Hacker News
new
|
past
|
comments
|
ask
|
show
|
jobs
|
submit
login
Saving memory using gradient-checkpointing
(
github.com/openai
)
52 points
by
stablemap
on Jan 15, 2018
|
hide
|
past
|
favorite
|
3 comments
eggie5
on Jan 16, 2018
|
next
[–]
It takes about 4GB of memory to train a VGG network w/ batch size 8 on Tensorflow. Could I use a larger batch size at the expense of computation time w/ this module?
iaw
on Jan 17, 2018
|
parent
|
next
[–]
Yes, potentially a batch-size 64 with a 25% increase in run-time based on the figures they reported.
supermdguy
on Jan 16, 2018
|
prev
[–]
I wonder if this would work when using keras with the tensorflow backend. Has anyone tried it?
Guidelines
|
FAQ
|
Lists
|
API
|
Security
|
Legal
|
Apply to YC
|
Contact
Search:
