Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,44 @@
+---
+license: mit
+tags:
+  - reinforcement-learning
+  - autoencoder
+  - games
+---
+
+# OpenFront tile-state autoencoder
+
+Fully-convolutional autoencoder over [OpenFront.io](https://openfront.io)
+tile-ownership state, intended as a frozen spatial observation encoder for RL
+agents. 64x compression: 16x16 spatial downsampling into 64 latent channels
+per region, works on any map size.
+
+- Input: per-tile owner slot (static per-game slot assignment, 8-dim learned
+  embedding) + terrain channels (land, magnitude, fallout)
+- Loss: border-weighted cross-entropy reconstruction of owner slots
+- Trained on [djmango/openfront-snapshots](https://huggingface.co/datasets/djmango/openfront-snapshots)
+  (250 games, 10 maps, random 256x256 crops): 30k steps, batch 16, ~40 min on
+  an RTX 3070. 1.0M params.
+
+![training curves](loss_curve.png)
+
+Original vs reconstruction through the latent (World map, 17 players):
+
+![reconstruction](recon_world.png)
+
+Border-tile accuracy (the honest metric; overall accuracy is inflated by
+water): 98.7% on a 2-player endgame, 91.4% on World with 17 players, 90.1% on
+Africa with 27 players.
+
+Training code and usage: [djmango/openfront-ae](https://github.com/djmango/openfront-ae)
+
+```python
+import torch
+from ae.model import TileAutoencoder
+
+ckpt = torch.load("ae.pt", map_location="cpu", weights_only=False)
+model = TileAutoencoder(latent_c=ckpt["args"]["latent_c"])
+model.load_state_dict(ckpt["model_state_dict"])
+model.eval()
+z = model.encode(owner_slots, terrain)  # (B, 64, H/16, W/16)
+```