Narsil
commented
6 years ago Hmm yes it does create a bunch of files. There is a file for each move of every game of every model.
The interest is that the model only has to parse the directory of a model once (which is usually pretty fast) and can then open the files only once for each batch in training. During training the samples are taken randomly from any move of any game. Random access can be pretty slow pretty fast for huge data.
I could zip for past models as they are not used after some point (though Deepmind says they use the last 500k games which would correspond to 40M files in the current architecture). But it's not really my current focus as I feel there are still some optimizations that could be done.
Do you have any other idea to make it better ?
tianshuo
It ran for a couple of days and found several new best models. However, it also creates numerous files (502,586 items, totalling 5.6 GB). The models directory is large and the games directory has most of the files. Perhaps zipping would be worthwhile. In any case, I'm happy to restart it again after you have had a chance to make more improvements. Thanks again for sharing.