All Questions
1 question
1
vote
1
answer
614
views
Instead of accumulating the gradient, can we accumulate loss values?
I have read and used Gradient Accumulation as a method to handle large batch size on smaller memory restrictions. It is described as following:
...
- 41.4k