Latent Compression OOD Soft Sensor

Research models for predicting penicillin and biomass from Raman spectra on IndPenSim V3: frozen CDAE/CVAE/beta-VAE/PCA/PLS encoders and physics-informed LSTM models, with neural ODE and optional Mamba and liquid variants.

Install

pip install indpensim_bench

Python 3.10 or later is required. Install the PyTorch build appropriate for your hardware if you need a specific CUDA runtime. Optional variants:

pip install "indpensim_bench[liquid]"
pip install "indpensim_bench[cde]"

The mamba extra requires a compatible CUDA/Linux build environment.

Library usage

from indpensim_bench.config import Config
from indpensim_bench.algorithms import get_algorithm
from importlib.resources import files

config_path = files("indpensim_bench").joinpath(
    "configs/beta_vae_pi_lstm.yaml"
)
cfg = Config.from_yaml(config_path)
algorithm_class = get_algorithm("beta_vae_pi_lstm")

Configuration paths for datasets and checkpoints must be adapted to your local files. The dataset, trained checkpoints and fitted scalers are not included. Obtain IndPenSim V3 separately from the dataset provider.

Reproduce the benchmark

The wheel provides the Python library. The source distribution also includes the original src/, train.py, evaluate.py, run_seeds.py, scripts/, configuration files and the full benchmark README. Extract the source archive and run these workflows from its root using the original README instructions. The original requirements file records the research environment; installing the library uses the smaller dependency list in pyproject.toml.

Existing checkpoints serialized with imports under src should be loaded from the original research source tree; automatic checkpoint migration is not provided.

Source repository and full documentation