Thomas Gamauf
Sep 7, 2018 · 1 min read

So you are training on whole sequences, which in turn consist of an number of channels. For a setup like this a SequenceExample “as channels” probably is the right choice. You would create a Tensorflow iterator on your TFRecord Dataset, which will then deliver whole sequences as described in your last comment to your model. Batching, shuffling, and so on can be done automatically by Dataset methods as described here and here.

You hand the iterator over to fit function (e.g. Keras fit — I haven’t worked with pure TF models up to now, so I can’t comment on it) and the fit function then iterates over the data the iterator provides until the model is trained.

What I don’t see yet is how the class probabilities fit into this. Do you use them as targets/labels?

    Thomas Gamauf

    Written by

    Ideas are like rabbits. You get a couple, learn how to handle them, and pretty soon you have a dozen. – John Steinbeck