1.0.45 no longer exists


I was using fast.ai 1.0.45 locally and to deploy a webservice on Azure, but a few days ago, that version disappeared. This caused the new deployment I was trying to make to fail.

Is there a github page I should follow that announces in advance when versions get deleted?

This kind of disappearance tends to have a negative impact, especially when one needs to retrain a model all over again, with another version, to allow for the deployment to work.

Thank you for any information you can share.

Looking at git it doesn’t appear that they normally delete versions. Perhaps it was a mix-up. You can always check out the 1.0.45.dev0 branch yourself and create your own local branch. There appears to be only one commit after sgugger created 1.0.45.dev0 so you either can check out the commit without those changes or check out the one with those changes and see if it works for you.

initial 1.0.45.dev0

git checkout 7b3413c7e3fba77f263528dd197641fce9439352

1.0.45.dev0 with commit

git checkout 6b7883c71e09ecfedee00fc559351649c9454546

Found this which seems to describe what happened to 1.045

sorry about that, I deleted 1.0.45, thinking that it wasn’t released, but it was 1.0.44 that was skipped, but it’s exactly the same version and pypi doesn’t allow re-uploading files :frowning: Will try to be more careful in the future.

It’s still on conda just moved it into the test channel to avoid confusion. https://anaconda.org/fastai/repo/files?type=any&label=test

Thank you, Steven and Stas, for your helpful and clarifying replies.

For now, I am using 1.0.44 because 1.0.46 has changes that no longer work for me (ImageItemList --> ItemList and loading data with less than 64 images now breaks on Windows, even when I use a batch size = dataset size).

Stas, I appreciate you pointing me to the test channel, and implementing a safer procedure for new version releases.

1 Like

Just wondering, did you try ImageList. I think from 1.0.45 or 46, ImageItemList is now called ImageList

You can take a look of the datablock API definition


Hello Hao,
I indeed tried ImageList, but because my dataset was smaller than the default batch size of 64, it failed. It even failed when I forced the batch size to be = the size of my dataset.