I guess it is more about using some digital notebook, i.e. OneNote or Evernote maybe. To easily dump the files, attach logs, diagrams, etc.
I think itās just a traditional text file with the resutls of the experiments copy-pasted.
Not really. I just see combining .3 of the Feb 2nd and .7 of the Feb 15th for Feb 11thā¦ its probably just the 2 combined images that is confusing me.
Any sort of version control system is essential, even for teams of 1. That way you can go back to any point in the history of your project. And sometimes you need to.
I use Evernote for this exact purpose. Every research project has a notebook. I keep records of changes, plans for future changes, and try to record results along the way.
What about just cloning notebooks after each experiment and keeping .ipynb files in version control system? I am going with this strategy and it seems to be working so far. Is there a reason why copying results into a text file would be better?
āThat weāll transfer in a future lessonā could become a meme of this course
As I understand it, itās a linear combination of things in vector space. Images are already tensors so thatās easy. For categorical data you would need to convert all categoricals to embeddings and whatnot.
So if you have day_of_month
and month
categoricals you would need to pass them through embeddings first.
transformer xl has state
Regarding applying mixup to NLP or other fields where you have categorical inputs, is it necessary to have pretrained embeddings first to do the mixups on? It doesnāt seem obvious to me how to do mixup when youāre dynamically learning the embedding weights as well. And you certainly donāt wanna do mixup on one-hot encoded inputs before embedding them.
Side note: much respect for both Jeremy and Rachel for tonight - both look visibly unwell.
I tried doing that. I need some dropout with my notes.
So I use a physical notebook for āstrategic non-code notesā then move onto one-note for coding. Separated by project. As I finish my project, I publish my findings in Github and an article to myself, my associated one-note gets deleted.
If it was important enough to remember, it is important enough to publish.
What is required to use ULMFiT with SentencePiece?
Do we have to re-train the full model using SP for tokenization?
Jeremy is speaking about parameters of a script, so they are easy to copy and paste. A notebook (as long as you save the commit/date of the corresponding libraries), it should work.
We are using ULMFit in prod, and it performs great even with small data set.
Yes, you need a pretrained model with the same tokenization.
Any examples of ULMFiT being applied to multi-task learning problems (esp. seq2seq tasks)?
thanks! that helps, but then wouldnāt embedding drop out be similar?
Iām assuming Jeremy is about to discuss this, but if there are folks who have done it with SentencePiece, Iād be interested to know how they did it.