Little generalization, otherwise it’s straight from 7x7x512 to no of classes (say 2) . With a layer in between, it helps the model to choose exact features to classify it into one of the classes
2 Likes
do you need to have 2 bias (user/movie)? or is 1 bias is ok?
It’s better to have one per user (cause some are harsher than others) and one per movie (cause some movies are bad and others are good).
So in that example he had 5 numbers for activations/weights - as I understand, this is the number of neurons in the layer. How is this number decided? If you had too many you would only need one layer and it would just be an exact map, so the model is useless for new data.
Is there some heuristic for number of neurons per layer?
4 Likes
we can use python package “chardet” to determine encoding of a file.
5 Likes
so does it divides the cycle into those many parts
say x ,x+divfactor,x+div factor in a cycle
Please wait with the questions on 1cycle, Jeremy is going to explain it later.
2 Likes
How to use weight decay(regularization technique) in fastai library?
1 Like
@lesscomfortable This question had 6 likes! Thanks
Does fastai provide API for downloading “Movielens 100k data dataset”?
Added. What is the loss function by the way?
It is nll_loss as par learn.loss_func
I think you mean one number for user bias, and one number for movie bias. Not one bias per user and one bias per movie, right?
Each user gets their individual bias (one number) and each movie their individual bias (one number again).
2 Likes
What does the skip_end argument do? When should it be used?
6 Likes
how do you determine the bias number?
It is learned by gradient descent
not clear about the details of latent factors… if this can be explained more here or asked to…
I think it’s to omit the end part of the graph so that it’s easier to see the earlier part of the graph.
2 Likes
Skip end is just telling the plotter to skip the last 3 values for the curve.
3 Likes