MicroRTS 32x32

d01c6e3 “Deprecate” _Microrts-squnet-map12-selfplay
eccf7f1 Evaluate map32 every 5M steps, map64 10M steps
c4bed6d map64 uses 64 channels across all levels
9ef4d96 Agent support for multiple models based on map size
37989c2 Fix ACTION_TYPE_TO_ACTION_INDEXES checking
97dd419 Restore prior model structure to be able to run old models
7b56dad Separate combat produce reward between types
74171a4 No val-clip squnet

fddc017 128 channels across all levels for map32 and 64
c0a4e6c Need to remember to update both references
c1f666d Catch exceptions from non-RAI AIs
c80ad07 Fix unhandled argument crash
51aaf61 Compiled resources observation into jar
4b86d47 Add player and opponent resources to observation
477fa93 map64: Make network narrower at the top
bec93ad map64: 16 envs to meet map training invariants
ee41e07 map64: Half envs and rollout steps
607b2e2 48 minibatches for 64x64 maps…
e0eba44 64x64 requires 32 minibatches…

aa45849 Reduce minibatch size for 32 and 64 maps for memory
cbad659 16, 32, and 64 Microrts squnet hyperparameters
855e79b Critic heads use same strides as used by the unet
ab7d32a wide map12 and fix to critic_heads num_channels
8618698 Remove unnecessary env make_kwarg
ef3675c Forgot to set bot alternating player during training
ccc490f Fix squnet-map12 eval env overrides
083c565 Set squnet-map12 to use 8x8 map for eval
6d5785f Microrts-squnet-map12 uses one minibatch
9189e1a min grid size is 12
1d93b6f Use ModuleList to register list of modules with Pytorch
f990525 Accidentally deleted a hyperparam copy reference
d78da94 SqueezeUnet training on 8x8-12x12 maps
b1f0d91 Down-res convolutition should have GELU non-linearity