sethtroisi / leela-zero Goto Github PK

I'm developing on top of NTT123/udp_server in poorly named branch w__upd_server_any_instance

improvements

• Theano neural net refactored into a class (TheanoLZN)
• server.py refactored into named methods
• server pulls input from first batch_size instances read (instead of 0-8,9-15,16-24...)
• timing information at lower granularity

Example Timing Info

	minibatch iteration 30000x2 = 60000 playouts
	collecting data: 25.1s = 835.2 micros/iter, samples [0.4 ms, 0.4 ms, 0.5 ms, 0.4 ms, 0.4 ms] => 418.3 micros/iter
	running net: 90.2s = 3.0 millis/iter, samples [2.9 ms, 2.8 ms, 3.5 ms, 2.7 ms, 2.8 ms] => 2.9 millis/iter