I hope we will have some remote sensed data examples. I wonder if this handles 4 band or multispecteral data and in what format?
1 Like
How do I load saved weights/checkpoint and continue training from there?
Is the library by default will utilise multi-gpu in parallel?
7 Likes
I setup all the things in google colab and fasta ai libray is working but i need to use num_worker=0 in dataloader (in pytroch) can i do that in fast.ai
1 Like
Fast AI library is really cool. I use Keras a lot but the simplicity is so cool. For instance, callbacks used in Keras 2.x
Learn.save() learn.load(). See notebook
2 Likes
3 Likes
What is the suggested batch size when using one-cycle policy? What’s the reason behind it?
1 Like
What units is the loss expressed in?
1 Like
Imagenet weights vs coco weights?
1 Like
The maximum you can support on your GPU. Jeremy will present the LR Finder in the next lesson and explain how to pick an appropriate learning rate.
5 Likes
The top loss plot’s 4th values are really close to 1. Does that mean that the probability of the “correct” class is close to 1? Then the loss shouldn’t be high. I probably misunderstood something.
3 Likes
Why is the probability of the actual class 1.00 in top_loss images that jeremy showed ?
4 Likes
How do we used most_consfused to improve our model?
4 Likes
What do the numbers on top of plot top losses represent?
1 Like
How fit is different from fit_one_cycle? Is the later one a shortcut for the former one but with OneCycleScheduler callback or something?
8 Likes
All links to documentation will be fixed by tomorrow.
Are we going to learn on what and how to tweak the defaults chosen and set for us by the fast.ai library?
Yes. 
1 Like