I'm running in a bit of trouble with my bot on the LoCM contest, but i believe it applies to other contests as well.
I'm trying to fit a neural network offline, write the trained weights in the code, and load them into keras/tensorflow on turn 1 (I'm using python2).
The problem is that the lines to import tensorflow and keras take a lot of time and my bot timeouts most of the time.
t_start = time.time()
import tensorflow as tf
from tensorflow import keras
print >> sys.stderr, "Time to import: %.2f" % ((time.time() - t_start)*1000.0)
Standard Error Stream:
Time to import: 2853.28
It varies between 2.5 and 2.8 seconds.
I know it's not that hard to re-implement the predict method from keras, but given that NN are hot right now, I think it's a bit sad to have every player who wants to try NNs re-implement it.
So yeah, if anyone has a solution to that issue, it would be very welcome
PS: On a similar note, keras requires h5py to store and load models, I've re-implemented an alternative to do that using jsons since h5py is not on the servers. I guess it would be nice to have h5py available.