Fixed bug in _target_modality_is_real

[souramoo]

I was training a transformer model with a real-value output. I had my problems file set up properly, but when using .infer to load a trained model checkpoint and use it to run inference on some input, I was getting very obscure errors to do with tensor sizes. After some debugging it became apparent that the tensor2tensor framework was trying to look for a class to output, and did not accept that I wanted a real-valued output.

Turns out this is due to a small typo in t2t_model - when checking if the output modality is real, it decided it was not and this caused the decoder part of the library to use fast_decode rather than the slow version that seems suitable for real valued predictions. It checks if the output modality name begins with "Real_" to determine this.

Unfortunately, the name of these output modalities are things like real_l2_loss_modality with a lowercase r...

Hope this helps other people with the same issue!

[afrozenator]

Wow! Thanks a lot @MOOSD ! Will merge in shortly!

[ashu5644]

Hi, @MOOSD, I have trained a language model based on transformer using t2t. I am trying to do inference using model.infer() method by using tf.session() and feed_dict method in tensorflow but I am getting random outputs each time I run model.infer(). Can you suggest how did you do inference using .infer method instead of using estimator.predict() ?