On 2015-12-09 4:59, Michael Markefka wrote:
Thank you for the feedback, everyone.Regarding the CPU-GPU roundtrips, I'm wondering whether it'd be possible to recursively apply the output matrix to the prior input matrix to update board positions within the GPU and without any actual (possibly CPU-based) evaluation until all branches come up with game ending states. I assume illegal moves would mostly fall away when sticking to the top ten or top five move considerations provided by the CNN. As for performance, I could imagine initialization being relatively slow, but wouldn't be surprised if the GPU-based CNN performance could offer a branch size, running through many parallel boards with comparatively minor performance impact, where this outweighed the initial overhead again. Whether this would provide a better evaluation function than MCTS I don't know, but just like Alvaro I would love to see this tried, even if just to rule it out for the moment. I've got a GTX 980 Ti on a 4790k with 16 GB at home. For a low key test I could run Windows (CUDA installed and running, tested with pylearn2) or Ubuntu from a live setup on USB and would be willing to run test code, if somebody provided a package I could simply download and execute. All the best Michael On Tue, Dec 8, 2015 at 7:52 PM, Álvaro Begué <alvaro.be...@gmail.com> wrote:Of course whether these "neuro-playouts" are any better than the heavy playouts currently being used by strong programs is an empirical question. But I would love to see it answered... On Tue, Dec 8, 2015 at 1:31 PM, David Ongaro <david.ong...@hamburg.de> wrote:Did everyone forget the fact that stronger playouts don't necessarily lead to an better evaluation function? (Yes, that what playouts essential are, a dynamic evaluation function.) This is even under the assumption that we can reach the same number of playouts per move. On 08 Dec 2015, at 10:21, Álvaro Begué <alvaro.be...@gmail.com> wrote: I don't think the CPU-GPU communication is what's going to kill this idea. The latency in actually computing the feed-forward pass of the CNN is going to be in the order of 0.1 seconds (I am guessing here), which means finishing the first playout will take many seconds. So perhaps it would be interesting to do something like this for correspondence games, but not for regular games. Álvaro. On Tue, Dec 8, 2015 at 12:03 PM, Petr Baudis <pa...@ucw.cz> wrote:Hi! Well, for this to be practical the entire playout would have to be executed on the GPU, with no round-trips to the CPU. That's what my email was aimed at. On Tue, Dec 08, 2015 at 04:37:05PM +0000, Josef Moudrik wrote:Regarding full CNN playouts, I think that problem is that a playout is a long serial process, given 200-300 moves a game. You need to construct planes and transfer them to GPU for each move and read result back (at least with current CNN implementations afaik), so my guess would be that such playout would take time in order of seconds. So there seems to be a tradeoff, CNN playouts are (probably much) better (at "playing better games") than e.g. distribution playouts, but whether this is worth the implied (probably much) lower height of the MC tree is a question. Maybe if you had really a lot of GPUs and very high thinking time, this could be the way. Josef On Tue, Dec 8, 2015 at 5:17 PM Petr Baudis <pa...@ucw.cz> wrote:Hi! In case someone is looking for a starting point to actually implement Go rules etc. on GPU, you may find useful: https://www.mail-archive.com/computer-go@computer-go.org/msg12485.html I wonder if you can easily integrate caffe GPU kernels in another GPU kernel like this? But without training, reimplementing the NN could be pretty straightforward. On Tue, Dec 08, 2015 at 04:53:14PM +0100, Michael Markefka wrote:Hello Detlef, I've got a question regarding CNN-based Go engines I couldn't find anything about on this list. As I've been following your posts here, I thought you might be the right person to ask. Have you ever tried using the CNN for complete playouts? I know that CNNs have been tried for move prediction, immediate scoring and move generation to be used in an MC evaluator, but couldn't find anything about CNN-based playouts. It might only be feasible to play out the CNN's first choice move for evaluation purposes, but considering how well the performance of batch sizes scales, especially on GPU-based CNN applications, it might be possible to setup something like 10 candidate moves, 10 reply candidate moves and then have the CNN play out the first choice move for those 100 board positions until the end and then sum up scores again for move evaluation (and/or possibly apply some other tried and tested methods like minimax). Given that the number of 10 moves is supposed to be illustrative rather than representative, other configurations of depth and width in position generation and evaluation would be possible. It feels like CNN can provide a very focused, high-quality width in move generation, but it might also be possible to apply that quality to depth of evaluation. Any thoughts to share? All the best Michael On Tue, Dec 8, 2015 at 4:13 PM, Detlef Schmicker <d...