Unstable training when saving zero-shot weights

[joshmyersdean]

Hello!

I am trying to reproduce the pascal part results when going from base to novel categories and am experiencing different results from saving my own weights vs. using the ones you have provided in datasets/metadata.

For example, when using the saved weights you provide these are my results after 2000 iterations (1 gpu).

And these are my results after saving new weights using tools/pascal_part_base_clip_name.py.

The only difference seems to be in the precision between these weights (in the 1e-4 place). Do you know why there would be such a discrepancy?

Thank you!
Josh