TensorFlow vs Caffe

3,428 views
Skip to first unread message

Evan Weiner

unread,
Nov 9, 2015, 9:06:04 PM11/9/15
to Caffe Users
TensorFlow was just released (tensorflow.org). I haven't been able to fully digest it yet and wondering if others can compare it to Caffe? 

Matt Ragoza

unread,
Nov 9, 2015, 9:32:52 PM11/9/15
to Caffe Users
It's probably too early for anyone outside of Google to have fully explored it's capabilities but I'm also very much interested in seeing how they compare. As a side note, it looks like Yangqing Jia, original author of Caffe, is listed on the TensorFlow white paper as a co-author. My expectations are high.

Steven Clark

unread,
Nov 10, 2015, 12:22:48 PM11/10/15
to Caffe Users
Competition is a grand thing, but it will be hard to resist the might of Google. About the only thing that would make me reluctant to switch would be if benchmarks show TF to be significantly slower than Caffe for the same model architecture / GPU configuration. TensorFlow certainly seems to have an elegant syntax, similar to Theano.
Message has been deleted

ath...@ualberta.ca

unread,
Nov 10, 2015, 3:43:22 PM11/10/15
to Caffe Users
Yanqqing, could you please comment on this? :)  Does TensorFlow supercede Caffe in every way?

Evan? What have you heard?

Oluwatobi Olabiyi

unread,
Nov 10, 2015, 7:51:13 PM11/10/15
to Caffe Users
Well TensorFlow is still usind CudNN6.5 (R2) while Caffe is already using CudNN7 (R3). Although in the long run, TF may supersede Caffe but it's currently behind at least from GPU implementation perspective.  

Steven Clark

unread,
Nov 11, 2015, 5:03:40 PM11/11/15
to Caffe Users
This thread is relevant:
 
"Thanks @soumith! Yes, our only launch criterion for convnets was 'GoogLeNet within distance from CuDNN[R2]', and we've punted on a lot of performance work, including upgrading CuDNN, until after the initial release. Expect a lot of movement on that front in the coming weeks."
 
So they haven't quite optimized it yet. Soon, hopefully.

cooling

unread,
Nov 12, 2015, 12:38:07 PM11/12/15
to Caffe Users
Is this accurate for Caffe? 


Or are they not using CUDNN with caffe? 

Evan Shelhamer

unread,
Nov 12, 2015, 1:14:57 PM11/12/15
to cooling, Caffe Users
None of the Caffe results in this set of benchmarks make use of Caffe + cuDNN -- all the cuDNN benchmarks are with Torch. That is why Caffe is listed as "Caffe (native)". They are pure Caffe benchmarks on purpose according to https://github.com/soumith/convnet-benchmarks/blob/master/caffe/install.sh#L13.

Evan Shelhamer



--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
 To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/a9866be3-eaa7-49bc-8a03-6241fd3df0b8%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

Evan Shelhamer

unread,
Nov 12, 2015, 1:16:04 PM11/12/15
to cooling, Caffe Users
Oh and to be clear that's fine -- I think it is explained in the benchmark description.

Evan Shelhamer


cooling

unread,
Nov 12, 2015, 1:43:57 PM11/12/15
to Caffe Users, div...@gmail.com
Seems unfair still. We should push for Caffe+CUDNN.
Reply all
Reply to author
Forward
0 new messages