@physik.de> wrote:-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Hi, as somebody ask I will offer my actual CNN for testing. It has 54% prediction on KGS 6d+ data (which I thought would be state of the art when I started training, but it is not anymore:). it has: 1 2 34 libs playing color1 2 34 libs opponent colorEmpty points last move second last move third last move forth last move input layers, and it is fully convolutional, so with just editing the golast19.prototxt file you can use it for 13x13 as well, as I did on last sunday. It was used in November tournament as well. You can find it http://physik.de/CNNlast.tar.gz If you try here some points I like to get discussion: - - it seems to me, that the playouts get much more important with such a strong move prediction. Often the move prediction seems better the playouts (I use 8000 at the moment against pachi 32000 with about 70% winrate on 19x19, but with an extremely focused progressive widening (a=400, a=20 was usual). - - live and death becomes worse. My interpretation is, that the strong CNN does not play moves, which obviously do not help to get a group life, but would help the playouts to recognize the group is dead. (http://physik.de/example.sgf top black group was with weaker move prediction read very dead, with good CNN it was 30% alive or so :( OK, hope you try it, as you know our engine oakfoam is open source :) We just merged all the CNN stuff into the main branch! https://bitbucket.org/francoisvn/oakfoam/wiki/Home http://oakfoam.com Do the very best with the CNN Detlef code: if (col==Go::BLACK) { for (int j=0;j<size;j++) for (int k=0;k<size;k++) { for (int l=0;l<caffe_test_net_input_dim;l++) data[l*size*size+size*j+k]=0; //fprintf(stderr,"%d %d %d\n",i,j,k); int pos=Go::Position::xy2pos(j,k,size); int libs=0; if (board->inGroup(pos)) libs=board->getGroup(pos)->numRealLibs()-1; if (libs>3) libs=3; if (board->getColor(pos)==Go::BLACK) { data[(0+libs)*size*size + size*j + k]=1.0; //data[size*size+size*j+k]=0.0; } else if (board->getColor(pos)==Go::WHITE) { //data[j*size+k]=0.0; data[(4+libs)*size*size + size*j + k]=1.0; } else if (board->getColor(Go::Position::xy2pos(j,k,size))==Go::EMPTY) { data[8*size*size + size*j + k]=1.0; } } } if (col==Go::WHITE) { for (int j=0;j<size;j++) for (int k=0;k<size;k++) {//fprintf(stderr,"%d %d %d\n",i,j,k); for (int l=0;l<caffe_test_net_input_dim;l++) data[l*size*size+size*j+k]=0; //fprintf(stderr,"%d %d %d\n",i,j,k); int pos=Go::Position::xy2pos(j,k,size); int libs=0; if (board->inGroup(pos)) libs=board->getGroup(pos)->numRealLibs()-1; if (libs>3) libs=3; if (board->getColor(pos)==Go::BLACK) { data[(4+libs)*size*size + size*j + k]=1.0; //data[size*size+size*j+k]=0.0; } else if (board->getColor(pos)==Go::WHITE) { //data[j*size+k]=0.0; data[(0+libs)*size*size + size*j + k]=1.0; } else if (board->getColor(pos)==Go::EMPTY) { data[8*size*size + size*j + k]=1.0; } } } if (caffe_test_net_input_dim > 9) { if (board->getLastMove().isNormal()) { int j=Go::Position::pos2x(board->getLastMove().getPosition(),size); int k=Go::Position::pos2y(board->getLastMove().getPosition(),size); data[9*size*size+size*j+k]=1.0; } if (board->getSecondLastMove().isNormal()) { int j=Go::Position::pos2x(board->getSecondLastMove().getPosition(),size); int k=Go::Position::pos2y(board->getSecondLastMove().getPosition(),size); data[10*size*size+size*j+k]=1.0; } if (board->getThirdLastMove().isNormal()) { int j=Go::Position::pos2x(board->getThirdLastMove().getPosition(),size); int k=Go::Position::pos2y(board->getThirdLastMove().getPosition(),size); data[11*size*size+size*j+k]=1.0; } if (board->getForthLastMove().isNormal()) { int j=Go::Position::pos2x(board->getForthLastMove().getPosition(),size); int k=Go::Position::pos2y(board->getForthLastMove().getPosition(),size); data[12*size*size+size*j+k]=1.0; } } -----BEGIN PGP SIGNATURE----- Version: GnuPG v2.0.22 (GNU/Linux) iQIcBAEBAgAGBQJWZvOlAAoJEInWdHg+Znf4t8cP/2a9fE7rVb3Hz9wvdMkvVkFS 4Y3AomVx8i56jexVyXuzKihfizVRM7x6lBiwjYBhj4Rm9UFWjj2ZvDzBGCm3Sy4I SpG8D01VnzVR6iC1YTu3ecv9Wo4pTjc7NL5pAxiZDB0V7OTRklfZAYsX4mWyHygn cr1pIb79/9QfBf/johmuutXJIwYfVG9ShR1+udbxs3aU3QDAbJJ4eTs8oj+NqFpg JolEEEg3wY693e77SqbUbjxR3kSsysoz9h1nKnR/ZjHByqlwNvSz9ho9eU0rKhaK GSQ22/c1VPIZhr24FYBbYNYweOzDtonLpuUFCPSnYVels3h/I/LlqV3MeDo6wuZ2 QCPp5+11o4JzvEt7A4zfJCtEOEH0W2/+IjRcIkAVOo65OV/pPsz2EjHehMU6PC6m vXA/kPx0jqUm1qSb0qCgMq5ZvSqfpcCY7JOlkEwkDBS1fty9sU0hqst3zXR0KGtn rFuoREmQYi/mkjZfS2Q4AHiZUDbDZUKzRegUA+gR/eKAmJsmWeTDEI9ZAXgxL0cB p1HGBNDEUKGk+ruq0gIe5vYygyBcJV0BbbBnweDjeZnlG8vLUAVoMF6V/q3gkZb1 P61rfE4d9dohfGBsZ+UWltRyWMj09ieR2G2zCDpIXyxEuoV6CTAlLzDuhmqFa2ma Fp3lK/uLhOucXwBtStdx =E47K -----END PGP SIGNATURE----- _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go-- Petr Baudis If you have good ideas, good data and fast computers, you can do almost anything. -- Geoffrey Hinton _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go-- Petr Baudis If you have good ideas, good data and fast computers, you can do almost anything. -- Geoffrey Hinton _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
