Train a transformer to produce Hadamard matrices. Install core dependency with:

pip install -r requirements.txt

Install one logging backend:

pip install -r requirements-wandb.txt

or

pip install -r requirements-tensorboard.txt

check.py has additional dependencies:

pip install -r requirements-check.txt

For wandb logging, you can override defaults with:

export WANDB_ENTITY=your_entity
export WANDB_PROJECT=your_project

Modify the parameters in params.py (size of matrix, transformer parameters, etc) then run:

python hadamard.py

Data are saved in subdirectories of training/; runs can be resumed with the parameter resume in params.py. A symlink latest/ is created to point to the latest data. Hadamard matrices for each run are saved in hada.txt. Use check.py to check that they are indeed hadamard matrices, e.g.,

python check.py latest/hada.txt

Logging is in two forms:

• there is wandb or tensorboard logging, which summarises three types of information:
  • Loss in the training process.
  • Mean scores of each generation.
  • Proportion of Hadamard matrices in each generation.
• There is a stats.txt that contains complementary information, including tallies of the generations at which matrices have been produced. Each generation has two entries:
  • sample which is the sample generated by the transformer (or at the start, a random sample).
  • selected which is a subset of the data after improvement and top score selection.

There is also a standalone program exactgs.py to perform exact enumeration of GS type Hadamard matrices for small n, e.g., for n=64:

python exactgs.py 64