You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Given I have a set of (performance, config) pairs, I would like to use this to warm-start SMAC. There is no user friendly way to do this, the only way possible is to restart from the same directory from which it was started before which is not a possibility here. To do so would require having a valid file structure set up, a lot of overhead of figuring out and definitely not the intended way to warm-start from past runs that exist elsewhere.
While trying to find the model update part in smbo.py, I could not find anywhere that remotely tells me where the model update happens. It's rather obscure and happens here:
# For each intensification of the incumbent, we also want to intensify the next configuration
# We simply add it to the queue and intensify it in the next iteration
try:
config=next(self.config_generator)
config_hash=get_config_hash(config)
self._queue.append((config, 1))
logger.debug(f"--- Added a new config {config_hash} to the queue.")
Yup it's rather hard to spot but it happens on the call to config = next(self.config_generator). Essentially the update to the model happens in a rather obscure file called config_selector.py, which is not at all where one one expects the model to be updated.
I don't have any concrete proposals of how to fix this without trying it out, but essentially the proposal would be to lift up this logic to smbo.py somehow as this has a neat little comment that it describes the BO loop. This would make hacking around in SMAC a lot more manageable.
The text was updated successfully, but these errors were encountered:
Given I have a set of
(performance, config)
pairs, I would like to use this to warm-start SMAC. There is no user friendly way to do this, the only way possible is to restart from the same directory from which it was started before which is not a possibility here. To do so would require having a valid file structure set up, a lot of overhead of figuring out and definitely not the intended way to warm-start from past runs that exist elsewhere.While trying to find the model update part in
smbo.py
, I could not find anywhere that remotely tells me where the model update happens. It's rather obscure and happens here:SMAC3/smac/intensifier/intensifier.py
Lines 223 to 229 in e64e191
Yup it's rather hard to spot but it happens on the call to
config = next(self.config_generator)
. Essentially the update to the model happens in a rather obscure file calledconfig_selector.py
, which is not at all where one one expects the model to be updated.I don't have any concrete proposals of how to fix this without trying it out, but essentially the proposal would be to lift up this logic to
smbo.py
somehow as this has a neat little comment that it describes the BO loop. This would make hacking around in SMAC a lot more manageable.The text was updated successfully, but these errors were encountered: