I think the problem is that widgets are not supported in Google Colab.
As you are storing the images in Google Drive you could just go in there and delete some. However the widget is worth a go so I would recommend setting up a local Jupyter environment, connect to your Google Drive and then run the widget. Once you have run the widget a csv is generated. I would recommend then going back to Google Colab to complete the lesson (as it is likely to be faster than your local environment).
I am using Colab,so went ahead and removed the noisy images from the directories. Now it started erroring in the plot top losses and other functions. So I’m having to go all the way back to the beginning and create a random seed + data bunch.
Is there another way to update the data bunch? Could I make that call later?
Unfortunately it has been a while since I looked at this so I cannot recommend a different way to update the data bunch. The way I actually did it at the time was to set up a local development environment, then use the csv which is created to filter the data.
I have encountered this problem too, in fast.ai part 1 v3 lesson2-download. I am currently switching to Paperspace Gradient. As far as I can tell, it is Jupyter Notebook, not JupyterLab (which Google Colab runs on and is incompatible with widgets).
According to this page, JupyterLab is the successor to Notebooks. I have not checked out fastai v2 nor the part 1 v4, which is currently being filmed, so I cannot say how they are dealing with this incompatibility. Perhaps someone who works on the code can comment as to whether the library v2 and the lesson v4 migrates to Lab, or drops Colab support.