Austin

I agree the other posters. Loving the book so far, but an answer key would be reassuring. Thanks for all your work!

Hi Carlo, thanks for your reply. I've been learning a lot going through your library. Unfortunately, I'm not aware of a way to get a url of the full image...

> Hi Thanks for raising this issue. > > The import error is because the script is importing `architecture.py` and `architecture.py` is importing `block.py`. Cloupickle (package we use to serialize...

@coryMosaicML you know all the things about this?

I see what you're getting at and i believe this would alleviate having to explicitly set or wrap models to take initializers on init, but I'm not in love with...

As an aside, I do agree the conventional pytorch model defined initialization schemes aren't perfect. They are often not obvious (implicitly set as i mentioned) and they aren't flexible. Here's...

The interface for ComposerModel is left to the user, but I do like the idea of passing initializers to the model as python functions. Upside is that they're swappable again....

> Alternatively, we can have the recommended way be to have the user call `composer.utils.reproducibility`, and then remove these arguments from the trainer's `__init__`. However, I think this would make...

I'll do my best. TLDR: the progress bar should at least tell us total training time and model metrics at a glance. our current progress bar looks like this: There...

I'm not sure yet about the first two. Definitely going to be ongoing discussions. For example, with unet in particular it's probably better to move the preprocessing to the loader....