I'm saving multiple checkpoints when training model and saving final model as well.
When evaluating I do model.eval() along with torch.no_grad().
This is where is gets weird:
If I evaluate all my checkpoints back to back, the loss is different Vs just evaluating just one model.
@singhay I think this would be a problem with the script and not the library. You should just console out log the path for the desired checkpoint, and reconcile that with the single eval
Hi,
I'm saving multiple checkpoints when training model and saving final model as well.
When evaluating I do
model.eval()
along withtorch.no_grad()
. This is where is gets weird: If I evaluate all my checkpoints back to back, the loss is different Vs just evaluating just one model.Here's how to run the script
Just eval: remove
--do_train
flag To only run single eval remove--eval_all_checkpoints
Single eval
Multiple eval: where all checkpoints are evaluated and the last one should be equal to above but it's not
Here's my toy dataset of two records
source gist