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
I want to evaluate_network() on two shuffles that have a different number of snapshots in their train folders. The differing number of snapshots saved is a result of exceptions or time-outs while training on colab. Both shuffles have a particular snapshot in common that I would like to use to compare the two. It is not currently intuitive to do this.
I could run evaluate_network() for each shuffle individually specifying the snapshotindex in the config file but doing this is not especially intuitive as one has to determine what the snapshotindex actually is. I wasn't able to find an explanation in the DLC documentation although from what I can see in the repo it can be determined by listing the snapshots in the order of increasing iterations and determining the index corresponding the desired snapshot. I believe there is a simpler alternative to this.
A parameter explicitly specifying the exact snapshot number (eg. 'snapshot-120000') in the evaluate_network() function would simplify this greatly.
Alternatively adding documentation clarifying how the snapshotindex is defined (beyond the values of '-1' and 'all') could be helpful.
The text was updated successfully, but these errors were encountered:
If this feature is of any interest to DLC I have a branch in my fork that enables passing a list of snapshot names using a snapshots_to_evaluate parameter to evaluate_network() and override the snapshotindex parameter from the cfg. The default functionality still remains if the parameter is not used. Let me know!
I find it helpful to be able to pass specific, named snapshots to the evaluate_network rather than "all" or specific indices. Snapshot indices are not necessarily constant in which snapshot they refer to since the snapshot files can be created and deleted.
Hi @Tetra-quark, I like that feature a lot (especially since snapshots are sorted by the number of iterations rather than timestamps, which is sometimes problematic when finetuning). I'd encourage you to open a PR and document the feature, so we can review it. 😊
I want to
evaluate_network()
on two shuffles that have a different number of snapshots in their train folders. The differing number of snapshots saved is a result of exceptions or time-outs while training on colab. Both shuffles have a particular snapshot in common that I would like to use to compare the two. It is not currently intuitive to do this.I could run
evaluate_network()
for each shuffle individually specifying thesnapshotindex
in the config file but doing this is not especially intuitive as one has to determine what thesnapshotindex
actually is. I wasn't able to find an explanation in the DLC documentation although from what I can see in the repo it can be determined by listing the snapshots in the order of increasing iterations and determining the index corresponding the desired snapshot. I believe there is a simpler alternative to this.A parameter explicitly specifying the exact snapshot number (eg. 'snapshot-120000') in the
evaluate_network()
function would simplify this greatly.Alternatively adding documentation clarifying how the snapshotindex is defined (beyond the values of '-1' and 'all') could be helpful.
The text was updated successfully, but these errors were encountered: