accuracy
is a metric, not a loss function. You couldn’t train well with it because it has no gradients (they’re zero everywhere pretty much). The usual loss function used is cross entropy, but it depends on your task.
The function isn’t coded for that so you would have to tweak its implementation.
I want to remind everyone that as of last week, the docs now have a working search function as well.
yes VGG is there
https://docs.fast.ai/vision.models.html
https://pytorch.org/docs/stable/torchvision/models.html
You are right. I was in the branch release 1.0.15, thanks
They’re not plugged in the create_cnn
function yet, as they’re a bit outdated, but it shouldn’t be too much work to make them work with fastai.
Yes, if you have three classes {A, B, C}
you can train three binary classifiers:
{A, not A}
{B , not B}
{C, not C}
Could you give more detail how you choose the threshold?
if we have just yes or no for a given object ,is no of class one or two?
Thank you!
You try and see which value works better
That’s two classes
Single classification with two classes. Multi class is when an object can have several tags.
Yeah, I didn’t seem to see them in the docs. But, I like the ones present already. They seem to be really cool.
LR plot is plot vs Validation loss or Training loss
what is the loss function used for planets?
For the deployment app, we need to " Upload your trained model file (for example stage-2.pth
) to a cloud service like Google Drive or Dropbox". How can we directly upload that saved .pth to google drive? Or how to download it to local machine and then upload it?
Thank you.
It’s kind of the same since it’s training for the first time on data it hasn’t seen, but it’s the training loss.
pip install --upgrade fastai
Did you try this?