I’m running into an issue where I’m using the information in chapter 2 of the book to create an image recognition model and when I get to the point where I’m running:
cleaner = ImageClassifierCleaner(learn)
cleaner
I get a new cell with a single image and a large scroll bar. If I scroll to the right on the bar other images will flicker in and out but I can never select them for cleaning. I am running things locally. I’ve tried this in Brave, Chrome, and Edge. I’ve tried using the fastai fast setup with mamba (mambo?), as well as using my own install of things using a virtualenv and pip installing things. I get the same results regardless of using jupyter-lab or jupyter notebook. I’ve tried this both under Windows as well as WSL (I much prefer WSL though since wow it’s so much faster).
The code works as expected on the Kaggle workspace but I have a 4090 on my local machine and training is 5X faster locally than in the cloud so would prefer to be able to use my own machine for it. I can still do the course materials so it’s not a crisis but I would like to be able to clean my data if possible for the model.
I’m using Python 3.10.12. I’ve got fastai 2.7.12, jupyterlab 4.0.3, and ipwidgets 8.0.7. I’ve got the latest NPM and Node installed as of two days ago. I’m not seeing any errors and interp.plot_top_losses(5, nrows=1) is showing 5 images correctly.
Here is a picture of what it looks like (I can’t show the flickering as it shows while I’m moving the bar but they hide again when I’m done):
Great suggestion! On Kaggle it’s using fastai 2.7.12, ipywidgets 7.7.1, and jupyterlab 3.6.5. Sadly, if I install all of these (using a new virtualenv for fresh everything) once I get to the point where I run the training:
to at least see what’s going on and try and find bad images so I can at least do some clean up and then retrain once things get removed. Was just hoping there was something easy I might be missing that I just didn’t know yet.
Hey Jeremy, I was having the same problem but I managed to fix it by tweaking the jupyter packages version using pip to match the Google Collab version. My local machine is running on Fedora 38. The following is my requirements.txt.
Thanks for that! Sadly, when I try with those I just get Kernel restarted errors with nothing in the console saying why when I try and run learn.fine_tune(10). For now I’ll just have to run without this feature I guess. Hopefully there will be some future updates in the packages that fix things!
Unfortunately I did not. I know that some libraries take longer to update and I’m just hoping that eventually people will run into the same issue and work on the underlying stuff to make it work and update things and it will just start working.