python version hfo.step() is very slow

[liyuanl6666]

Can you check why the python version hfo.step() is so slow? I'm trying to use the python interface; found that the step() function is much much slower than the C version (with the caffe code): for example, the time used to generate actions with actor network is about 40 seconds (across multiple episodes), and the time to execute hfo.step() is about 50 seconds... this is way much slower than the c++ version.

[drallensmith]

50 seconds for each execution of hfo.step()? That does not accord with my experience. Could you try running some of the Python examples and see how long they're taking per episode (which has multiple invocations of hfo.step(), of course)?

[liyuanl6666]

42.94 seconds spent on hfo.step() in 2000 plays(each play has roughly 100 steps) While for the c++ version, the code can finish 2000 plays within 32 seconds(not only the time used on hfo.step, but all).

[mhauskn]

In general the python interface is going to be a slower since the native code is c++ and python calling the native c++ through a ctypes interface. Python's call to step is here:

https://github.com/LARG/HFO/blob/master/hfo/hfo.py#L177

and the ctypes implementation is:

https://github.com/LARG/HFO/blob/master/hfo/hfo_c_wrapper.h#L40

I'm not too sure how to further optimize the routine, but if you have ideas, I'd be happy to listen.

On Wed, Sep 20, 2017 at 8:09 PM, liyuanl6666 notifications@github.com wrote:

42.94 seconds spent on hfo.step() in 2000 plays(each play has roughly 100 steps) While for the c++ version, the code can finish 2000 plays within 32 seconds(not only the time used on hfo.step, but all).

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/LARG/HFO/issues/45#issuecomment-331038512, or mute the thread https://github.com/notifications/unsubscribe-auth/AABNOYRReGkhcHKqCj266MUH1p7m459Wks5skdNWgaJpZM4PeCZz .

[drallensmith]

Also:

• hfo.step(), from a python profiling standpoint, is going to include everything being done by the c++ code.
• What actions is the python code using (how high-level) vs the c++ code? Higher-level actions (with the somewhat odd exception of intercept vs go_to_ball) tend to have more processing needed.
• Similarly, what state interface is the python code using and what state interface is the c++ code using?