TensorFlow训练产生3个ckpt文件，哪个正确？ （Music_RNN_RBM）

Question

I'm giving a shot at Music_RNN_RBM

I've trained the model and, as instruction said a ckpt file is generated, problem is that in the directory for each epoch saved I've three .ckpt .data .index and .meta example:

model.ckpt-151800.data-00000-of-00001
model.ckpt-151800.index
model.ckpt-151800.meta

Now, next step would be to generate trough

python rnn_rbm_generate.py <path_to_ckpt_file>

I've tried to do like this for each type of file

python3 rnn_rbm_generate.py model.ckpt-151800.index

But I get this error each time

Traceback (most recent call last):
  File "rnn_rbm_generate.py", line 52, in <module>
    main(sys.argv[1])
  File "rnn_rbm_generate.py", line 43, in main
    saver.restore(sess, ckpt.model_checkpoint_path)
AttributeError: 'NoneType' object has no attribute 'model_checkpoint_path' 

I don't get what I should do

Answer 1

As you discovered, each checkpoint consists of a .data , .meta , and a index file.

However, TensorFlow also creates a checkpoints file, which keeps track of all the checkpoints in the output directory. It contains a link model_checkpoint_path to the most recent checkpoint which is what your model is looking for.

So to answer your question, the path you are providing should contain this checkpoints file.

Answer 2

There's a function to load the latest one . Use it to load it, and create a proper saved model, checkpoints are used as back-up, or to continue training afterwards, not to serve. The 'checkpoints' file prouast mentioned is loaded in by the function I mentioned, it's information about the files that are being created, and can't be loaded. Questions, ask away

EDIT: besides the code code in your repo indicates that you can also provide the full folder, instead of a file.