That’s our advice yes.
3 Likes
Had the same problem. You’ll need to update your fastai library.
In the camvid example, how do we create our own labeled datasets?
Thanks, that worked!
We could probably build a cnn model of lr plots and labeled data of manually chosen rates!
What does accuracy mean for Pixelwise segmentation? Is it
#correctly classified pixels / #total number of pixels?
9 Likes
Instead of a segmentation dataset, I’ve got a text file with x and y coordinates of the location of an object in my image. How do I go about setting up my images/text file in the same way Jeremy does for the camvid image problem for training and getting x/y coord predictions?
Could some one remind again how to find the right learning rate by looking at the plot?
2 Likes
In my understanding, for stage one, you usually find the point with the steepest slope. I could be wrong.
4 Likes
My rule of thumb i to go 1/10th of the minimum, so here 3e-2. Then you can test a few values around to find the best.
7 Likes
I thought that it is what we do for stage-2 after unfreeze, where there is no clear long steep slope anymore.
Jeremy told you his trick, I’m just telling you mine ;).
11 Likes
Ah, I see. Definitely appreciate it. I just thought we are discussing what is discussed in the lecture. Sorry about that.
isn’t 1/10th of minimum (1e-1) => 1e-2 ? How is it 3e-2 ? magic number which works best which jeremy mentioned in previous videos ?
1 Like
Is it the same before and after unfreeze?
Where do you get the “3” from here? The x axis has everything like “1e-03”. This is confusing me.
1 Like
Curious to see a sample multi-class prediction code so I could return a set of labels.
Read the log scale carefully. I assure you the minimum at 3e-1
It’s a log scale, between each 1e-something, you have 8 graduations for the digits.
Well, this may be a bit too meta, but maybe we can have the fast.ai API to upload all the lr_find result, together with the accuracy of the model from immediate before/after trainings, then use DL somehow to use lr_find result as an input and try to have it output a learning rate?