Dataset preprocessor for the KGS go dataset, eg according to Clark and Storkey input planes
Python HTML

Table of Contents generated with DocToc


Dataset preprocessor for the KGS go dataset, eg according to Clark and Storkey input planes


The goal of this project is to take the data from the kgsgo website, and make it available into a somewhat generic format, that can be fed into any go-agnostic learning algorithm. The guidelines used for the creation of this project is to be able to somewhat reproduce the experiments in the Clark and Storkey paper, and also somewhat targetting the Maddison et al paper.


  • python (tested with 2.7, but including lots of python3 compatibility headers, but gomill would need a bit of tweaking, if we want to run with full, real python 3)
  • internet connection
  • about 20GB disk space

v1 vs v2 format

  • both formats contain the same data
  • v2 format is arguably a more generic format, eg it contains metadata inside describing the arrangement and type of the data

v1 format


These are written for linux. They may need some slight tweaking for Windows


git clone --recursive
cd kgsgo-dataset-preprocessor


  • the datasets are downloaded from, into data subdirectory
  • processed into .dat files, with the same name as the unzipped zip files, just with .dat instead of .zip, also in data directory
  • finally, consolidated together, into two three .dat file, in the data subdirectory:
    • kgsgo-test.dat: data from 100 games, randomly selected, and itemized in
    • kgsgo-train10k.dat: data from 10,000 games, randomly selected, and non-overlapping with the test games
    • (and we can easily create a larger training data set, with everything except the test games, by just adding 3 lines in, at around line 315, but I figure I'm going to try with train10k first, before filling up my hard drive :-D

Data format of resulting file

  • it's a binary file, consisting of 1 or more fixed size records, followed by bytes 'E' 'N' 'D'
  • Each record represents one labelled training example. Each training example has the following format:
    • 2 bytes 'G' 'O', just to help verify alignment during loading
    • 2 bytes containing the label:
      • First byte is the moverow, 0 to 18
      • Second byte is the movecol 0 to 18
      • To get the label for this example, 0 to 360, you can calculate:
        • row * 19 + column
    • 361 bytes (19 by 19) containing the training features, representing a 19x19x8 cube
      • Each byte represents a vertical cross-section through the cube, 8 layers, with each bit as the value, 0 or 1
      • The bytes are laid out in rows, with 19 bytes per row, for a total of 19 rows, each of 19 bytes

Data processing applied

  • The data cubes created are derived from Go games, downloaded from
  • each training example comprises one position in the game, with the label being the subsequent next move in the game
  • goal is to predict the next move in the game (which might not always be the game-theoretic optimal move)
  • The planes of each training data cube are derived as follows, with plane order starting from the least significant byte in each byte of training data:
    • plane 0: location has our own piece, with 1 liberty
    • plane 1: location has our own piece, with 2 liberties
    • plane 2: location has our own piece, with 3 or more liberties
    • plane 3: location has opponent piece, with 1 liberty
    • plane 4: location has opponent piece, with 2 liberties
    • plane 5: location has opponent piece, with 3 or more liberties
    • plane 6: location would be an illegal simple-ko
    • plane 7: always 1 (represents the valid area of the board, since we might be zero-padding the images during CNN training)


When I run it, I get md5sums:

850d2c91b684de45f39a205378fd7967  kgsgo-test.dat
80cfa39797fa1ea32af30191b2fb962c  kgsgo-train10k.dat

If it's different, it doesn't necessarily matter, but if it's the same, it's a good sign :-)

v2 format

v2 format vs v1 format

After writing v1 format as detailed above, I noticed some things I'd prefer to do differently. Therefore, v2 format modifies these things, but without changing anything detailed above. If you continue to use, then the data produced will be unchanged. In addition the filenames produced by v2 do not overwrite those produced by the earlier version.

v2 changes the following:

  • run by doing python
  • data files have a header, 1024 bytes, with the following format, eg:
  • where:
    • n= gives number of examples
    • int means the data is in ints, cf float, if in floats
    • bpp= gives the number of bits per pixel/point
  • each data example is prefixed with 'GO' as before
  • label is provided as a 4-byte integer, comprising nextmoverow * 19 + nextmovecol, in intel-endian
  • data is arranged as in a sql 'group by' clause of: example, plane, row, column
  • data is provided as a bitmap, eg where the 8 bits of the first byte represent the first 8 columns of the first row of the first plane
  • bits are arranged so that if you wrote out the bytes in binary, the 1s and 0s would be arranged in order, ie in sql group by order of: plane, row, column
  • bit continue to write into bytes, with no padding added, until the end of each example, eg if there are 4 plane, with imagewidth 1, imagewidth 1, and the first plane is 1, and the other are 0, then we'd have bits 1000
  • the final byte of each example is 0 padded, on the right hand side, so, in the example in the previous sentence, the byte would become 10000000
  • finally, compared to the previous version, only 7 planes are stored, the plane that is all 1s is omitted

Running v2 format processor


Available options:

  • dir specify the target directory, where the data will be downloaded to, and the datasets generated
  • sets specify which sets to generate. There are three sets available:
    • test: the test set, generated from 100 randomly selected, but fixed, games
    • train10k: training set, generated from 10,000 randomly selected, but fixed, games, non-overlapping with the test games
    • trainall: training set, generated from all games, excluding, and non-overlapping with, the test games
  • eg you can do:
python dir=data sets=test,train10k,trainall

(this is the default in fact, if you run with no arguments)

md5 sums

When I run this, I get the following md5 sums. If these are different for you, it's not necessarily an issue. If they are the same, this is a good sign :-)

57382be81ef419a5f1b1cf2632a8debf  kgsgo-test-v2.dat
6172e980f348103be3ad06ae7f946b47  kgsgo-train10k-v2.dat
20440801e72452b6714d5dd061673973  kgsgo-trainall-v2.dat

File sizes:

$ ls -lh kgsgo-*v2.dat
-rw-rw-r-- 1 ubuntu ubuntu 5.8M Feb  8 05:58 kgsgo-test-v2.dat
-rw-rw-r-- 1 ubuntu ubuntu 601M Feb  8 06:13 kgsgo-train10k-v2.dat
-rw-rw-r-- 1 ubuntu ubuntu  11G Mar  7 15:58 kgsgo-trainall-v2.dat

Example loader

Third-party libraries used

Related projects

I'm building a convolutional network library in OpenCL, aiming to train this, at ClConvolve