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
We currently document the number of parameters in the Weights meta-data and we display these on our summary tables. Other values might be of interest for users:
Model size in MB. This will usually be ~= 4 * num_params, except for quantized models. Either-way, it's worth computing and documenting directly (Note: Adding FLOPs and size to model metadata #6936 added docs for the file size, which may (or may not?) be different)
Memory requirements for 1 image (forward and backward pass). This can be estimated manually but can quickly become untractable, so it's best to do that automatically with a script. We should be able to use memory_stats module for this (accounting for the caching allocator, etc). I'll need to think about this more to make sure what we report is meaningful.
We currently document the number of parameters in the
Weight
s meta-data and we display these on our summary tables. Other values might be of interest for users:~= 4 * num_params
, except for quantized models. Either-way, it's worth computing and documenting directly (Note: Adding FLOPs and size to model metadata #6936 added docs for the file size, which may (or may not?) be different)memory_stats
module for this (accounting for the caching allocator, etc). I'll need to think about this more to make sure what we report is meaningful.CC @datumbox
The text was updated successfully, but these errors were encountered: