Jeremy just mentioned during the live stream that you can adjust dropout per layer and how you do this by passing an array in for ps, like so: learn = ConvLearner.pretrained(arch, data, ps=[0.,0.2], ...)
4 Likes
Has anyone managed to get the Rossman code working in Crestle? Iām getting an SSL error installing feather.
Is binning a common practice to force continuous values into categorical groups?
1 Like
Yep, sometimes there are so many categories that itās not practical to use them all ā¦ especially when there isnāt a huge amount of variation between them.
Is it normal to set a column as continuous because itād be more costly to set it as categorical?
Can you double-count a column by using it categorically and continuously?
2 Likes
Just wait ā¦
Isnāt binning a form of feature engineering? If so isnāt that kind of counterintuitive to use with deep learning?
I wouldāve considered it normalization and/or part of the process in data prep.
probabilities
Will we capture the trend along the years when we code the year column as categorical?
1 Like
This note helped me a lot:
training, validation, accuracy
0.3, 0.2, 0.92 = under fitting, cycle is too short (cycle_mult=2?)
0.2, 0.3, 0.92 = over fitting
2 Likes
Maybe if you want to do a more general approach yeah, it is counter-intuitiveā¦ But if you want to win at kaggle competitionsā¦ haha
1 Like
Guessing we cant do data augmentation with structured data?
2 Likes
Is there a difference between using the type Category vs using LabelEncoder from sk-learn ?
@jeremy Can you add a link to the new paper about binning continuous features you discussed briefly
7 Likes
@arjunrajkumar, I would guess we can, for example, we could change the temperature of the day by a little bit, since it probably wouldnāt affect the sales of the day.
Thatās a big assumption 
1 Like
Note sure, I would have to look at this data.