0
$\begingroup$

In my experience, both reducing batch_size and increasing epochs can decrease loss amount. But I like to know is there any difference to using which one? Has decreased loss amount same meaning and it's not important you reached it with what way(I mean has no effects on results)?

For ex, I got same loss amount 2.5e-4 with both the following case:

1. batch_size = 1 , epochs = 100
2. batch_size = 60 , epochs = 1000

Are they same result?

$\endgroup$
1
$\begingroup$

The concept of overfitting does not generalize over a specific combination of batch size and epochs. It depends on your data and the architecture of your model

A friend of mine ran into these scenarios with a CPU based image classifier:

1) If I use more epochs ,it may take me a lot of time to come to a desirable outcome.

2) If I prefer small batch-sizes over small epochs , It might take less time time to compute , but not reach the desirable outcome by that epoch limit.

I used a GPU and my results were different. Using low epochs , and better convolutional architecture , I reached a better accuracy with not so small batch sizes.

I increased epochs, and my accuracy improved until I felt I reached overfitting. I increased batch-sizes and my accuracy was not increasing at a decent rate.

I had to come to balance at which my model was acceptable.

Its a balance , yes. A balance which I am afraid, the designer needs to take care of. Not always they are inversely proportional. The data set and the architecture makes all the difference in this debate I am afraid.

$\endgroup$

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.