Probably an ignorant question, but could you use the resnet34 model to apply transfer learning to tabular data?
Does random initialization work?
There is no precompute=True in fastai v1.
2 Likes
Not really, since that model expects images as an input.
It works if you don’t have any pretrained model. If you do, then it is better to load these weights and finetune from there. In any case @sgugger 's formula will be used to update the weights as the network trains.
1 Like
You have no other choice for new layers.
A model pretrained on imagenet was randomly initialized before training.
1 Like
File b’data/ml-100k/u.data’ does not exist in the “lesson4-collab”. Do I need to download and unzip " http://files.grouplens.org/datasets/movielens/ml-100k.zip"?
How does progressive resize works? When the image size increase, will the weights and activations dimensions increase as well?
yes, you can look for affine transformations in Linear Algebra. Functions that rotate, move, etc
1 Like
No, because a convolution doesn’t expect a given size. Jeremy will explain them more in detail in next lesson.
2 Likes
To try to make sure I understand this…for an embedding 2, 3, 1, would the matrix multiplication be
[0 1 0
0 0 1
1 0 0]?
1 Like
Is this how word embeddings work?
Your last row is off, but yes otherwise.
Word embeddings are exactly the same.
2 Likes
How can we visualize latent features?
3 Likes
Jeremy is going to show you in a few minutes 
1 Like
Ziiing to Scientology! Funny
Any rule of thumb for embedding sizes?
6 Likes
who is calculating / providing the bias value ?
2 Likes