Figure 3 — Software environments and hand-off through artifacts

Vector PDF

File: fig03_environments.pdf

Suggested caption. Isolation of conflicting dependencies. The VAE pool generator runs in a TensorFlow environment without PyTorch. Assembly, Diffusion-TS, Chronos, TimesFM, Timestral, evaluation and the statistics run in PyTorch environments that never imports TensorFlow. The only link between the environments is the Parquet files they write.

What the figure shows

Environment	Requirements file	Runs	Main dependencies
base	envs/base.txt	raw, prepare	numpy, scipy, pandas, pyarrow, psutil
tsgm	envs/tsgm.txt	pools --pool-generators tsgm_vae	base + TensorFlow, Keras 3, tsgm 0.1.0 (no torch)
sdv (optional)	envs/sdv.txt	pools --pool-generators sdv_par	base + SDV
core	envs/core.txt	generate (assembly), fidelity, utility, collect, hypotheses	base + PyTorch, scikit-learn, XGBoost, statsmodels, ruptures, hmmlearn, elkai, OR-Tools
diffusion_ts	envs/diffusion_ts.txt	generate --methods diffusion_ts	core + einops + Diffusion-TS code
chronos	envs/chronos.txt	generate --methods chronos	core + chronos-forecasting
timesfm	envs/timesfm.txt	generate --methods timesfm	core + timesfm 3.0.2
timestral	envs/timestral.txt	generate --methods timestral	core + timestral

Notes for the text

• The pool is the interface between generator and assembler. The generator writes (N, T, 3) windows to artifacts/pools. generate only reads them, so every same-pool method sees exactly the same windows.
• Each generator declares its environment and the modules it needs. If the modules are missing and the pool does not exist yet, the stage skips that generator and prints which environment to use.
• Docker: target tsgm runs stage 2, target bench runs the rest. In Colab each notebook has its own runtime. run_in_env(...) creates a venv from envs/<env>.txt and runs the stage as a subprocess.