如何在张量流中从* .meta恢复训练？

Question

In the latest version of tensorflow, when I save the model I find two files are produced: model_xxx and model_xxx.meta.

Does model_xxx.meta specify the network? Can I resume training using model_xxx and model_xxx.meta without specify the network in the code? What about training queue structure, are they stored in model_xxx.meta?

Answer 1

Not sure if this will work for you, but at least for DNNCLassifiers you can specify the model_dir parameter when creating it and that will construct the model from the files and then you can continue the training.

For the DNNClassifiers you specify model_dir when first creating the object and the training will store checkpoints and other files on this directory. You can come then after that, and create another DNNClassifier specifying the same model_dir and that will restore your pre-trained model.