python - 如何使用 TensorFlow 在 Returnn 中加载经过训练的网络的权重

Question

当尝试加载经过训练的多个时期保存的权重时使用以下代码返回网络:

import tensorflow as tf
from returnn.Config import Config
from returnn.TFNetwork import TFNetwork

for i in range(1,11):
    modelFilePath = path/to/model/ + 'network.' + '%03d' % (i,)

    returnnConfig = Config()
    returnnConfig.load_file(path/to/configFile)
    returnnTfNetwork = TFNetwork(config=path/to/configFile, train_flag=False, eval_flag=True)

    returnnTfNetwork.construct_from_dict(returnnConfig.typed_value('network'))

    with tf.Session() as sess:
        returnnTfNetwork.load_params_from_file(modelFilePath, sess)

我收到以下错误:

Variables to restore which are not in checkpoint:
global_step_1

Variables in checkpoint which are not needed for restore:
global_step

Probably we can restore these:
(None)

Error, some entry is missing in the checkpoint

Answer 1

问题在于，您每次在循环中都重新创建 TFNetwork，并且每次都会为全局步骤创建一个新变量，该变量必须以不同的方式调用，因为每个变量都必须有一个唯一的名称。

你可以在循环内做这样的事情:

tf.reset_default_graph()