Open andnp opened 1 week ago
Tagging
One can tag on experiment name, part name, or part version. How do you propose we tackle this? The easiest use case would be tagging an entire experiment after a major deadline with most recent version of all experiment parts.
Suppose atari-online
is an experiment made of two parts:
Part Name Versions
QRC V0, V1, V2, V3
DQN V0, V1
Simply tagging the atari-online
experiment should tag the latest table version
exp tag atari-online "icml deadline"
How do you propose we save a record of tags? Is this something to be stored in metadata
database? I think we want each tag to be associated with at most one version of a part. With this restriction, tags and their corresponding experiment parts can be stored in something like the following:
Tags -> List of parts
aamas-deadline -> [dqn-v0]
icml-deadline -> [qrc-v3, dqn-v1]
Tagged experiments should be easily recoverable for rerun or analysis of results data.
I agree with proposal for resetting and deleting results. We just need to discuss how to incorporate experiment-name
, experiment-part
and version numbers, and come up with reasonable default behaviour for not providing any of them.
The preferred way to build an experiment definition is through the python scripting interface, e.g.:
However, there are several features that are far more convenient to interact with via cmdline.
Tagging
Very useful for tracking deadlines or other major events.
Resetting results
Currently, we will reuse results as much as possible (which is a good thing). However, sometimes we find bugs in the code that invalidate prior results. It isn't currently easy to selectively delete these. It also isn't currently possible to not reuse them. This puts the user in a bad position.
Expected behavior: this would create a new version,
v3
, that has the exact same configurations asv2
except all new configuration ids.Deleting results
For instance, in the case of a bug where results are fully meaningless or in case the results db becomes too large.