科技回声

Wondering if non-standard interpreters like pypy and cpython can be used to speed up training.

I got stuck on this problem a while ago. I was doing reinforcement learning and the environment needed to run in pypy to make it fast enough. I ended up using Pyro4 to bridge pypy (environment) and CPython (keras).

It seems like there's some progress, including a merged PR to tensorflow that should end up making this work, but I'm confused as to the exact status:<a href="https://github.com/tensorflow/tensorflow/issues/252" rel="nofollow">https://github.com/tensorflow/tensorflow/issues/252</a>

Asked about this before. I was told no one cared since the real work is done in C and would not benefit from pypy. So the python level optimizations would not affect much.

> non-standard interpreters like pypy and cpython<p>Isn't CPython the default and most standard interpreter?

Wondering if non-standard interpreters like pypy and cpython can be used to speed up training.

Asked about this before. I was told no one cared since the real work is done in C and would not benefit from pypy. So the python level optimizations would not affect much.

> non-standard interpreters like pypy and cpython<p>Isn't CPython the default and most standard interpreter?

Ask HN: Keras on PyPy?

4 条评论

Ask HN: Keras on PyPy?

4 条评论