---
license: other
tags:
  - trajectory-prediction
  - diffusion
  - flow-matching
---

# SRA — Spatial Reasoning Adapter (code)

Code for running **SRA**, a modular future-interaction graph that plugs into three stochastic
trajectory predictors — **MID** (DDPM), **LED** (leapfrog-DDPM) and **MoFlow** (flow matching) —
on **NBA**, **Soccer** and **Football**.

## 👉 Start here: [`SETUP.md`](SETUP.md) — download, install, run

It covers the required directory layout, environment setup, the exact training command for each
host × dataset, the E2 ablation settings, environment-variable switches, the adapter contract,
and known gotchas.

## Documents

- [`SETUP.md`](SETUP.md) — end-to-end setup from scratch (env, data, paths, first run)
- [`RUNNING.md`](RUNNING.md) — how to run every host × dataset
- [`GAMEFORMER_SRA.md`](GAMEFORMER_SRA.md) — GameFormer+SRA negative result (does SRA generalize to feedforward models?)
- [`sample_data/README.md`](sample_data/README.md) — bundled 100-scene NBA smoke-test subset

## Contents

```
MoFlow/   flow-matching host + the SRA graph module + E4 baseline modules
MID/      DDPM host
LED/      leapfrog-DDPM host
```

**Important:** `MID/` and `LED/` import the SRA graph from a *sibling* `MoFlow/` directory at
runtime — keep the three folders side by side. See §0 of `RUNNING.md`.

## Not included

- **Datasets** (NBA / soccer / football `.npy` files) — see §2 of `RUNNING.md` for the expected paths.
- **Checkpoints and training logs** — code only.