Improve tensorboard monitoring #514
Conversation
| summary_writer = get_tensorboard_writer() | ||
| if summary_writer: | ||
| with summary_writer.as_default(step=get_step_number()): | ||
| if tf.rank(query_points) == 2: |
There was a problem hiding this comment.
why wouldnt this be true?
There was a problem hiding this comment.
Because the type system doesn't tell me so :-)
Seriously: does it have to be true (i.e. do we assume/assert it anywhere) or is just always true in practice?
| f"{tag}.observation.best_overall", | ||
| np.min(datasets[tag].observations), | ||
| step=step, | ||
| tf.summary.histogram( |
There was a problem hiding this comment.
histogram makes sense only for batches, no?
There was a problem hiding this comment.
I mean it still works for batch size one (though it doesn't provide any more information that the best_in_batch scalar plot).
There was a problem hiding this comment.
looks ok to me
what I found needed was recording some info about the run, like size of the data at the beginning (i.e. num initial points), budget aka number of steps, and information about the model and acq rule - this can all be filed as text, can serve as a useful metadata when comparing several runs
good suggestion, though i'd rather add text summaries in a separate pr |
Some initial improvements, based on suggestions at #401: