t2t notebook attention visualization broken

[tomsbergmanis]

Description

When running https://colab.research.google.com/github/tensorflow/tensor2tensor/blob/master/tensor2tensor/notebooks/hello_t2t.ipynb
code for en2de translation and visualizing all attentions heads input-output view attentions are nonsensical: for each output word all 8 attention heads to the first input word "The". Clearly, that is not right.
The prerendered visualization shows a very different attentions that do not match those produced by code.

I got this issue by simply running the notebook.

[martinpopel]

Can you try installing an older version of T2T in Colab (e.g. pip install tensor2tensor==1.4.0) and check if there is the same problem?
I have never seen this kind of problem, but I remember the visualization code was always buggy: for some sentences there were attention positions with no tokens shown or the other way round. I was just lazy to inspect it properly and report an issue.

Note that the enc-dec (input-output) attention may not correspond to what we would expect as alignment (it may be e.g. "shifted by one", cf. #898) - thus it is difficult to distinguish these unexpected attentions from visualization errors.

[tomsbergmanis]

Thank you for your answer.

After installing pip install tensor2tensor==1.4.0 code fails to import trainer_lib:
ImportErrorTraceback (most recent call last)
in ()
8 from tensor2tensor import problems
9 from tensor2tensor.layers import common_layers
---> 10 from tensor2tensor.utils import trainer_lib
11 from tensor2tensor.utils import t2t_model
12 from tensor2tensor.utils import registry

ImportError: cannot import name trainer_lib

[tomsbergmanis]

The oldest version code actually runs with is tensor2tensor==1.6.0, but the buggy attentions are still the same.
Any other suggestions?

[repoloper]

+1

Not sure if related, but one problem I am seeing is that attention is not using the decode_hparams somehow and the translations are different when I use t2t-decoder vs. the one in the attention visualization.

[drscotthawley]

I don't even get any display of anything. No matter what I select for the dropdowns, there are no layers to select and no output from this portion of the notebook.