当变量形状改变时从检查点恢复 [英] Restoring from checkpoint when variable shape changes

问题描述

我无法恢复包含改变形状的变量的检查点模型.例如这个简单的模型:

I'm unable to restore checkpointed models that include variables that change shape. For example with this simple model:

var = tf.get_variable(initializer=tf.constant_initializer([0]), shape=[1], trainable=False, name='var')
op = tf.assign(var, [1, 2], validate_shape=False)
saver = tf.train.Saver(reshape=False)

如果我运行 op 然后保存模型，当我尝试恢复它时，我会收到以下错误:

if I run op and then save the model, when I try to restore it I get the following error:

Assign requires shapes of both tensors to match. lhs shape= [1] rhs shape= [2]
 [[Node: save/Assign = Assign[T=DT_FLOAT, use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/cpu:0"](var, save/restore_slice)]]

这似乎与不断变化的形状和 Saver 试图验证形状有关.如果我在构建 Saver 时将 reshape 设置为 True，根据文档应该可以解决这个问题，我反而得到这个错误:

which seems to have to do with the changing shape and Saver trying to validate shape. If I set reshape to True when constructing the Saver, which according to the docs should solve this problem, I instead get this error:

Input to reshape is a tensor with 2 values, but the requested shape has 1
 [[Node: save/Reshape = Reshape[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"](save/restore_slice, save/Reshape/shape)]]

我倾向于认为这是一个错误.

I'm inclined to think that this is a bug.

推荐答案

Saver 的 reshape 选项仅在形状具有相同的元素总数时才有效.例如，它可以让您从形状为 [] 的数据中加载形状为 [1] 的变量，或形状为 [15, 7]<的变量/code> 来自形状为 [5, 21] 的数据.如果形状以这种方式不兼容，则您必须构建一个新图.

The reshape option to Saver only works if the shapes have the same total number of elements. For example, it will let you load a variable with shape [1] from data with shape [], or a variable with shape [15, 7] from data with shape [5, 21]. If the shapes aren't compatible in this way, you'll have to build a new graph.

这篇关于当变量形状改变时从检查点恢复的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！