I have a model that was trained on an older version of fastai, but with torch 0.4 and cuda 8.0x.
I’ve set up another machine that has the standard fastai library and setup from the paperspace script. For some reason, I am getting completely different results.
I’m running:
the same model,
with the same data
from the same file
The only difference is that it’s a different computer. I’m sure there are differences in the underlying libraries - fastai is a different version (although I did try reverting the git repo) - etc. But I can’t understand why I’m getting completely different results. Can anyone help?
Kind regards,
Steve
edit - the previous revision of fast ai was e8841f79eb399ac641dac1fe3ba05fb5f6b8b93f
I am guessing the filesystems might be different. Have a look at data.test_ds.fnames (or data.trn_ds.fnames or data.val_ds.fnames) in both the computers and check if they are the same. It mainly differs when using ImageClassiferData.from_paths.
Though it would be better to clarify what “completely different results” mean.
Thanks very much - that fixed it - or rather identified it. Now, at least, I know what was wrong.
For the record, completely different meant that results on one system were completely uncorrelated with results on the other. It was mystifying. I would have expected it to fail completely instead of being wrong.