If you ever get around to it, a nice add would be adding a section called: “What to do with git-lfs goes wrong”
I messed something up with adding files to git-lfs and was getting errors I couldn’t resolve around “missing” files. Ended up just cloning the HF Spaces repo and starting from scratch. Heck, that might be the answer, haha.
For the homework to read the book, I had to muddle around a bit to get 01_intro.ipynb running smoothly on kaggle.com. So I’m documenting here in case it helps anyone, and I might learn from others suggesting improvements.
What do you think about Jupyter Lab? Seems like we mostly use Jupyter Notebooks in the course. I was thinking that the former one is about to become an improved version of the latter. But I guess it didn’t become a ubiquitous approach.
Is there a some kind of auto-augmentation in fastai? (Sorry if this was mentioned somewhere already.) Last time I tried this approach in some other framework, it wasn’t easy to set up.
Update
I mean, something like a backprop guided augmentations derived from the data. But I guess it is not relevant for this lesson anyway, something to ask about in a different thread.
At what point do you add more data to the dataset? Ie if the model had trouble identifying multiple teddy bears, do you immediately add more images of that?
@Raymond-Wu This really depends on what you’re trying to accomplish. In general, adding more data helps if you have already chosen a good underlying neural net architecture to work on your problem.
Does anyone know what the best bang for buck GPUs are on GCP? I picked a T4 because its got lots of ram and the price is pretty similar to the P4. I was under the impression the older K80s aren’t that fast for the price. If i’m going to pay for one on the lower end what does everyone recommend?
Putting this question here from the chat: If you went looking for photos of grizzlys and black bears online (assuming there wasn’t a dataset already made and labelled), what is the best way to ensure these photos aren’t misclassified?.
Jeremy then showed the Image Classifier Cleaner, and Nick said it pays to visually inspect when using these “open” image searches. Results can deteriorate drastically with both the inherent ambiguity of your topic or your search query. Sheik Mohamed Imran said we would have to manually get the losses for the data and sort it.Or you van peek into the code used for the GUI, has the same logic
