search

rwth-i6 / pytorch-to-returnn

Make PyTorch code runnable within RETURNN
3 stars 6 forks source link

Reduced printing of numpy arrays #16

Closed vieting closed 3 years ago

vieting commented 3 years ago

It's maybe personal preference, but I don't like if all values for all numpy arrays are always printed, especially when a lot of ConstantLayers are used. I therefore added a threshold and if a numpy array contains more values, only an abbreviated string is printed.

albertz commented 3 years ago

I'm not sure on this. This pprint is not intended for debugging output (although also useful for that), but for printing the RETURNN net dict such that you can copy & paste it into your RETURNN config and it will directly work.

You are right that this deviates from the original __repr__ or __str__ of a Numpy array (and the original pprint also just uses the original __repr__ and has no extra logic for Numpy arrays).

But there is already the second problem with your approach, that this is inconsistent as well to the original ndarray.__repr__. Numpy has options for this, specifically e.g. via numpy.set_printoptions.

I wonder also, you changed a print to pprint here. But the original print already used the original Numpy mechanism, which should also have done that?

These prints (all via the converter) should anyway be cleaned up. We might use the Python logging mechanism. This was/is mostly for debugging. You should be able to control that with a verbosity setting.

vieting commented 3 years ago

I tried numpy.set_printoptions but still found that quite verbose. But yes, for my case now the easiest thing might be to reduce the verbosity level.