Generally speaking, we all know it's to save spaces with incremental learning.

According to the ques in stackoverflow , it also said that.

But what's the disadvantages?

What I know from my experiments is two points below:

  1. Train with subsets of data but shouldn't be too small. I prepared very small datasets and the predict result is very worse.

  2. When training for a very long time, some elder behavors will be forgotten due to the multiple training epochs.

That's all from my experience when training with xgboost incrementally.

Or anything else?

  • $\begingroup$ An epoch general refers to as one full pass through the training data. So the training set is repeated each epoch. Each epoch consists of many batches, usually. $\endgroup$ – jonnor Apr 18 at 0:11

It is more complicated. Instead of just having all the entire dataset in memory and doing performing a single fit, there are now more things to do:

  • Need a way to load data in batches from disk
  • Need to decide a new hyperparameter, the batchsize

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.