Dense(32, activation='relu'),
layers.Dense(10)])
network.compile(optimizer=optimizers.Adam(lr=0.01),
loss=tf.losses.CategoricalCrossentropy(from_logits=True),
metrics=['accuracy']
)
####使用过去的模型
network.load_weights('weights.ckpt')
print('loaded weights!')
network.evaluate(ds_val)
保存之前是0.9666，加载weight后为0.9601。
不是完全一样的原因：因为神经网络不只是有w、b的参数，还有一些额外的影响因子在，会存在小的差别。如果想完全还原出差别，则需要保存整个模型。
2 保存整个模型
效率会低一些
network.save('model.h5')
恢复：network = tf.keras.models.load_model('model.h5')
实战
import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics
def preprocess(x, y):
"""
x is a simple image, not a batch
"""
x = tf.cast(x, dtype=tf.float32) / 255.
x = tf.reshape(x, [28*28])
y = tf.cast(y, dtype=tf.int32)
y = tf.one_hot(y, depth=10)
return x,y
batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())
db = tf.data.Dataset.from_tensor_slices((x,y))
db = db.map(preprocess).shuffle(60000).batch(batchsz)
ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz)
sample = next(iter(db))
print(sample[0].shape, sample[1].shape)
network = Sequential([layers.Dense(256, activation='relu'),
layers.Dense(128, activation='relu'),
layers.Dense(64, activation='relu'),
layers.Dense(32, activation='relu'),
layers.Dense(10)])
network.build(input_shape=(None, 28*28))
network.summary()
network.compile(optimizer=optimizers.Adam(lr=0.01),
loss=tf.losses.CategoricalCrossentropy(from_logits=True),
metrics=['accuracy']
)
network.fit(db, epochs=3, validation_data=ds_val, validation_freq=2)
network.evaluate(ds_val)
####保存整个模型
network.save('model.h5')
print('saved total model.')
del network
print('loaded model from file.')
####使用整个模型
network = tf.keras.models.load_model('model.h5', compile=False)
network.compile(optimizer=optimizers.Adam(lr=0.01),
loss=tf.losses.CategoricalCrossentropy(from_logits=True),
metrics=['accuracy']
)
x_val = tf.cast(x_val, dtype=tf.float32) / 255.
x_val = tf.reshape(x_val, [-1, 28*28])
y_val = tf.cast(y_val, dtype=tf.int32)
y_val = tf.one_hot(y_val, depth=10)
ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val)).batch(128)
network.evaluate(ds_val)
3 导出模型
导出为标准的可以给其他语言使用的模型。
tf.saved_model.save(m,'/temp/saved_model/')
imported = tf.saved_model.load(path)
f = import.signatures["serving_default"]
#测试
print(f(x=tf.ones([1,28,28,3])))
Keras实战(CIFAR10自定义网络实战)
一共有10个大类。准确度最好80%，一般50%。
My Dense layer
自定义层。
下图是标准的Dense层：
这次刻意地将Bias(b)去掉。
实战
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
import tensorflow as tf
from tensorflow.keras import datasets,layers,optimizers,Sequential,metrics
from tensorflow import keras
def preprocess(x,y):
#优化1：将x范围从[0,255]变成[-1,1]
x = 2 * tf.cast(x,dtype=tf.float32) / 255 - 1
y = tf.cast(y,dtype=tf.int32)
return x,y
batchsz = 128
# [32,32,3]
#初始：测试集x[50000,32,32,3] y[50000,1,10] x:0~255
(x,y),(x_val,y_val) = datasets.cifar10.load_data()
#y的label是[50000,1,10]，因此需要消掉多余的维度【1】
#squeeze:将原始input中所有维度为1的那些维都删掉 [1, 2, 1, 3, 1, 1]=>[2, 3]
y = tf.squeeze(y)
y_val = tf.squeeze(y_val)
#one-hot可以在这里；也可以在preprocess；也可以在GradientTape里，皆可
y = tf.one_hot(y,depth=10) #[50k,10]
y_val = tf.one_hot(y_val,depth=10) #[10k,10]
#此处还没有preprocess，因此x[0,255]
print('datasets:',x.shape,y.shape,x_val.shape,y_val.shape,x.min(),x.max())
#datasets: (50000, 32, 32, 3) (50000, 10) (10000, 32, 32, 3) (10000, 10) 0 255
train_db = tf.data.Dataset.from_tensor_slices((x,y))
#预处理；shuffle打乱10000为buffer随便输；batch
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 = next(iter(train_db))
print('batch:',sample[0].shape,sample[1].shape)
#自定义层
class MyDense(layers.Layer):
#代替标准的 layers.Dense()
def __init__(self,inp_dim,outp_dim):
super(MyDense,self).__init__()
#w
self.kernel = self.add_variable('w',[inp_dim,outp_dim])
#b，在这里去掉
# self.bias = self.add_variable('b',[outp_dim])
def call(self,inputs,training=None):
x = inputs @self.kernel
return x
#自定义网络，5层
class MyNetwork(keras.Model):
def __init__(self):
super(MyNetwork,self).__init__()
#优化2：增加参数量。结果过拟合，后面修改回去了。
self.fc1 = MyDense(32*32*3,256)
self.fc2 = MyDense(256,128)
self.fc3 = MyDense(128,64)
self.fc4 = MyDense(64, 32)
self.fc5 = MyDense(32, 10)
#training表示训练模式还是测试模式
def call(self,inputs,training=None):
"""
:param inputs:[b,32,32,3]
:param training:
:return:
"""
#[b,32,32,3] => [b,32*32*3]，这里reshape意味着你需要接收[b,32,32,3]的Tensor
x = tf.reshape(inputs,[-1,32*32*3])
# [b,32*32*3] => [b,256]
x = self.fc1(x)
x = tf.nn.relu(x)
# [b,256] => [b,128]
x =self.fc2(x)
x = tf.nn.relu(x)
# [b,128] => [b,64]
x = self.fc3(x)
x = tf.nn.relu(x)
# [b,64] => [b,32]
x = self.fc4(x)
x = tf.nn.relu(x)
# [b,32] => [b,10]
x = self.fc5(x)
#返回logits
return x
network = MyNetwork()
#lr一般是1e-3，设置大一点可能可以训练快一点，设置le-2，实验证明还是不行设会le-3；因为是多分类因此使用CategoricalCrossentropy
network.compile(optimizer=optimizers.Adam(lr=1e-3),
loss=tf.losses.CategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
#5个epoch；validation_data为test_db；每次epoch都验证一次
#优化3：增加epoch
network.fit(train_db,epochs=15,validation_data=test_db,validation_freq=1)
#模型保存，优先采用权值保存
network.evaluate(test_db)
network.save_weights('87result/ckpt/weight.ckpt')
del network
print('saved to ckpt/weights.ckpt')
#使用权值
network = MyNetwork()
network.compile(optimizer=optimizers.Adam(lr=1e-3),
loss=tf.losses.CategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
network.load_weights('87result/ckpt/weights.ckpt')
print('loaded weights from file.')
network.evaluate(test_db)
结果：
loss: 0.6486 - accuracy: 0.7690 - val_loss: 1.8323 - val_accuracy: 0.5208 到后面overfitting(过拟合)了，因为没用卷积神经网络、Resnet。
5356
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
