Sample data — NBA (100 scenes)

A tiny smoke-test subset so you can verify the code runs end-to-end before wiring up the full datasets. Not for reproducing any reported number.

sample_data/nba/original/nba_train.npy   # (100, 30, 11, 2) float32, 258 KB
sample_data/nba/original/nba_test.npy    # (100, 30, 11, 2) float32, 258 KB

Layout: (scenes, frames, agents, xy) — 30 frames @ 5 Hz = 10 past + 20 future (4.0 s), 11 agents (10 players + ball, ball = index 10), absolute court coordinates.
Each file is the exact first-100 prefix of the corresponding full split (trajs[:100]), which is the same slice the loaders take — so scene indices line up with the full dataset.
Full splits are 32 500 train / 12 500 test scenes.

Use it

--data_dir must point at the parent of original/ (the loader appends original/ itself):

cd MoFlow
CUDA_VISIBLE_DEVICES=0 python fm_nba_graph_v6.py \
  --cfg cfg/nba/cor_fm.yml --exp smoke \
  --data_dir ../sample_data/nba \
  --n_train 100 --n_test 100 \
  --batch_size 8 --epochs 1 \
  --fm_in_scaling --tied_noise --top_n_neighbors 5 --uncertainty_weight 0.01

Verified to load and collate:

past_traj                 (B, 11, 10, 6)
past_traj_original_scale  (B, 11, 10, 6)
fut_traj_original_scale   (B, 11, 20, 2)

With only 100 scenes the model cannot learn anything meaningful — expect high ADE/FDE. This subset exists purely to confirm the data path, model construction, training step and evaluation loop all execute.

Provenance

Derived from the NBA player-tracking (SportVU) movement data as preprocessed by prior trajectory-prediction work. Redistributed here only as a minimal fixture for code testing; please refer to the original data source for licensing and terms of use.