Very nice article! I saw that the author chose an exponential function to fit the learning curve. Is this always considered the best choice or it depends on the problem (e.g. dataset, architecture)?
A big question you need to answer when approaching an image classification problem: how many samples are you going to need to get to your required test metrics? Solution: use a starter dataset to model how accuracy increases with data size. Find out how:https://keras.io/examples/keras_recipes/sample_size_estimate/ …
-
-
-
There are diminishing returns to adding more data so this type of function is a good fit. Though many different variants could be valid depending on the data regime
End of conversation
New conversation -
-
-
It would be nice if we could collect results of this method for many datasets/models to see how well it predicts the point of convergence in general.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Just finished the colab. Learning from this code was very educational. Thanks. Hope your little baby boy is doing well and you and Mom are getting some sleep. :)
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Can the same solution work on deciding the size of test set? :)
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Interesting, but I’m not convinced that using the same test set of ~365 samples throughout the procedure is ideal. I would consider sampling different train/test splits every iteration.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
This is one of the most elegant tutorial I’ve ever read. I’m enchanted
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.