Re: [Computer-go] CNN for winrate and territory

[Detlef Schmicker]

Exactly the one from the cited paper:


The best network had one convolutional layer with 64 7x7
filters, two convolutional layers with 64 5x5 filters, two lay-
ers with 48 5x5 filters, two layers with 32 5x5 filters, and
one fully connected layer.


I use caffe and the definition of the training network is:

name: "LogReg"
layers {
  name: "mnist"
  type: DATA
  top: "data_orig"
  top: "label"
  data_param {
source: "train_result_leveldb/"
batch_size: 256
  }
  include: { phase: TRAIN }
}
layers {
  name: "mnist"
  type: DATA
  top: "data_orig"
  top: "label"
  data_param {
source: "test_result_leveldb/"
batch_size: 256
  }
  include: { phase: TEST }
}

layers {
  name: "slice"
  type: SLICE
  bottom: "data_orig"
  top: "data"
  top: "data_territory"
  slice_param {
slice_dim: 1
slice_point : 2
  }
}

#this part should be the same in learning and prediction network
layers {
  name: "conv1_7x7_64"
  type: CONVOLUTION
  blobs_lr: 1.
  blobs_lr: 2.
  bottom: "data"
  top: "conv2"
  convolution_param {
num_output: 64
kernel_size: 7
pad: 3
weight_filler {
  type: "xavier"
  }
  bias_filler {
  type: "constant"
  }
}
}

layers {
  name: "relu2"
  type: TANH
  bottom: "conv2"
  top: "conv2u"
}

layers {
  name: "conv2_5x5_64"
  type: CONVOLUTION
  blobs_lr: 1.
  blobs_lr: 2.
  bottom: "conv2u"
  top: "conv3"
  convolution_param {
num_output: 64
kernel_size: 5
pad: 2
weight_filler {
  type: "xavier"
  }
  bias_filler {
  type: "constant"
  }
}
}

layers {
  name: "relu3"
  type: TANH
  bottom: "conv3"
  top: "conv3u"
}

layers {
  name: "conv3_5x5_64"
  type: CONVOLUTION
  blobs_lr: 1.
  blobs_lr: 2.
  bottom: "conv3u"
  top: "conv4"
  convolution_param {
num_output: 64
kernel_size: 5
pad: 2
weight_filler {
  type: "xavier"
  }
  bias_filler {
  type: "constant"
  }
}
}

layers {
  name: "relu4"
  type: TANH
  bottom: "conv4"
  top: "conv4u"
}

layers {
  name: "conv4_5x5_48"
  type: CONVOLUTION
  blobs_lr: 1.
  blobs_lr: 2.
  bottom: "conv4u"
  top: "conv5"
  convolution_param {
num_output: 48
kernel_size: 5
pad: 2
weight_filler {
  type: "xavier"
  }
  bias_filler {
  type: "constant"
  }
}
}

layers {
  name: "relu5"
  type: TANH
  bottom: "conv5"
  top: "conv5u"
}

layers {
  name: "conv5_5x5_48"
  type: CONVOLUTION
  blobs_lr: 1.
  blobs_lr: 2.
  bottom: "conv5u"
  top: "conv6"
  convolution_param {
num_output: 48
kernel_size: 5
pad: 2
weight_filler {
  type: "xavier"
  }
  bias_filler {
  type: "constant"
  }
}
}

layers {
  name: "relu6"
  type: TANH
  bottom: "conv6"
  top: "conv6u"
}


layers {
  name: "conv6_5x5_32"
  type: CONVOLUTION
  blobs_lr: 1.
  blobs_lr: 2.
  bottom: "conv6u"
  top: "conv7"
  convolution_param {
num_output: 32
kernel_size: 5
pad: 2
weight_filler {
  type: "xavier"
  }
  bias_filler {
  type: "constant"
  }
}
}

layers {
  name: "relu7"
  type: TANH
  bottom: "conv7"
  top: "conv7u"
}


layers {
  name: "conv7_5x5_32"
  type: CONVOLUTION
  blobs_lr: 1.
  blobs_lr: 2.
  bottom: "conv7u"
  top: "conv8"
  convolution_param {
num_output: 32
kernel_size: 5
pad: 2
weight_filler {
  type: "xavier"
  }
  bias_filler {
  type: "constant"
  }
}
}

layers {
  name: "relu8"
  type: TANH
  bottom: "conv8"
  top: "conv8u"
}

layers {
  name: "flat"
  type: FLATTEN
  bottom: "conv8u"
  top: "conv8_flat"
}

layers {
  name: "split"
  type: SPLIT
  bottom: "conv8_flat"
  top: "conv8_flata"
  top: "conv8_flatb"
}

layers {
  name: "ip"
  type: INNER_PRODUCT
  bottom: "conv8_flata"
  top: "ip_zw"
  inner_product_param {
num_output: 361
weight_filler {
  type: "xavier"
  }
bias_filler {
  type: "constant"
  }
   }
}

layers {
  name: "sigmoid"
  type: SIGMOID
  bottom: "ip_zw"
  top: "ip_zws"
}

layers {
  name: "ip2"
  type: INNER_PRODUCT
  bottom: "conv8_flatb"
  top: "ip_label"
  inner_product_param {
num_output: 121
weight_filler {
  type: "xavier"
  }
bias_filler {
  type: "constant"
  }
   }
}


