log a few images to tensorboard

[notiho]

No description provided.

[bencomp]

Thanks for making my comment in #482 into this PR. The code looks even more understandable than I expected :) (I haven't tested it myself, but I assume it works.)

I wonder if it makes sense to log the number of the sample in the train_set, i.e. the result of np.random.randint instead of or next to the i, so that it can be looked up more easily? If the train_set is shuffled or if the dataset is an Arrow file, then this number would not map to an image that you can look up easily.

Perhaps the logging of images could be made optional too, by adding a CLI option that sets the max number of logged images to 0 by default. You could refer to that number instead of 16.

[PonteIneptique]

Just an additional comment: this would be a wonderful thing to also log the prediction with it :)
But either, thank you for this PR !

[mittagessen]

I wonder if it makes sense to log the number of the sample in the train_set, i.e. the result of np.random.randint instead of or next to the i, so that it can be looked up more easily? If the train_set is shuffled or if the dataset is an Arrow file, then this number would not map to an image that you can look up easily.

Because the samples might be substituted on the fly if loading them fails for any reason the indices might not actually correspond to the order of the dataset. And as a note you can unpack a binary dataset with contrib/extract_lines.py now but if you're defining fixed splits in the file the indices in the extracted file names won't match the ones used during training even in the absence of errors.

[notiho]

I now also added logging for validation images and their prediction. How do you feel about adding another parameter to control the image logging?