Please add your in-class comments or questions here!
International fellows can chat on your private group
@sbhaaf can’t access the Part 2 forums. Can someone add him?
“You mentioned @sbhaaf but they won’t be notified because they do not have access to this category. You will need to add them to a group that has access to this category.”
where is the private group?
See the international fellows thread - each group is responsible for creating their own forum private group.
I have the same question . I am in group 2, not sure how to join the group.
Is it possible for jeremy to turn off forum notifications. I find them a little distracting when they pop up in the stream as he’s discussing a topic.
Does Jeremy have Flux on? The screen looks very orange.
Can we also turn off the lights above the screen?
I had a question in regards to style transfer for cartoons. With cartoons, when we think of transferring the style what we really mean is transferring the contours of the cartoon to redraw the content in that style. This is obviously not what style transferring is doing here. How might one implement this? I’ve tried actually selecting activations that correspond with edges and such as indicated by one of the conv visualization papers, and comparing outputs from specifically those activations. I’ve been able to create an image with the edges of the style image, but I have no idea how one might use this to alter the content to have similar style edges
When a forum notification pops up, Jeremy can click the gear icon to disable future ones, I think.
Shouldn’t the per layer weights for the style loss be learned by the optimizer?
@amanmadaan What would your goal be for optimizing per layer weights (since we are just trying to create interesting images)?
Different layers contribute differently to the style loss, perhaps not all layers should be given equal weightage while optimizing?
EDIT: The original paper uses a constant for all these weights “(wl =1/5 in those layers, wl = 0 in all other layers)”
EDIT 2: I was referring to the values of style_wgts.
@amanmadaan In the example Jeremy showed, he gave higher weights to later layers. Giving different weights to different layers creates different effects (but you would have to define the “optimal effect” for an optimizer to learn “optimal layer weights”)
If we mix out both style and content loss together, how can we make sure if we are interested more in the style or content?
Why multiple images?
how many images do we need to get good results for the CNN style transfer?
can I use any images? just pick randomly 100k images from google / imagenet?
Haha I have to wonder, can we just stick a CNN between any two things and it will learn the relationship?