jason-yoo-108/lifelong-bouncing-balls-c

Lifelong Bouncing Balls (C)

One of the five single-video-stream datasets introduced in Lifelong Learning of Video Diffusion Models From a Single Video Stream (arXiv:2406.04814). Code: https://github.com/plai-group/lifelong-vdm.

Same simulation as Lifelong Bouncing Balls (O): two balls bounce deterministically in a 2D box and cycle through red → yellow → red → green on every collision. The "(C)" version adds non-stationarity: the blue channel of every ball color increases at a constant rate over the training stream, so red, yellow and green become fuchsia, white and aqua by the end of the video. Details in the stream therefore never repeat. The test stream contains balls in all previously observed colors. Solving it requires learning the deterministic dynamics and color transitions from a correlated stream with unrepetitive details; in the paper, streaming methods without a replay buffer forget the early colors.

	Train stream	Test stream
Frames	1,000,000	1,000,000
Resolution	32 x 32 RGB	32 x 32 RGB
Frame rate	10 FPS (~28 h)	10 FPS (~28 h)
Chunks	1,000 x 1,000 frames	1,000 x 1,000 frames

Files

config.json                 {"T_total": 1000000, "chunk_size": 1000, "resolution": 32, "seed": 0,
                             "num_balls": 2, "color_shift": true, "frame_dtype": "float32", ...}
train/0/0.npy … 999.npy     float32 arrays of shape (1000, 32, 32, 3), values in [0, 1], RGB
train/0/ball_positions.npy  float64 array of shape (1000000, 2, 2): ball centre coordinates
                            per frame (frame, ball, coordinate), the ground truth used for minADE
test/0/0.npy … 999.npy      same layout as train
test/0/ball_positions.npy

Chunk i holds stream frames i*1000 … i*1000+999, so frame t lives in train/0/{t // 1000}.npy[t % 1000]. Chunks are consecutive; nothing is shuffled. The 0 directory level is the generator seed used by the loaders in the code base. The frames are stored here as float32.

Loading

import numpy as np
chunk = np.load("train/0/0.npy")          # (1000, 32, 32, 3) float32 in [0, 1]
frame_12345 = np.load(f"train/0/{12345 // 1000}.npy")[12345 % 1000]

To use the dataset with the paper's code (it expects the directory at datasets/ball_nstn):

pip install -U huggingface_hub
hf download jason-yoo-108/lifelong-bouncing-balls-c --repo-type dataset --local-dir datasets/ball_nstn
# or: python datasets/download.py ball_nstn

Provenance and license

Fully synthetic; generated with datasets/ball.py --color_shift (seed 0) from the code repository, which can regenerate it. Released under CC BY 4.0.

Citation

@article{yoo2024lifelong,
  title   = {Lifelong Learning of Video Diffusion Models From a Single Video Stream},
  author  = {Yoo, Jason and He, Yingchen and Naderiparizi, Saeid and Green, Dylan and van de Ven, Gido M. and Pleiss, Geoff and Wood, Frank},
  journal = {arXiv preprint arXiv:2406.04814},
  year    = {2024}
}