Files
worldmodel/JEPA/lejepa-identifiability/experiments/configs/scaling.yaml
T
gaojie c66855adfc
Sync to site1 / sync (push) Has been cancelled
refactor: 将子模块转为普通目录,移除外部 git 依赖
- 移除 JEPA/lejepa-identifiability 子模块 gitlink
- 移除 research/multiply/MultiPLY 子模块 gitlink
- 删除 .gitmodules(不再有外部 URL 依赖)
- 两个目录内容作为普通文件纳入主仓库追踪
- 删除各自内部 .git 目录,消除嵌套 git 仓库
2026-06-05 17:14:01 +08:00

36 lines
846 B
YAML

# Scaling experiment (paper figure)
# python run.py --config configs/scaling.yaml --N 16 --seed 0
# python run.py --config configs/scaling.yaml --N 16 --seed 0 --mode infonce
# python run.py --config configs/scaling.yaml --N 16 --seed 0 --mode whiten
experiment: scaling
out: results/scaling
# Data
source_dist: gaussian
num_eval: 10000
# Training (shared)
steps: 20000
lr: 3.0e-3
batch_size: 256
rho: 0.95
log_every: 500
# Encoder
encoder: matched
n_layers: 4
mode: lejepa # default; override with --mode
# Mode-specific defaults (used based on --mode)
lamb: 1.0e-6 # for lejepa
lamb_whiten: 0.5 # used when mode=whiten
sigma: 1.0 # for infonce
# Mixing
mixing: coupling
# Sweep dimensions
dims: [2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
seeds: [0, 1, 2, 3, 4]
K: 3 # parallel encoder runs per (N, seed); pick lowest loss