- Published on
Show HN: I Turned DOOM into a Transformer
- Authors
- Name
- DeepHorus
DOOMFORMER, a new open-source project introduced on Show HN, runs neural networks inside DOOM maps using voodoo dolls as computing elements. It compiles network layers into map geometry, including doll crowds, weight lines and staircase gates. Every exposed value is read back and checked against NumPy with zero tolerance.
Available via pip for Python 3.11+ on Linux and macOS, the tool offers quick demos. A “hello” run verifies a small MLP in about two seconds. A digit-drawing mode takes roughly 15 seconds, with a marine walking the digit as the level reads it.
In the main MNIST demonstration, 1,355 voodoo dolls run a network in a single map. The winning digit’s doll crosses a line that lights a seven-segment display. DOOM computes 248 exposed values in 273 game tics, about eight seconds of real-time play at 35 tics per second.
The project reports exact agreement: all 248 values match an integer NumPy reference at tolerance zero. Yet it cautions that exactness does not guarantee correct classification. The float model scores 97.58 percent, and the network also trained on scripted marine strokes.
A Python-free museum.wad demo runs in Freedoom and GZDoom with 136 voodoo dolls and seven floor pads. Pressing MUTATE rewires one connection during play, turning the same 7 into a 1; RESTORE returns it to 7. The compiler also works around engine quirks such as ignored floor-mover requests, teleport line cancellations and BSP placement limits.
The repo includes a facts command that tests engine behaviors in about 10 seconds, and a self-test version runs in GZDoom, LZDoom, UZDoom and ZDoom 2.8.1. It builds on earlier DOOM-map computers such as counter.wad and conway.wad, adding a neural-network compiler and exact checks. The authors stress that verification proves only the inputs that ran, not full network behavior.
