[HW4] Problem 2.1 & 3.1 #161
-
|
Hello Professor, Sorry for the trivial question, but in problems 2.1 & 3.1 is it ok to train on a small subset of the datasets (like they've done in tutorials), instead of training on the full datasets? Training on the full datasets takes a significant amount of time on weak machines (and on collab too) :( Thanks! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
If it takes a week, I think you are doing something wrong. If you are training around 3 epochs, it should take a few hours at most. If it takes too much time, consider training with 1 epoch only. That being said, you won't be penalized for training with a subset as long as it achieves a reasonable accuracy (that is, EM bigger than 50%). |
Beta Was this translation helpful? Give feedback.
If it takes a week, I think you are doing something wrong. If you are training around 3 epochs, it should take a few hours at most. If it takes too much time, consider training with 1 epoch only. That being said, you won't be penalized for training with a subset as long as it achieves a reasonable accuracy (that is, EM bigger than 50%).