|
I realize this is a bit of a long shot, but if anyone here used both Torch7 and Theano/Pylearn2, would you be willing to compare and contrast them? I'm not looking for information I can easily read on their web sites, such as that Torch7 is based on LuaJIT, while Theano/Pylearn2 builds on NumPy/Python and includes symbolic differentiation, and that both can use CUDA. Instead, I'm interested in things like a. Quality (bugs in math, performance bugs, bugs causing crashes), b. Extensibility (if dropout and maxout hadn't been invented yet, would users who are not developers of these frameworks be able to add them themselves?) c. Any other things that aren't obvious from reading the web sites, but you think are important to know. Lastly, are there other neural network frameworks similar to Theano/Pylearn2 and Torch7 that could compete with them? |
|
I know this is an old thread, but I came across it while search theano on this board. Since it's unanswered, I'll chime in, in the event someone else has the same question. The best reference I found for comparing theano to Torch7 was this paper: Theano: new features and speed improvements http://arxiv.org/abs/1211.5590 It's written by the LISA folks (creators of pylearn2 which uses theano). But I found it to be pretty unbiased. |
|
for category c: I was told that the Deep Learning teams at Google, Facebook and Microsoft use Torch7. Theano is by LISA, which is headed by Yoshua Bengio, who vows not to join a big company so he can keep pushing research his own way. As a result, one can expect that Torch7 will become the established toolkit in industry (the real world?). Pity though, I like python... |
|
I also heard big compagnie also use torch7, but some of them also use Theano. Most of the answer I saw on torch7 mailing list for some advanced feature is that they are implemented in private repo. So not available. p.s. I'm a Theano author. If torch7 or any other software release important new implementation, they can be wrapped in Theano. We did that with cuda convnet. Personally, I think the main difference is that Theano is a compiler and torch7 a library. |
