billzorn / mtg-rnn Goto Github PK

The most recent release of AllSets.json from MTGJson has some changes that break the script.

1. Rarities are encoded in lower-case letters only
2. Mythic Rare is stored as "mythic"
3. Time Spiral Timeshifted rarity has been moved to rare
4. Basic Lands are now common.

Other than a couple silver-bordered sets that need excluding to avoid all the unicode characters, it seems to work when the lower-case rarities are implemented for common through mythic. The last two merely affect how the net will associate rarities, and are less important.

I'm guessing somewhere along the way that the order that fields are fed to the neural network changed, and the script was never updated. Priming with the rarity parameter (I think) actually primes the name, and so on.
A sound plan would probably be to use the numbers that follow the fields as an identifier layer (those are less likely to change than the field order) and have the script use those as a mapping to the correct fields. This should reduce the amount of maintenance that would need to be done when column orders are changed in the future.

Not sure if this is even a bug with the repo, since it seems like a Lua error. Just wondering if anyone's run into this issue before.

Command

th train.lua -rnn_size 256 -num_layers 3 -seq_length 200 -data_dir data/mtgencode-std -checkpoint_dir cv/mtgencode-std/ -eval_val_every 1000 -seed 7767

Result

...
997/12000 (epoch 4.154), train_loss = 0.55129399, grad/param norm = 1.9592e-01, time/batch = 2.6165s
998/12000 (epoch 4.158), train_loss = 0.55412990, grad/param norm = 2.0948e-01, time/batch = 2.6173s
999/12000 (epoch 4.163), train_loss = 0.55656064, grad/param norm = 2.2091e-01, time/batch = 2.5914s
evaluating loss over split index 2
1/13...
PANIC: unprotected error in call to Lua API (not enough memory)

• OS X 10.11.6 El Capitan
  • (I know, I know, the repo says to use Linux, but everything seems to run fine)
• Memory: 16GB 1600 MHz DDR3
• Graphics: Intel Iris Pro 1536 MB

cltorch has changed their installation process. It instead goes through distro-cl to get cltorch and clnn.

I haven't gleaned all the details, but the process in the readme on the front page has been obsoleted.