#only learning framework
layers {
  name: "flat"
  type: FLATTEN
  bottom: "data_territory"
  top: "flat"
}
layers {
  name: "loss"
  type: EUCLIDEAN_LOSS
  bottom: "ip_zws"
  bottom: "flat"
  top: "lossa"
}
layers {
name: "accuracy"
type: ACCURACY
bottom: "ip_label"
bottom: "label"
top: "accuracy"
}

layers {
  name: "loss"
  type: SOFTMAX_LOSS
  bottom: "ip_label"
  bottom: "label"
  top: "lossb"
}



Am 08.02.2015 um 11:43 schrieb Álvaro Begué:

What network architecture did you use? Can you give us some details?



On Sun, Feb 8, 2015 at 5:22 AM, Detlef Schmicker > wrote:


Hi,

I am working on a CNN for winrate and territory:

approach:
 - input 2 layers for b and w stones
 - 1. output: 1 layer territ

Re: [Computer-go] CNN for winrate and territory

[Hugh Perkins]

Detleft wrote:
> The idea is, I can do the equivalent of lets say 1000 playouts with a call to 
> the CNN for the cost of 2 playouts some time...

That sounds like a good plan :-)
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Re: [Computer-go] CNN for winrate and territory

[Álvaro Begué]

What network architecture did you use? Can you give us some details?



On Sun, Feb 8, 2015 at 5:22 AM, Detlef Schmicker  wrote:

>  Hi,
>
> I am working on a CNN for winrate and territory:
>
> approach:
>  - input 2 layers for b and w stones
>  - 1. output: 1 layer territory (0.0 for owned by white, 1.0 for owned by
> black (because I missed TANH in the first place I used SIGMOID))
>  - 2. output: label for -60 to +60 territory leading by black
> the loss of both outputs is trained
>
> the idea is, that this way I do not have to put komi into input and make
> the winrate from the statistics of the trained label:
>
> e.g. komi 6.5: I sum the probabilites from +7 to +60 and get something
> like a winrate
>
> I trained with 80 positions with territory information through 500
> playouts from oakfoam, which I symmetrized by the 8  transformation leading
> to >600 positions. (It is expensive to produce the positions due to the
> playouts)
>
> The layers are the same as the large network from Christopher Clark
> , Amos Storkey
>  :
> http://arxiv.org/abs/1412.3409
>
>
> I get reasonable territory predictions from this network (compared to 500
> playouts of oakfoam), the winrates seems to be overestimated. But anyway,
> it looks as it is worth to do some more work on it.
>
> The idea is, I can do the equivalent of lets say 1000 playouts with a call
> to the CNN for the cost of 2 playouts some time...
>
>
> Now I try to do a soft turnover from conventional playouts to CNN
> predicted winrates within the framework of MC.
>
> I do have some ideas, but I am not happy with them.
>
> Maybe you have better ones :)
>
>
> Thanks a lot
>
> Detlef
>
>
> ___
> 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] CNN for winrate and territory

[Detlef Schmicker]

Hi,

I am working on a CNN for winrate and territory:

approach:
 - input 2 layers for b and w stones
 - 1. output: 1 layer territory (0.0 for owned by white, 1.0 for owned 
by black (because I missed TANH in the first place I used SIGMOID))

 - 2. output: label for -60 to +60 territory leading by black
the loss of both outputs is trained

the idea is, that this way I do not have to put komi into input and make 
the winrate from the statistics of the trained label:


e.g. komi 6.5: I sum the probabilites from +7 to +60 and get something 
like a winrate


I trained with 80 positions with territory information through 500 
playouts from oakfoam, which I symmetrized by the 8 transformation 
leading to >600 positions. (It is expensive to produce the positions 
due to the playouts)


The layers are the same as the large network from Christopher Clark 
, Amos Storkey 
 : 
http://arxiv.org/abs/1412.3409



I get reasonable territory predictions from this network (compared to 
500 playouts of oakfoam), the winrates seems to be overestimated. But 
anyway, it looks as it is worth to do some more work on it.


The idea is, I can do the equivalent of lets say 1000 playouts with a 
call to the CNN for the cost of 2 playouts some time...



Now I try to do a soft turnover from conventional playouts to CNN 
predicted winrates within the framework of MC.


I do have some ideas, but I am not happy with them.

Maybe you have better ones :)


Thanks a lot

Detlef

___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go