jeudi 28 avril 2016

Tensorflow LSTM model testing

I'm new to LSTM and Tensorflow, and I'm trying to use an LSTM model to learn and then classify some huge data set that I have. (I'm not worried about the accuracy my intention is to learn). I tried to implement the model in a similar way as in the PTB word prediction tutorial that uses LSTM. The code in the tutorial (http://ift.tt/1SUgZSz) uses the below line to run the session using the model

 cost, state, _ = session.run([m.cost, m.final_state, eval_op],
                                 {m.input_data: x,
                                  m.targets: y,
                                  m.initial_state: state})

I modified this for my example as below (to get the logits and work with it):

  cost, state, _,output,logits = session.run([m.cost, m.final_state, eval_op, m.output,m.logits],
                                 {m.input_data: x,
                                  m.targets: y,
                                  m.initial_state: state})

So my questions if someone could help are as below:

  • How can the model built while training be used for testing? What exactly is happening when 3 models are being used by the tutorial one for each test, train and validation?
  • What about the targets while testing(if I don't know them, say in a classification problem). What changes in the run_epoch () can be done in a way to use the model built during training.
  • Just another question: It's difficult to debug tensorflow graphs ( and I found it difficult to understand the tensorboard visualizer too) And I didn't find good resource for learning tensorflow (the website seems to be lacking structure/ documentation) What other resources/ debugging methods are there?

Thanks.

Aucun commentaire:

Enregistrer un commentaire