search

harbecke / HexHex

AlphaGo Zero adaptation for Hex
GNU General Public License v3.0
20 stars 4 forks source link

storing training data internally #22

Closed harbecke closed 5 years ago

harbecke commented 5 years ago

The training data is now only saved after the whole repeated_self_training script has run. There are several changes to the sample_config that you have to adopt, most importantly the old train_samples_pool_size is now samples_per_model * num_data_models. samples_per_model indicates how many samples a model generates, and num_data_models of how many of the last models we use data from.

The data is now saved under the same name as the model (in the data folder). You can start training with this data if you set load_initial_data=True (it should not be required to have the correct amout of data). Otherwise the initial training data is generated by a Random model.

Please let me know if this works on your system!

(the important commit is only 383431b, i just wanted to merge anyway and use a pull request)

simonant commented 5 years ago

The creation of the not already existing puzzle data does not work for me. The problem seems to be that config cannot be deepcopied. When adding boardsize to config instead of puzzle_config everything seems to work. I do not know how to properly fix this.