Wow, it’s been a while since I’ve had some progress to report.
I’ve had some extended-family medical stuff going on (happily all resolved for now) that slowed my studies down for a bit.
I also had some minor technical challenges to work through, the kind which honestly sometimes felt like a waste of time, but hey, it led to some learning too.
Package managers and environments
In the second fastai video (2022 course), Jeremy runs through a quick and easy local installation of python and fastai. Because of what I already had set up, this caused me confusion on several points.
I wanted to follow what was shown and install Mambaforge. But I already had Anaconda installed, I installed on top of that, and my conda environments stopped working properly.
I tried some half measures to fix things, but I ended up unistalling Anaconda and Mamba, and installing Mambaforge to get things working again.
Also, it was advised in the video to install various packages in the base environment. This was counter to how I had used conda environments before, and both the conda and mamba docs advise against it, so I wasn’t sure what to do with that.
I asked about it in the fastai forum for Lesson 2, and the response from several people was basically, “aaah, don’t worry about it, we do it all the time.” So it’s probably fine, but I’ll stick to creating new environments rather than installing everything into the base environment.
Another thing that sent me digging for more information: the mamba docs advised to use conda activate to activate an environment while mamba itself, after creating an environment, tells you to use mambba activate. What the what??
After a bit of hunting, I found this exhange which explains that mamba activate is implemented for most shells and is fine to use, and that the documentation pre-dates this and needs to be updated.
It also threw me off that the video has you installing fastai on Windows, Linux, or Mac, while the fastai docs state that it’s not supported for Mac. I think this comes down to a lack of GPU support, but I don’t really know.
I gathered from the video that it should still run on Mac and be ok for non-GPU-intensive tasks, like using an already-trained model to classify a few examples. I haven’t tried it yet, and I may just stick to running fastai on kaggle.com, in a Colab notebook, or on Gradient Paperspace for now.