Hacker News new | ask | show | jobs
by state_less 3392 days ago
Nice job! The samples sound good.

Is there an implementation of this to check out? It seems like you needed to write some custom, low-level code to implement this in real-time. Which libraries did you use to generate the ANNs and do the inferences?
1 comments
PieSquared 3392 days ago
We are not currently releasing any code, but hopefully the paper on arxiv is enough to make it easy to reproduce the result.

We use TensorFlow for writing and training the model and c++ with a lot of hand optimizations for inference, with assembly kernels written with PeachPy (which is an awesome piece of software!)

link
modeless 3392 days ago
I didn't know Baidu uses TensorFlow! Is that the main framework there now or are others in use?
link
PieSquared 3392 days ago
Baidu as a company doesn't use TensorFlow (as far as I know). We have our own high-performance and easy to use open source framework called PaddlePaddle [0], which is quite powerful and flexible.

However, the Baidu Research Silicon Valley AI Lab (SVAIL) allows researchers and research teams to use whatever frameworks they want to, and we have projects using TensorFlow, Torch, our own SVAIL-internal internal high-performance RNN framework, and PaddlePaddle. Using our own framework sometimes allows us to work on very high-performance implementations of various primitives and techniques that would be harder to do without complete control over the source code.

[0] http://www.paddlepaddle.org/

link