Question

如何在下面显示的tensorflow 2.0中将所选变量保存在文件中，然后将其加载到其他代码中的某些已定义变量中？

class manyVariables:
    def __init__(self):
        self.initList = [None]*100
        for i in range(100):
            self.initList[i] = tf.Variable(tf.random.normal([5,5]))
        self.makeSomeMoreVariables()

    def makeSomeMoreVariables(self):
        self.moreList = [None]*10
        for i in range(10):
            self.moreList[i] = tf.Variable(tf.random.normal([3,3]))

    def saveVariables(self):
        # how to save self.initList's 3,55 and 60th elements and self.moreList's 4th element

Answer 1

我不确定这是否是您的意思，但是您可以为要保存和恢复的变量专门创建一个tf.train.Checkpoint对象。请参见以下示例：

import tensorflow as tf

class manyVariables:
    def __init__(self):
        self.initList = [None]*100
        for i in range(100):
            self.initList[i] = tf.Variable(tf.random.normal([5,5]))
        self.makeSomeMoreVariables()
        self.ckpt = self.makeCheckpoint()

    def makeSomeMoreVariables(self):
        self.moreList = [None]*10
        for i in range(10):
            self.moreList[i] = tf.Variable(tf.random.normal([3,3]))

    def makeCheckpoint(self):
        return tf.train.Checkpoint(
            init3=self.initList[3], init55=self.initList[55],
            init60=self.initList[60], more4=self.moreList[4])

    def saveVariables(self):
        self.ckpt.save('./ckpt')

    def restoreVariables(self):
        status = self.ckpt.restore(tf.train.latest_checkpoint('.'))
        status.assert_consumed()  # Optional check

# Create variables
v1 = manyVariables()
# Assigned fixed values
for i, v in enumerate(v1.initList):
    v.assign(i * tf.ones_like(v))
for i, v in enumerate(v1.moreList):
    v.assign(100 + i * tf.ones_like(v))
# Save them
v1.saveVariables()

# Create new variables
v2 = manyVariables()
# Check initial values
print(v2.initList[2].numpy())
# [[-1.9110833   0.05956204 -1.1753829  -0.3572553  -0.95049495]
#  [ 0.31409055  1.1262076   0.47890127 -0.1699607   0.4409122 ]
#  [-0.75385517 -0.13847834  0.97012395  0.42515194 -1.4371008 ]
#  [ 0.44205236  0.86158335  0.6919655  -2.5156968   0.16496429]
#  [-1.241602   -0.15177743  0.5603795  -0.3560254  -0.18536267]]
print(v2.initList[3].numpy())
# [[-3.3441594  -0.18425298 -0.4898144  -1.2330629   0.08798431]
#  [ 1.5002227   0.99475247  0.7817361   0.3849587  -0.59548247]
#  [-0.57121766 -1.277224    0.6957546  -0.67618763  0.0510064 ]
#  [ 0.85491985  0.13310803 -0.93152267  0.10205163  0.57520276]
#  [-1.0606447  -0.16966362 -1.0448577   0.56799036 -0.90726566]]

# Restore them
v2.restoreVariables()
# Check values after restoring
print(v2.initList[2].numpy())
# [[-1.9110833   0.05956204 -1.1753829  -0.3572553  -0.95049495]
#  [ 0.31409055  1.1262076   0.47890127 -0.1699607   0.4409122 ]
#  [-0.75385517 -0.13847834  0.97012395  0.42515194 -1.4371008 ]
#  [ 0.44205236  0.86158335  0.6919655  -2.5156968   0.16496429]
#  [-1.241602   -0.15177743  0.5603795  -0.3560254  -0.18536267]]
print(v2.initList[3].numpy())
# [[3. 3. 3. 3. 3.]
#  [3. 3. 3. 3. 3.]
#  [3. 3. 3. 3. 3.]
#  [3. 3. 3. 3. 3.]
#  [3. 3. 3. 3. 3.]]

如果要将所有变量保存在列表中，则可以将makeCheckpoint替换为以下内容：

def makeCheckpoint(self):
    return tf.train.Checkpoint(
        **{f'init{i}': v for i, v in enumerate(self.initList)},
        **{f'more{i}': v for i, v in enumerate(self.moreList)})

请注意，您可以具有“嵌套”检查点，因此，更一般而言，您可以具有为变量列表创建检查点的函数，例如：

def listCheckpoint(varList):
    # Use 'item{}'.format(i) if using Python <3.6
    return tf.train.Checkpoint(**{f'item{i}': v for i, v in enumerate(varList)})

那么你就可以拥有这个：

def makeCheckpoint(self):
    return tf.train.Checkpoint(init=listCheckpoint(self.initList),
                               more=listCheckpoint(self.moreList))

Answer 2

在下面的代码中，我将一个名为变量的数组保存到一个.txt文件中，该文件具有您选择的名称。该文件将与python文件位于同一文件夹中。打开函数中的“ wb”表示以截断形式写入（因此删除文件中以前的所有内容）并使用字节格式。我用pickle处理保存/解析列表。

import pickle

    def saveVariables(self, variables): #where 'variables' is a list of variables
        with open("nameOfYourFile.txt", 'wb+') as file:
           pickle.dump(variables, file)

    def retrieveVariables(self, filename):
        variables = []
        with open(str(filename), 'rb') as file:
            variables = pickle.load(file)
        return variables

要将特定内容保存到文件中，只需将其作为saveVariables中的variables参数添加即可，

myVariables = [initList[2], initList[54], initList[59], moreList[3]]
saveVariables(myVariables)

要从具有特定名称的文本文件中检索变量：

myVariables = retrieveVariables("theNameOfYourFile.txt")
thirdEl = myVariables[0]
fiftyFifthEl = myVariables[1]
SixtiethEl = myVariables[2]
fourthEl = myVariables[3]

您可以在类中的任何位置添加这些功能。

但是，为了能够访问示例中的initList / moreList，您应该从它们的函数中返回它们（就像我对variables列表所做的那样）或将它们设为全局。

如何在TensorFlow 2.0中保存和加载所选变量？

2 个答案: